ECoG of BCI Thesis

ECoG signal processing for Brain Computer Interface
with multiple degrees of freedom for clinical application

Marie-Caroline Schaeffer
To cite this version:

Marie-Caroline Schaeffer. ECoG signal processing for Brain Computer Interface with multiple degrees
of freedom for clinical application. Medical Physics [physics.med-ph]. Université Grenoble Alpes,
2017. English. �NNT : 2017GREAS026�. �tel-01763451�
HAL Id: tel-01763451

https://tel.archives-ouvertes.fr/tel-01763451
Submitted on 11 Apr 2018
HAL is a multi-disciplinary open access L’archive ouverte pluridisciplinaire HAL, est

archive for the deposit and dissemination of sci- destinée au dépôt et à la diffusion de documents
entific research documents, whether they are pub- scientifiques de niveau recherche, publiés ou non,
lished or not. The documents may come from émanant des établissements d’enseignement et de
teaching and research institutions in France or recherche français ou étrangers, des laboratoires
abroad, or from public or private research centers. publics ou privés.
THÈSE
Pour obtenir le grade de
DOCTEUR DE LA COMMUNAUTE UNIVERSITE
GRENOBLE ALPES
Spécialité : MBS – Modèles, méthodes et algorithmes en biologie,
santé et environnement
Arrêté ministériel : 25 mai 2016
Présentée par
Marie-Caroline SCHAEFFER
Thèse dirigée par Tetiana AKSENOVA (EDISCE), CEA
préparée au sein du Laboratoire CLINATEC

dans l'École Doctorale Ingénierie pour la Santé, la Cognition et
l’Environnement
Traitement du signal ECoG pour

Interface Cerveau Machine à grand
nombre de degrés de liberté pour
application clinique
Thèse soutenue publiquement le 6 Juin 2017,
devant le jury composé de :
Prof. Alim-Louis BENABID
Président
Dr. Laurent BOUGRAIN
Membre
Prof. François CABESTAING
Rapporteur
Dr. Marco CONGEDO
Membre
Prof. Saïd MOUSSAOUI
Rapporteur
Prof. Laura SACERDOTE
Membre
i
Acknowledgements
I feel very grateful to the members of my committee, Prof. Alim-Louis BENABID,
Dr. Laurent BOUGRAIN, Dr. Marco CONGEDO and Prof. Laura SACERDOTE.
My special thanks go to Prof. François CABESTAING and Prof. Saïd MOUSSAOUI
for having accepted to act as reporters of this doctoral work.
My deepest gratitude goes to Dr. Tetiana AKSENOVA who has supervised this
doctoral work. I wouldn’t have been able to complete this work without her constant
guidance, insight, availability, patience and kindness. It has been a great privilege
to work and progress under her supervision.
I would also like to thank Guillaume CHARVET who co-supervised this doctoral
work, and Prof. Alim-Louis BENABID and Corinne MESTAIS who head CLI-
NATEC’s BCI project. Their suggestions have always been helpful and illuminating.
It is a great pleasure to me to extend my thanks to all the associates of CLINATEC
with whom I worked, had insightful discussions with or with whom I simply enjoyed
meals or coffee breaks. Their presence has helped me along more than once.
Finally, I would like to thank my family and friends for their unwavering support
throughout these last three years.
iii
Preface
The present doctoral work, which has been supervised by Dr. T. AKSENOVA and
co-supervised by G. CHARVET, has been prepared at CLINATEC, an Edmond J.
SAFRA Biomedical research centre located on the MINATEC Campus at CEA/LETI
in Grenoble. CLINATEC associates medical research and technological innovation to
translate new solutions to patients, with the purpose of accelerating the development
and clinical validation of innovative medical devices. This doctoral work has been
completed within the framework of CLINATEC’s motor Brain-Computer-Interface
project, which is supervised by Prof. A.-L. BENABID, C. MESTAIS and G.
CHARVET. The goal of this project is to bring the proof of concept that it is feasible
for a tetraplegic subject to control complex effectors, for example a 4-limb exoskeleton,
thanks to the monitoring and decoding of his brain activity. CLINATEC’s "BCI
and Tetraplegia" 5-year clinical trial has recently been authorized by the French
regulatory agencies. Signal processing challenges specific to the clinical deployment
of CLINATEC’s Brain-Computer-Interface system, namely asynchronous mono-
limb decoding, asynchronous sequential multi-limb decoding and decoding accuracy
during active states, have been addressed in the present doctoral thesis.
Parts of the presented work have been published in journals or conference
proceedings, or exposed during oral presentations:
Journal
Schaeffer, M. C., & Aksenova, T. (2017). Switching Markov Decoders for Asyn-
chronous Trajectory Reconstruction from ECoG signals in Monkeys for BCI Applica-
tions. Journal of Physiology-Paris. doi: 10.1016/j.jphysparis.2017.03.002; Published
online: 10 March 2017.
International conference proceedings

Schaeffer, M. C., & Aksenova, T. (2016, September). Hybrid Trajectory Decoding
from ECoG Signals for Asynchronous BCIs. In International Conference on Artificial
Neural Networks, Barcelona, Spain (pp. 288-296). Springer International Publishing.
Patent pending
Schaeffer, M. C., & Aksenova, T. (June 2016). Interface neuronale directe à décodage
continu utilisant un mélange markovien d’experts.
Oral communications
M.C. Schaeffer, T. Aksenova. Markovian Mixture of Experts for 3D wrist movement
decoding from ECoG recordings. GDR 2904 Multielectrode systems for Neuroscience:
6th annual meeting, 5-8 Jan 2016, Grenoble-Autrans. (speaker)
iv
M.C. Schaeffer, E. Labyt, V. Rohu, N. Tarrin, I. Vergara, S. Cokgungor, A. Eliseyev,

G. Charvet, C. Mestais, A.L. Benabid, T. Aksenova. Hand movement decoding from
magnetoencephalographic signals for BCI applications. Société de Neurophysiologie
Clinique de Langue Française (SNCLF) 2015, 1-3 Juillet 2015, Grenoble.
M.C. Schaeffer, S. Cokgungor, N. Tarrin, V. Rohu, E. Labyt, C. Mestais, A.L. Ben-

abid, T. Aksenova. Decoding of 3D hand movements from Magnetoencephalography
recordings of brain neuronal activity: preliminary results. BioHealth Computing
Workshop, 1-3 Septembre 2014, Grenoble. (speaker)
v
Résumé
Introduction Les Interfaces Cerveau-Machine (ICM) sont des systèmes qui per-
mettent à des patients souffrant d’un handicap moteur sévère d’interagir avec leur
environnement en utilisant leur activité cérébrale pour contrôler des effecteurs ex-
térieurs. Plusieurs étapes sont généralement nécessaires pour convertir l’activité
cérébrale du patient en commandes permettant de contrôler un effecteur extérieur.
Des caractéristiques spécifiques aux intentions de l’utilisateur sont tout d’abord
extraites de son activité cérébrale, préalablement acquise et digitalisée. Un décodeur
est appliqué sur ces caractéristiques cérébrales, et permet de les convertir en esti-
mations des intentions du sujet. Après une étape optionnelle de post-traitement
susceptible d’améliorer la qualité des estimations, les intentions décodées du sujet
sont converties en commandes utilisées pour contrôler le ou les effecteur(s) de l’ICM,
par exemple des orthèses ou des prothèses de membres inférieurs ou supérieurs dans
le cas d’ICM motrices.
L’objectif de l’ICM motrice de CLINATEC est de permettre à des patients
tétraplégiques de recouvrer de façon chronique une indépendance motrice en mod-
ulant leur activité électrocorticographique (ECoG) pour contrôler des effecteurs
complexes tels qu’un exosquelette 4 membres. Des challenges spécifiques au dé-
ploiement clinique d’ICMs multi-effecteurs ont été étudiés dans la présente thèse de
doctorat.
Un problème majeur des ICMs cliniques est la capacité à proposer aux utilisateurs
un contrôle asynchrone sur l’effecteur. Contrairement aux ICMs synchrones qui sont
périodiquement contrôlables par les utilisateurs et requièrent donc la présence d’un
opérateur pour activer et désactiver le système, les ICMs asynchrones sont disponibles
en continu. Lorsqu’une ICM fonctionne en mode asynchrone, il est particulièrement
souhaitable de limiter des activations erronées de l’effecteur pendant les périodes dites
de Non-Contrôle (NC). Un deuxième challenge résulte de la possible présence d’un
effecteur multi-membres, qui rend nécessaire la généralisation du contrôle asynchrone
mono-membre au cas multi-membres. Une stratégie d’activation séquentielle a été
considérée dans la présente thèse, avec le but d’améliorer la robustesse du système
et de faciliter le contrôle cérébral. Au-delà de la limitation des activations erronées
du système, il est nécessaire d’éviter des mouvements parallèles résiduels d’effecteurs
momentanément non-contrôlés. Finalement, la capacité des utilisateurs à exécuter
des mouvements contrôlés par leur activité cérébrale est compromise quand les
estimations de leurs intentions de mouvements ne sont pas suffisamment proches de
leurs véritables intentions. La précision du décodage pendant les périodes dites de
Contrôle Intentionnel (Intentional Control, IC) est donc essentielle.
Décodeurs pour ICM motrices précises et asynchrones Différents outils al-

gorithmiques sont utilisés dans la communauté des ICMs pour assurer une conversion
pertinente des signaux cérébraux des sujets en estimations de leurs intentions, tout
particulièrement des modèles permettant d’estimer la valeur de variables discrètes
(classifieurs) ou continues (modèles de régression ou filtres bayésiens) à partir des
vi
signaux cérébraux. Deux stratégies de décodage sont principalement exploitées par

les ICMs motrices. La première consiste à demander au sujet d’effectuer des tâches
mentales (tout particulièrement des imaginations motrices ou des tâches cognitives,
par exemple des calculs mentaux), qui sont généralement chacune associées à une
direction possible de l’effecteur. La seconde vise à directement extraire des sig-
naux cérébraux les paramètres cinématiques caractérisant la trajectoire désirée de
l’effecteur. Les décodeurs correspondants sont ici désignés sous le terme de décodeurs
cinématiques. Il a été suggéré que les décodeurs cinématiques permettent d’obtenir
des ICMs motrices précises et intuitives. Les décodeurs cinématiques exploitent des
décodeurs continus, principalement des modèles linéaires statiques ou des filtres
bayésiens comme le filtre de Kalman.
Ces décodeurs cinématiques classiques échouent généralement à fournir des
estimations neutres (i.e., associées à une vitesse nulle) pendant les états NC. De ce
fait, des stratégies ont été spécifiquement élaborées pour gérer les états NC, c’est-
à-dire pour limiter des activations erronées de l’ICM pendant les états NC. Deux
approches principales ont émergé pour assurer un décodage asynchrone mono-membre
performant, à savoir la gestion des états NC au niveau du décodeur cinématique
ou d’un potentiel opérateur de post-traitement. Quand la gestion des états NC
est faite au niveau du décodeur, elle est généralement réalisée via l’introduction de
non-linéarités dans le décodeur. L’utilisation de modèles linéaires par morceaux a
particulièrement été rapportée. Ces modèles par morceaux se basent sur une variable
latente discrète pour introduire des non-linéarités dans un décodeur générique, par
exemple un modèle de regression ou un filtre bayésien. Dans le cas de mixtures de
régressions, des décodeurs discrets sont utilisés pour estimer les valeurs prises par la
variable latente à partir des signaux cérébraux.
Un unique modèle est en revanche entraîné et/ou appliqué sur les signaux
cérébraux NC et IC quand la gestion des états NC est faite au niveau du post-
traitement. De nouveau, la valeur d’une variable latente discrète est généralement
utilisée pour sélectionner l’opérateur de post-traitement à appliquer sur les estima-
tions de mouvements.
Le décodage asynchrone multi-membres a été rarement étudié dans les études
ICMs. L’utilisation d’un modèle par morceaux a néanmoins été rapportée pour le
décodage de mouvements séquentiels multi-doigts.
Finalement, les tentatives pour améliorer la précision du décodage pendant les
périodes IC se concentrent généralement sur l’utilisation de modèles non-linéaires,
en particulier des décodeurs non-linéaires classiques (par exemple, des réseaux
de neurones) ou linéaires par morceaux. Des méthodes de post-traitement sont
également utilisées pour améliorer certaines caractéristiques des estimations des
paramètres cinématiques d’intérêt, par exemple le bruit haute fréquence (lissage) ou
la précision spatiale.
Méthodes Des modifications de la dépendence entre signaux cérébraux et paramètres

cinématiques ont été constatées dans plusieurs études. Ces modifications ont par
vii
exemple été observées quand des sujets se trouvaient en états NC ou IC, lorsqu’ils ef-
fectuaient des movements unimanuels ou bimanuels, ou encore au cours de différentes
phases d’un mouvement du bras vers des cibles. Il a par conséquent été proposé dans
la présente thèse d’utiliser un modèle par morceaux, plus précisemment un modèle
markovien linéaire par morceaux (Markov Switching Linear Model, MSLM), pour
répondre au challenge d’un contrôle asynchrone mono- and multi-membre précis.
Ce modèle présente trois points clefs.
Le MSLM commute entre plusieurs modèles, à savoir un modèle NC et un ou
plusieurs modèles ICs. Contrairement aux approches basées sur le post-traitement qui
exploitent un unique modèle continu, le MSLM prend donc en compte de potentielles
modifications d’une dépendance linéaire entre signaux cérébraux et paramètres
cinématiques. Les études de la littérature mentionnées plus haut suggèrent que
la commutation entre modèles est susceptible d’être utile à la fois pour intégrer
la gestion des états NC (un modèle NC, un modèle IC) et de mouvements multi-
membres séquentiels (un modèle NC, un modèle IC par membre) et pour améliorer
la précision des estimations des paramètres cinématiques pendant les états IC (un
modèle NC, un modèle IC par phase de mouvement). La pertinence de chaque
modèle continu du MSLM, i.e. la valeur d’une variable latente discrète indiquant quel
modèle continu est approprié, est déduite à chaque instant des données cérébrales.
Elle permet de déduire une règle probabilistique utilisée pour pondérer les modèles
continus du MSLM.
Le MSLM a été développé dans le cadre des modèles de régression statiques, et
plus précisément comme une extension des Mélanges d’Experts (ME). Un modèle
linéaire discriminant entre les caractéristiques cérébrales et les paramètres cinéma-
tiques est conditionné par l’état courant d’une variable latente discrète. La valeur
de cette variable latente est directement déduite des signaux cérébraux et de leur
distribution durant chaque état possible. Cette propriété distingue le MSLM des
mixtures de filtres bayésiens dont l’utilisation a également été considérée dans la
littérature. Ces filtres sont basés sur des modèles génératifs, et la valeur de la variable
latente est déduite de l’adéquation entre chaque filtre et l’historique des signaux
cérébraux. De surcroît, le MSLM est capable de traiter des données de grande
dimension sans qu’il soit nécessaire de réaliser une étape préliminaire de réduction
de dimension, laquelle est généralement nécessaire pour des modèles génératifs.
A l’opposé des mixtures de modèles de régression précédemment utilisées dans
des études ICM, le MSLM réalise une détection d’état dynamique pour limiter les
fausses activations de l’effecteur. Plus précisément, il est supposé que la séquence
d’états latents (par exemple, NC et IC) est générée par une chaîne de Markov
d’ordre 1. Cette hypothèse est susceptible de réduire le nombre d’états mal détectés,
mais surtout d’améliorer les caractéristiques temporelles des fausses activations,
c’est-à-dire de privilégier des fausses activations plus longues mais également plus
rares. Un Modèle de Markov Caché (Hidden Markov Model, HMM) est utilisé pour
estimer la valeur de l’état discret. Le modèle des Mixtures d’Experts a été modifié
de façon à gérer des détections d’états dynamiques plutôt que statiques.
Des procédures d’apprentissage supervisé et non supervisé ont été présentées
viii
dans la présente thèse. La procédure d’apprentissage supervisé est pertinente quand

la valeur prise par la variable latente discrète est connue dans le jeu de données
d’entraînement, ce qui est par exemple le cas pour les états NC et IC. L’apprentissage
non-supervisé permet d’exploiter une variable latente discrète sans connaissance a
priori sur sa valeur dans le jeu de données d’entraînement. Il peut en particulier être
utilisé pour associer différents modèles IC à chacune des phases d’un mouvement
complexe. Cette procédure d’apprentissage non-supervisé est basée sur l’algorithme
espérance-maximisation (Expectation-Maximization, EM), et généralise à la fois
l’algorithme EM utilisé pour entraîner des HMMs et celui qui permet d’identifier les
paramètres de MEs.
Validation Deux jeux de données ont été utilisés pour explorer la capacité du
MSLM à réaliser du décodage ECoG asynchrone mono- et multi-membres précis. Ces
deux jeux de données sont libres d’accès. Le premier jeu de données est composé de
données précliniques acquises pendant que des Primates Non-Humains réalisaient des
mouvements asynchrones avec un bras (mono-membre) pour atteindre des cibles. Le
second jeu de données rassemble des données cliniques enregistrées pendant que des
patients exécutaient des mouvements de doigts séquentiels (flexions et extensions).
Deux décodeurs correspondant à des stratégies précédemment proposées pour
intégrer la gestion des états NC dans des décodeurs cinématiques ont été choisis
pour évaluer la performance comparative du MSLM. Le premier décodeur alternatif
a été proposé pour mesurer l’intérêt d’utiliser une mixture de modèles. Il s’agit d’un
modèle linéaire générique (filtre de Wiener avec post-processing markovien, MpWF)
dont les estimations sont post-traitées pour limiter des activations erronées durant
les périodes NC. Le second modèle a été choisi pour vérifier que le cadre discriminant
choisi pour développer le MSLM est plus performant que le cadre génératif. Un
filtre de Kalman commutatif (Switching Kalman Filter, SKF) a été choisi pour
mener cette comparaison entre modélisations discriminante et générative. Les SKFs
commutent entre plusieurs filtres de Kalman dont la pertinence est estimée à partir
de leur adéquation avec l’historique des signaux cérébraux.
Un ensemble d’indicateurs de performance a été formé pour évaluer la capacité
du MSLM, du MpWF et du SKF à réaliser du décodage asynchrone précis. La
qualité de la gestion des états NC est généralement mesurée au moyen d’indicateurs
basés sur la matrice de confusion. Il a ici été proposé d’additionellement considérer
des indicateurs de performance dits "par blocs", à savoir le nombre de blocs de fausses
activations et leur durée moyenne. Ces indicateurs par blocs prennent en compte
la dynamique des fausses activations, et sont susceptibles de davantage refléter la
qualité de la gestion des états de NC telle que perçue par l’utilisateur pendant
des sessions de contrôle neuronal asynchrone en boucle fermée. Les indicateurs de
performance associés à la reconstruction de paramètres cinématiques pendant les
périodes IC portent généralement sur leur précision spatiale et temporelle (Coefficient
de Corrélation de Pearson, distance euclidienne ou de Manhattan). De légères
désynchronisations entre les trajectoires observées et estimées, c’est-à-dire de petites
ix
erreurs temporelles, sont susceptibles d’avoir un impact important sur ces indicateurs
de performance. De ce fait, il est ici proposé de calculer ces indicateurs à la fois
entre les trajectoires mesurées et estimées originales (désynchronisées) et entre les
trajectoires synchronisées. Les indicateurs calculés sur les trajectoires synchronisées
se concentrent sur l’erreur spatiale de l’estimation de trajectoire. Une approche a
été mise au point pour calculer des trajectoires synchronisées dans l’échelle de temps
originelle de la trajectoire mesurée. Finalement, une procédure d’analyse statistique
a été choisie pour mesurer la significativité des différences de performance constatées
entre les trois décodeurs considérés.
Implémentation et résultats Des décodeurs MSLM, MpWF et SKF ont été im-
plémentés pour réaliser des reconstructions de trajectoires asynchrones mono-membre
et multi-membres sur les jeux de données précliniques et cliniques, respectivement.
Des représentations temps-fréquence-espace de grande dimension ont tout d’abord
été extraites des signaux cérébraux. Les paramètres de décodeurs avec 2 états (un
état NC et un état IC) ont été identifiés au moyen de procédures d’entraînement
supervisé pour les données mono-membre précliniques. Dans le cas de sessions
associées à des mouvements complexes, un MSLM avec deux états IC a également
été entraîné via l’algorithme d’apprentissage non-supervisé proposé dans la présente
thèse. Des décodeurs avec 6 états (un état NC et cinq états IC) ont été calibrés en
utilisant la procédure d’apprentissage supervisé pour les données multi-membres
(multi-doigts) cliniques. Des études préliminaires ont été réalisées pour optimiser les
composants du MSLM, c’est-à-dire ses modèles continus et le décodeur discret chargé
d’estimer la valeur de la variable latente permettant de combiner les estimations des
modèles continus. La méthode de réduction de dimension utilisée pour rendre le
SKF compatible avec du décodage temps réel a également été optimisée.
Le décodeur MSLM a permis d’obtenir une gestion des états NC de meilleure
qualité que le MpWF et SKF à la fois pour le décodage asynchrone mono-membre
et multi-membres. Une amélioration de la précision des estimations des paramètres
cinématiques a été constatée pour les données cliniques multi-membres (multi-doigts).
La précision de la reconstruction des mouvements complexes observés dans certaines
sessions d’acquisition pré-cliniques a été améliorée quand plusieurs modèles IC ont
été identifiés.
Discussion La pertinence du modèle proposé pour réaliser un décodage asyn-

chrone précis de mouvements mono-membre et multi-membres a été validée sur
des données précliniques et cliniques. Plusieurs pistes d’amélioration pourront être
explorées dans le futur, notamment l’utilisation d’un décodeur discret hiérarchique,
le développement d’une procédure d’apprentissage non-supervisé variationelle pour
une optimisation automatique du nombres de modèles du MSLM et celui d’une
méthode d’apprentissage récursif pour une adaptation en ligne des paramètres du
MSLM. Une étape de validation essentielle va également être menée à bien lors
de l’essai clinique de l’ICM motrice de CLINATEC "BCI et tétraplégie", à savoir
x
l’évaluation de la performance du MSLM lors de sessions de contrôle en boucle

fermée d’un exosquelette par un sujet tétraplégique.
xi
Abstract
Brain-Computer Interfaces (BCI) are systems that allow severely motor-impaired
patients to use their brain activity to control external devices, and thereby to interact
with their environment. Several processing procedures are usually carried out to
translate the user’s neuronal activity into commands for effector control. Features
specifically related to the user’s intentions are first extracted from measures of the
user’s brain activity. A decoder is then applied to estimate the user’s intention from
these brain features. After being optionally enhanced by post-processing algorithms,
this estimate of the user’s intention is converted into commands used to drive the
BCI effector(s), e.g. upper- or lower-limb orthoses or prostheses in the particular
case of motor BCIs.
The goal of CLINATEC’s motor BCI project is to bring the proof of concept that
it is feasible for a tetraplegic subject to control complex effectors, for example a 4-
limb exoskeleton, thanks to the monitoring and decoding of his electrocorticographic
(ECoG) brain activity. Challenges specific to the clinical deployment of CLINATEC’s
BCI system are addressed in the present doctoral thesis. A major issue for BCI
clinical applications is the ability to provide users with accurate asynchronous con-
trol over the effector. Unlike synchronous BCIs which are periodically controllable
by users, asynchronous BCI decoders are continuously available. In asynchronous
settings, the limitation of spurious effector activations during No-Control (NC)
periods is particularly desirable. A second challenge arises from the multi-limb
effector embedded into CLINATEC’s BCI system, namely the extension of accu-
rate asynchronous decoding to multi-limb control. A sequential upper-/lower-limb
activation strategy has been chosen for CLINATEC’s BCI system with the aim of
improving the system’s stability and of facilitating neural control. The corresponding
decoding objective lies in avoiding parallel, residual movements of the momentarily
non-controlled limbs. Finally, the ability of users to execute brain-controlled move-
ments is compromised when the estimates of the user’s intention are not sufficiently
correlated with his true intentions. Decoding accuracy during Intentional Control
(IC) periods is thus crucial.
Kinematic decoders, i.e. decoders which extract continuously-valued kinematic
parameters from the neural signals, are generally expected to be precise and intuitive
for motor BCIs. Recent studies have suggested they may be applicable for ECoG
decoding. In the present doctoral thesis, a switching decoder, namely a Markov
Switching Linear Model (MSLM), has been proposed for the task of accurate,
asynchronous sequential multi-limb kinematic ECoG decoding. The MSLM strategy
to support both NC/IC periods and multi-limb control consists of switching between
NC and (possibly multiple) IC linear models to take into account the specific
behaviours expected during NC and IC states. The relevance of each model, i.e. of
each state, is estimated from the neural data and is used to combine the available
continuous models. Switches between IC models are considered to address the
issue of sequential multi-limb effector control (one IC model per limb) as well as to
improve decoding accuracy during IC periods (one IC model per movement phase).
xii
The MSLM has been developed in the frame of static regression models, and
more precisely as an extension of Mixture of Experts (ME) models. A linear model
between neural features and kinematic parameters is conditioned on the current state.
A probabilistic rule is deduced from the likelihood of each possible state, and is used
to weight the available regression models. In contrast with the previously reported
BCI switching regression models, the MSLM performs dynamic state detection
to limit spurious effector activations. More specifically, the sequence of NC and
(possibly multiple) IC states is assumed to be generated by a first-order Markov
chain. A Hidden Markov Model-based (HMM) discrete decoder is used for state
estimation. Both supervised and unsupervised training procedures are presented in
the present dissertation. The proposed unsupervised training procedure is based
on the Expectation-Maximization (EM) algorithm, and extends both HMM- and
ME- well known EM training algorithms. Unsupervised training permits to exploit
internal states without precise a priori knowledge on their value in the training data
set. In particular, it can be used to associate different IC models with different
phases of complex movements.
The performance of the MSLM decoder was assessed for two decoding tasks,
namely asynchronous wrist trajectory reconstruction (publicly available preclinical
data set) and multi-limb (multi-finger) trajectory reconstruction (publicly available
clinical data set). The MSLM was compared to a Wiener Filter with Markovian
post-processing (MpWF) decoder and to a Switching Kalman Filter (SKF). These
two decoders were chosen so as to represent two alternative strategies previously
exploited for the considered tasks. The supervised MSLM decoder was found to
outperform both the SKF and the post-processed MpWF decoder in terms of state
detection accuracy and/or continuous decoding during IC states, thus reducing the
number of spurious activations during asynchronous mono- and multi-limb decoding
and/or improving decoding accuracy during IC periods. Evaluation of the MSLM
decoder will be performed in CLINATEC’s coming clinical trial.
xiii
Abbreviations and notations

Abbreviations
ACC Accuracy
ANN Artificial Neural Network
AP Action Potential
AR Auto-Regressive
AUC Area Under the Curve
CCWT Complex-Continuous Wavelet Transform
CV Cross-Validation
BCI Brain Computer Interface
CRF Conditional Random Field
CSP Common Spatial Pattern
CTPR Correct True Positive Rate
DBN Dynamic Bayesian Network
DoF Degree of Freedom
DTW Dynamic Time Warping
ECG Electrocardiographic, Electrocardiography, Electrocardiogram
ECoG Electrocorticographic, Electrocorticography
EEG Electroencephalographic, Electroencephalography
EM Expectation-Maximization
EOG Electrooculographic, Electrooculography, Electrooculogram
ERP Event-Related Potential
FN False Negative
FP False Positive
FPR False Positive Rate
FES Functional Electrical Stimulation
fMRI Functional Magnetic Resonance Imaging
fNIRS Functional Near-Infrared Spectroscopy
GAM Generalized Additive Model
GLM Generalized Linear Model
HMM Hidden Markov Model
IC Intentional Control
ICA Independent Component Analysis
IOHMM Input-Output Hidden Markov Model
KF Kalman Filter
kNN k-Nearest Neighbours
LASSO Least Absolute Shrinkage and Selection Operator
LDA Linear Discriminant Analysis
LFP Local Field Potential
LGF Laplace-Gaussian Filtering
LMP Local Motor Potential
LR Logistic Regression
xiv
M1 Primary Motor Cortex

MAP Maximum A Posteriori
ME Mixture of Experts
MEA Microelectrode Array
MEG Magnetoencephalographic, Magnetoencephalography
ML Maximum Likelihood
MLLS Minimal Length Least Squares
MpWF Markovian post-processed Wiener Filter
MRP Movement-Related Potential
MSLM Markov Switching Linear Model
MUA Multi Unit Activity
NC No-Control
NHP Non-Human Primate
NMAE Normalized Mean Absolute Error
NRMSE Normalized Root Mean Squared Error
OLS Ordinary Least Squares
PCA Principal Component Analysis
PCC Pearson Correlation Coefficient
PCR Principal Component Regression
PLS Partial Least Squares
PLSR Partial Least Squares Regression
PLV Phase-Locking-Value
PM Premotor Cortex
PPC Posterior Parietal Cortex
PRESS Predicted Error Sum of Squares
pWF Post-processed Wiener Filter
QDA Quadratic Discriminant Analysis
RBF Radial Basis Function
SCP Slow Cortical Potential
SCP Sensorimotor Rhythm
SL Surface Laplacian
SLM Switching Linear Model
SKF Switching Kalman Filter
SMA Supplementary Motor Area
SNR Signal-to-Noise Ratio
SSVEP Steady-State Visual Evoked Potential
STFT Short Time Fourier Transform
SUA Single Unit Activity
SVM Support-Vector-Machine
SVR Support-Vector-Regression
TN True Negative
TNR True Negative Rate
TP True Positive
TPR True Positive Rate
xv
UKF Unscented Kalman Filter

WF Wiener Filter
Mathematical Notations
Matrices are denoted by bold upper-case romans (e.g., X), vectors by bold lower-case
romans (e.g., x) and scalar by lower-cases italics (e.g., x). xij and xi refer to the
(i, j)th entry of matrix X and to the ith element of vector x, respectively.
Contents
1 Introduction 1
1.1 Brain-Computer Interfaces . . . . . . . . . . . . . . . . . . . . . . . . 1
1.1.1 BCI overview . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.1.2 Brain pattern elicitation . . . . . . . . . . . . . . . . . . . . . 2
1.1.3 BCI components . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.2 Motor BCIs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
1.2.1 Towards clinical motor BCIs . . . . . . . . . . . . . . . . . . 6
1.2.2 Signal acquisition . . . . . . . . . . . . . . . . . . . . . . . . . 9
1.2.3 Transducers for motor BCIs . . . . . . . . . . . . . . . . . . . 21
1.2.4 Effectors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
1.2.5 Feedback for prosthesis or orthosis control . . . . . . . . . . . 30
1.2.6 Summary on current progress . . . . . . . . . . . . . . . . . . 30
1.3 CLINATEC’s motor BCI project . . . . . . . . . . . . . . . . . . . . 31
1.3.2 Transducer . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
1.3.3 Effector . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
1.3.4 Feedback . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
1.4 PhD objectives . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
1.4.1 Asynchronous control . . . . . . . . . . . . . . . . . . . . . . 35
1.4.2 Multi-limb control . . . . . . . . . . . . . . . . . . . . . . . . 35
1.4.3 Accurate control . . . . . . . . . . . . . . . . . . . . . . . . . 35
2 Data-driven transducers for motor BCIs 37

2.1 Notations and processing blocks . . . . . . . . . . . . . . . . . . . . . 38
2.2 Pre-processing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
2.3 Feature extraction . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
2.3.1 Neural feature extraction . . . . . . . . . . . . . . . . . . . . 42
2.3.2 Features for effector control . . . . . . . . . . . . . . . . . . . 46
2.3.3 Dimensionality reduction . . . . . . . . . . . . . . . . . . . . 47
2.4 Decoders for motor BCIs . . . . . . . . . . . . . . . . . . . . . . . . . 49
2.4.1 Generalities on decoder training . . . . . . . . . . . . . . . . 49
2.4.2 Discrete decoding - classifiers . . . . . . . . . . . . . . . . . . 50
2.4.3 Continuous decoding . . . . . . . . . . . . . . . . . . . . . . . 58
2.5 Post-processing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
2.5.1 Discrete output . . . . . . . . . . . . . . . . . . . . . . . . . . 63
2.5.2 Continuous output . . . . . . . . . . . . . . . . . . . . . . . . 64
2.6 Transducers for accurate asynchronous kinematic decoding . . . . . 64
2.6.1 Asynchronous mono-limb decoding . . . . . . . . . . . . . . . 64
2.6.2 Asynchronous sequential multi-limb decoding . . . . . . . . . 66
xviii Contents
2.6.3 Accurate decoding during IC states . . . . . . . . . . . . . . . 66

2.7 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67
3 Switching models for accurate asynchronous trajectory decoding 69

3.1 Switching models for BCI control . . . . . . . . . . . . . . . . . . . . 69
3.1.1 Context-related modifications of kinematic encoding in neural
features . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
3.1.2 Switching models . . . . . . . . . . . . . . . . . . . . . . . . . 70
3.1.3 Mixture of Experts structure for discriminative switching models 72
3.2 Markov Switching Linear Models (MSLM) . . . . . . . . . . . . . . . 75
3.2.1 Dynamic gating . . . . . . . . . . . . . . . . . . . . . . . . . . 75
3.2.2 Experts . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78
3.3 MSLM training . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79
3.3.1 Supervised training . . . . . . . . . . . . . . . . . . . . . . . . 79
3.3.2 Unsupervised training . . . . . . . . . . . . . . . . . . . . . . 81
3.4 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86
4 Data set presentation 89

4.1 Asynchronous preclinical data set . . . . . . . . . . . . . . . . . . . . 89
4.1.1 Behavioural task . . . . . . . . . . . . . . . . . . . . . . . . . 89
4.1.3 Data set characteristics . . . . . . . . . . . . . . . . . . . . . 92
4.2 Asynchronous multi-limb (multi-finger) clinical data set . . . . . . . 95
4.2.2 Data acquisition . . . . . . . . . . . . . . . . . . . . . . . . . 95
4.3 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99
5 Evaluation and comparison 101

5.1 Comparative approaches for supervised MSLMs . . . . . . . . . . . . 101
5.1.1 Markovian post-processed Wiener Filter . . . . . . . . . . . . 102
5.1.2 Switching Kalman Filter . . . . . . . . . . . . . . . . . . . . . 103
5.1.3 Static decoders . . . . . . . . . . . . . . . . . . . . . . . . . . 104
5.2 Performance indicators . . . . . . . . . . . . . . . . . . . . . . . . . . 107
5.2.1 NC support for asynchronous mono- and multi-limb decoding 107
5.2.2 Multiple IC state discrimination for sequential multi-limb
asynchronous decoding . . . . . . . . . . . . . . . . . . . . . . 108
5.2.3 Performance indicators for continuous dependent variable . . 108
5.2.4 Trajectory synchronization . . . . . . . . . . . . . . . . . . . 108
5.3 Statistical tests . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 112
5.4 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 112
Contents xix
6 Implementation 115
6.1 Feature extraction and pre-processing . . . . . . . . . . . . . . . . . 115
6.1.1 Kinematic parameter extraction . . . . . . . . . . . . . . . . 115
6.1.2 Neural signal feature extraction . . . . . . . . . . . . . . . . . 116
6.2 Decoder implementation . . . . . . . . . . . . . . . . . . . . . . . . . 117
6.2.1 Preclinical data set . . . . . . . . . . . . . . . . . . . . . . . . 117
6.2.2 Clinical data set . . . . . . . . . . . . . . . . . . . . . . . . . 123
6.3 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 125
7 Results 127
7.1 Preclinical data set . . . . . . . . . . . . . . . . . . . . . . . . . . . . 127
7.1.1 Integration of NC support into kinematic decoders . . . . . . 127
7.1.2 Multiple IC experts for kinematic reconstruction of complex
movements . . . . . . . . . . . . . . . . . . . . . . . . . . . . 142
7.2 Clinical data set . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 145
7.3 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 159
7.3.1 Mono-limb decoding . . . . . . . . . . . . . . . . . . . . . . . 160
7.3.2 Multi-limb decoding . . . . . . . . . . . . . . . . . . . . . . . 163
7.3.3 Absolute decoding performance . . . . . . . . . . . . . . . . . 164
7.4 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 165
8 Discussion and perspectives 167

8.1 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 167
8.1.1 Challenges for motor BCI clinical applications . . . . . . . . 167
8.1.2 Proposed methodology . . . . . . . . . . . . . . . . . . . . . . 169
8.1.3 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 171
8.1.4 Limits of the present work . . . . . . . . . . . . . . . . . . . . 173
8.2 Perspectives . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 174
8.2.1 Technical optimizations . . . . . . . . . . . . . . . . . . . . . 174
8.2.2 Closed-loop evaluation during CLINATEC’s clinical trial . . . 176
A Supplementary materials for MSLM training 179

A.1 Supervised training: decoupling of parameter estimation . . . . . . . 179
A.2 Unsupervised training: E-step . . . . . . . . . . . . . . . . . . . . . . 182
B MSLM: expert training procedure and gate selection 187

B.1 Expert training procedure selection . . . . . . . . . . . . . . . . . . . 187
B.1.1 Methods and implementation . . . . . . . . . . . . . . . . . . 188
B.1.2 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 189
B.1.3 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 192
B.2 Gate selection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 193
B.2.2 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 196
B.2.3 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 199
xx Contents
C Switching Kalman Filtering for asynchronous ECoG decoding 201

C.1 Introduction: Kalman filter . . . . . . . . . . . . . . . . . . . . . . . 201
C.2 Switching Kalman Filter . . . . . . . . . . . . . . . . . . . . . . . . . 202
C.2.1 Switching State-Space Model . . . . . . . . . . . . . . . . . . 202
C.2.2 Training . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 203
C.2.3 Application . . . . . . . . . . . . . . . . . . . . . . . . . . . . 203
C.3 Optimization of Kalman Filtering for high-dimensional ECoG signals 206
C.3.1 Methods and implementation . . . . . . . . . . . . . . . . . . 207
C.3.2 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 208
C.3.3 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 208
Bibliography 211
Chapter 1
Introduction
Contents
1.1 Brain-Computer Interfaces . . . . . . . . . . . . . . . . . . . . 1
1.1.1 BCI overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.1.2 Brain pattern elicitation . . . . . . . . . . . . . . . . . . . . . 2
1.1.3 BCI components . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.2 Motor BCIs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
1.2.1 Towards clinical motor BCIs . . . . . . . . . . . . . . . . . . 6
1.2.3 Transducers for motor BCIs . . . . . . . . . . . . . . . . . . . . 21
1.2.4 Effectors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
1.2.5 Feedback for prosthesis or orthosis control . . . . . . . . . . . 30
1.2.6 Summary on current progress . . . . . . . . . . . . . . . . . . 30
1.3 CLINATEC’s motor BCI project . . . . . . . . . . . . . . . . 31
1.3.2 Transducer . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
1.3.3 Effector . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
1.3.4 Feedback . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
1.4 PhD objectives . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
1.4.1 Asynchronous control . . . . . . . . . . . . . . . . . . . . . . 35
1.4.2 Multi-limb control . . . . . . . . . . . . . . . . . . . . . . . . 35
1.4.3 Accurate control . . . . . . . . . . . . . . . . . . . . . . . . . 35
1.1 Brain-Computer Interfaces

1.1.1 BCI overview
Brain-Computer Interfaces (BCI), also referred to as Brain-Machine Interfaces or
Direct Brain Interfaces [Graimann et al., 2009], are systems which permit users
to utilize their brain activity to control external devices without using their nat-
ural neuromuscular pathways [Leuthardt et al., 2006b] [Mak and Wolpaw, 2009].
BCIs are particularly being investigated for the sake of severely motor-impaired
patients. Common causes for serious motor dysfuntions are neuromuscular disorders
2 Chapter 1. Introduction
like amyotrophic lateral sclerosis or cerebral palsy, spinal cord injury (paraplegia,
tetraplegia) and stroke [Lebedev and Nicolelis, 2006]. The independence and well-
being of severely motor-impaired patients are threatened by their inability to control
muscles required for the execution of critical motor tasks, for example bowel control,
respiration, limb movements or verbal communication. BCIs aim at overcoming
some of these disabilities by establishing a new communication pathway between the
patient’s brain and an effector (e.g., a robotic arm, a speller or a wheelchair). Brain
effector control [Mak and Wolpaw, 2009] is expected to help patients recovering the
ability to interact with their environment. It is also anticipated that neurorehabili-
tation programs will benefit from the introduction of BCI-based approaches [Lew,
2012].
1.1.2 Brain pattern elicitation

The input of a BCI system is a measure of the user’s brain activity. BCIs are
based on the interpretation of brain activity patterns. Specific and measurable brain
activity patterns must thus be generated to trigger the execution of a particular
action by the effector. Different strategies are used to elicit this recognizable brain
activity.
A distinction is drawn between externally- and internally-paced brain patterns

[Waldert et al., 2009]. Externally-paced brain patterns are responses evoked by
a visual, auditory or somatosensory stimulus (Evoked Potential) or by an event
(Event-Related Potential, ERP). Steady-State Visual Evoked Potentials (SSVEP)
and the P300 wave are two examples of externally-paced brain patterns regularly
exploited for BCI control [Waldert et al., 2009]. The use of SSVEPs has for example
been reported for spelling devices [Nakanishi et al., 2014a] [Yin et al., 2015], hand
orthosis [Ortner et al., 2011] and wheelchair control [Müller et al., 2011] [Nguyen
et al., 2013]. It is elicited when the user looks at a visual stimulus which flickers at a
specific frequency. The stimulation frequency and its harmonics are observed in an
oscillation mainly measured over the user’s occipital brain area [Amiri et al., 2013].
A common BCI protocol thus consists of presenting the user with several targets
flashing at a different frequency [Amiri et al., 2013]. Each target is associated with
one action over the BCI effector, e.g. a specific wheelchair movement, or a letter
of interest when the BCI is used for communication purposes. A target-specific
SSVEP is elicited when the user directs his gaze towards the target of interest. A
different protocol is utilized to elicit the P300 wave, which is typically exploited for
communication BCIs (e.g., [Sellers et al., 2006] [Krusienski et al., 2006] [Kindermans
et al., 2012] but has also permitted users to control wheelchairs [Rebsamen et al.,
2006] and prostheses [Palankar et al., 2008]. This ERP systematically occurs 300ms
after presentation of a rare but expected visual stimulus [Waldert et al., 2009]. To
exploit the P300 wave, a grid filled in with letters is displayed on a screen. Users
are asked to focus on the letter they wish to spell, and the grid columns and rows
are successively highlighted [Amiri et al., 2013]. The P300 wave is elicited when a
1.1. Brain-Computer Interfaces 3
column or row containing the letter of interest is highlighted [Amiri et al., 2013].
By contrast, internally-paced BCIs rely on brain patterns voluntarily elicited
by users, e.g. Slow-Cortical Potentials (SCP) and Sensorimotor Rhythms (SMR)
[Waldert et al., 2009]. SCPs are slow potentials centred on the upper surface of the
brain, while SMRs are rhythms generated in the motor cortex, i.e. in brain areas
particularly devoted to motor control. Voluntarily modulation of SCPs and SMRs
is possible after training [Waldert et al., 2009], thus enabling users to control, for
example, prosthetic devices [Pfurtscheller et al., 2000] [Wang et al., 2013c], cursors
[Wolpaw and McFarland, 2004] or robots [Chae et al., 2012].
Brain pattern elicitation sometimes requires the execution of movements of

unimpaired body parts [Amiri et al., 2013], e.g. head or eye movements [Mak and
Wolpaw, 2009]. The corresponding BCIs are referred to as "dependent", in contrast
with independent BCIs which don’t require the ability to contract particular muscles
[Wolpaw et al., 2002] [Mak and Wolpaw, 2009]. For example, the elicitation of
visual evoked potential may require (possibly imperceptible) eye movements of the
user towards the visual stimuli [Mak and Wolpaw, 2009]. The applicability of such
dependent BCIs to severely motor-impaired patients with poor control over muscles
is limited [Wolpaw et al., 2002].
1.1.3 BCI components

Several components are necessary to translate the user’s cerebral activity into effec-
tor actions. The cerebral signal acquisition system, the transducer, the controlled
effector and the feedback provided to the user are the main components of a BCI
system [Schwartz et al., 2006] (Figure 1.1).
Figure 1.1: BCI main components: acquisition system, transducer, effector and
feedback.
Signal acquisition The acquisition system is used to sample, amplify and digitize
a measure of the user’s cerebral activity [Homer and Nurmikko, 2013] as reflected
by electrophysiological, magnetic or metabolic signals [Mak and Wolpaw, 2009].
Electrophysiological signals originate from the electrical currents generated by
neurons. They are acquired by means of microelectrodes arrays (MEA), electrocor-
ticographic (ECoG) or electroencephalographic (EEG) arrays [Lebedev and Nicolelis,

2006]. These acquisition systems measure electrical fields at different distances from
their sources, and therefore exhibit different degrees of invasiveness and spatial
resolutions. MEAs are invasive arrays which directly sample neurons’ electrical
activity from within the brain (intracortical recordings) [Homer and Nurmikko,
2013]. Electrocorticographic (ECoG) arrays acquire the cerebral activity at the
surface of the brain [Mak and Wolpaw, 2009]. In contrast with MEAs, ECoG arrays
are said to be semi-invasive [Rak et al., 2012]. Finally, electroencephalographic
(EEG) arrays measure the neural signals from the surface of the skull [Teplan, 2002].
EEG-based acquisition systems are therefore non-invasive [Rak et al., 2012].
Magnetoencephalographic (MEG) sensors sample the magnetic field generated
by the brain activity at a few centimeters above the skull [Buzsáki et al., 2012].
The cerebral activity is also reflected by the blood oxygenation level in the brain,
which is typically measured via fMRI (functional Magnetic Resonance Imaging) or
fNIRS (functional Near-Infrared Spectroscopy) [Mak and Wolpaw, 2009].
Current MEG and fMRI acquisition systems suffer from high bulkiness, high
cost and/or poor temporal resolution [Mak and Wolpaw, 2009] [Nicolas-Alonso and
Gomez-Gil, 2012]. fNIRS acquisition systems, on the other hand, hold promise for
BCI applications because of their non-invasiveness, portability and reasonable price
[Mak and Wolpaw, 2009]. The feasibility of fNIRS decoding has been demonstrated
in recent studies [Coyle et al., 2007] [Naseer et al., 2014] [Hong et al., 2015]. Offline
distinction between 3 intentions was achieved in [Hong et al., 2015]; binary cursor
control was reported in [Coyle et al., 2007] and [Naseer et al., 2014]. As 3D
EEG-based neural control was provided to users in several studies [LaFleur et al.,
2013], [McFarland et al., 2010], the performance of fNIRS-driven BCIs is not yet
comparable with the performance of EEG-based BCIs. Consequently, BCIs mostly
rely on electrophysiological sources of control [Mak and Wolpaw, 2009].
Transducer The BCI transducer translates brain activity measurements into

estimates of the user’s intention. It is generally composed of several signal process-
ing blocks. A first, optional step consists of enhancing the raw cerebral signals.
Algorithmical preprocessing methods permit to improve the Signal-to-Noise Ratio,
and/or to remove (at least partially) artifacts [Bashashati et al., 2007a]. Features
specifically related to the user’s intentions are then extracted from the cerebral sig-
nals [Mak and Wolpaw, 2009]. A decoder, also referred to as "translation algorithm"
[Yuan and He, 2014] or "feature translator" [Bashashati et al., 2007a], interprets
the brain features and issues an estimate of the user’s intention. Construction of
the decoder is required before utilization of the BCI system. In other words, a
model of dependence between neuronal patterns and associated intended actions
has to be established [Homer and Nurmikko, 2013]. The hypotheses used to build
a decoder depend on the problem at hand (signal quality and resolution, effector
etc.). Most BCI systems rely on user-specific decoders, i.e. on decoders whose
parameters are adapted for each user [Wolpaw et al., 2002]. This is typically done
1.2. Motor BCIs 5
by applying model identification algorithms on a data set that contains samples of

cerebral signals along with the corresponding intentions of the subject. After being
optionally enhanced by post-processing methods [Bashashati et al., 2007a], intention
estimates are conveyed to the effector’s controller.
Effector control The effector’s controller converts intention estimates into com-
mands which are sent to the effector. The type of effector integrated into a BCI
system depends on the objective of the BCI. BCIs mostly aim at providing patients
with the ability to communicate, exert control over their environment, displace
themselves, or recover some motor control over their limbs [Mak and Wolpaw, 2009].
Beside the traditional spellers, BCIs for communication often give users control over
a cursor displayed on a computer screen, so that they can write texts, select icons,
surf the Internet etc. BCIs for environmental control also rely on cursor control for
the adjustment of environmental variables [Mak and Wolpaw, 2009], e.g. light or
position of a motorized bed [Cincotti et al., 2008]. Navigation (locomotion) BCIs
provide users with control over wheelchairs [Mak and Wolpaw, 2009]. Finally, motor
BCIs aim at restoring limb mobility. They generally embed upper- or lower-limb
orthoses [Yuan and He, 2014] or prostheses. Considerable efforts are put into the
restoration of upper-limb mobility, in particular for the execution of reaching move-
ment with a brain-controlled robotic arm (e.g., [Wodlinger et al., 2015] [Aflalo et al.,
2015]).
Feedback Natural volitional motor control is permitted by the perception and

exploitation of different types of feedback [Suminski et al., 2010], e.g. proprioceptive,
visual, auditory or tactile feedback. BCI systems consequently supply users with
so-called closed-loop control over the effector, i.e. feedback is regularly delivered to
users. Most BCI systems exclusively provide users with visual feeback on the effector
state [Hochberg et al., 2012] [Collinger et al., 2013] [Wodlinger et al., 2015]. The
potential benefits of alternative feedback types have nevertheless drawn the attention
of several teams [Cincotti et al., 2007] [Wilson et al., 2012] [McCreadie et al., 2014]
[Perruchoud et al., 2016]. Auditory feedback was provided to users in [McCreadie
et al., 2014]; sensory feedback was delivered through vibrotactile or electrotactile
stimulation on users’ skin or tongue in [Cincotti et al., 2007] and [Wilson et al., 2012].
The utilization of more complex schemes like direct intracortical brain stimulation
[O’Doherty et al., 2011] has additionally been reported [Lebedev and Nicolelis, 2006]
[Flesher et al., 2016].
1.2 Motor BCIs

The present dissertation reflects work conducted within the framework of motor
BCIs, i.e. BCIs providing users with control over an orthosis or prosthesis [Mak and
Wolpaw, 2009]. Such BCIs raise hopes that limb mobility may be restored in severely
impaired patients, for example patients with tetraplegia. Specific technical challenges
are associated with the control of physical orthoses or prostheses. The present section
gathers a presentation of these challenges and of the technical solutions which have
been proposed to address them.
1.2.1 Towards clinical motor BCIs

The following set of technical features is specially considered when designing a motor
BCI.
Safety The safety of the signal acquisition system is critical for BCI applications
[Leuthardt et al., 2006b]. Moreover, in contrast with communication BCIs and their
virtual effectors, motor BCIs generally involve interactions between the user and a
physical effector. Specific risks result from this physical interaction, for examples
falls or the execution of anatomically impossible movements. As they jeopardize the
safety of the BCI users, it is crucial to avert them.
Chronicity Long-term validity is commonly aimed at during the development

of a BCI system [Lebedev and Nicolelis, 2006] [Carmena, 2013] [Leuthardt et al.,
2006b]. It implies the chronic acquisition of brain signals (i.e., over several decades).
Unavoidable signal instabilities, for example due to the user’s fatigue, are one of the
challenges associated with chronic BCIs [Wolpaw et al., 2002].
Decoding accuracy BCI systems benefit patients if the decoded actions reflect
the user’s intentions with an accuracy enabling him to interact with his environment
[Lebedev and Nicolelis, 2006] [Leuthardt et al., 2006b] [Marathe and Taylor, 2011]
[Marathe and Taylor, 2015] (see examples in Figure 1.2).
Multi-limb control Full independence of severely motor-impaired patients, e.g.

patients with tetraplegia, is possible if they are provided with control over both
lower- and upper-limb protheses or orthoses.
Degrees of Freedom Upper- and lower-limb prostheses have multiple controllable

Degrees of Freedom (DoF). 3D control of the endpoint of an upper-limb prothesis or
orthosis is a frequent objective of motor BCIs [Lebedev and Nicolelis, 2006].
Robustness Most BCI studies are currently completed in controlled laboratory

environments. Robustness of the transducer, i.e. stability of decoding accuracy in
noisy environments, is necessary for the practical BCIs [Leuthardt et al., 2006b].
Feedback It is crucial for BCI users to be regularly provided with feedback on the
effector state. Specific feedback characteristics have been shown to have a dramatic
impact on neural control performances. It has for example been demonstrated
that using several feedback modalities, e.g. visual and proprioceptive feedback,
can improve the quality of neural control [Suminski et al., 2010]. The ability to
1.2. Motor BCIs 7
Figure 1.2: Ability to execute reaching movements depending on the correlation R

between intended and decoded trajectory parameters (reproduced from [Marathe
and Taylor, 2011]). Decoded position or velocity parameters were used to control
either the position or the velocity of an effector. It was necessary for subjects to
learn a position-to-velocity map when decoded position parameters were used to
control the velocity of the effector. (1): Examples of reaching trajectories. (2):
Average duration of a reaching movement. (3): Percentage of missed targets. A.
Position decoding, position control. B. Position decoding, velocity control. C.
Velocity decoding, velocity control.
exert control over a motor BCI has also been found to be affected by unnatural,
important latencies between the user’s intention and the corresponding effector
reaction [Willett et al., 2013][Marathe and Taylor, 2015]. The duration required for
the execution of reaching tasks has been found to be increased by around 180ms for
every 100ms-long pure delay introduced into the transducer [Willett et al., 2013].
Similar results were reported in [Marathe and Taylor, 2015], where the introduction
of a 300ms-long delay into the transducer increased the duration required to execute
reaching movements by more than 300ms. An important latency also limits brain
plasticity, i.e. the ability of the user to adapt to the BCI transducer [Leuthardt
et al., 2006b]. BCI developers therefore endeavour to limit the latency between
movement intention and execution of this movement by the effector.
Real-time decoding update The effector controller is regularly fed with updated
estimates of the user’s intention. The suitability of update frequencies depends on
the effector [Leuthardt et al., 2006b]. In contrast with communication BCIs which
can be highly beneficial to patients even when they exhibit comparatively low update
rates, motor BCIs require a high update rate to profit users. Most BCI systems
rely on a frequency rate of 10 to 20Hz [Kim et al., 2006a] for cursor or prosthetic
control in interaction-free environments. Lower frequency updates, e.g. 5Hz for arm
prosthesis control [Yanagisawa et al., 2012a], were nevertheless reported. It has been
suggested that an update rate of 20Hz is necessary for object manipulation [Kim
et al., 2006a]. The update frequency defines the maximal duration the transducer
can use to translate brain signals into user intention’s estimates. The computational
complexity of the transducer’s processing blocks is consequently constrained by the
update frequency and the hardware processing power.
Asynchronicity Most BCI demonstrations are conducted using synchronous, cue-

paced control paradigms. User intentions are not processed outside predefined, cued
windows [Graimann et al., 2009]. By contrast, an asynchronous or self-paced BCI
decoder is continuously available to the user [Graimann et al., 2009]. Potential BCI
users express their desire for stand-alone BCI systems, i.e. systems which utilization
does not require the presence of a technician [Blabe et al., 2015]. Asynchronicity is
thus a technical feature essential for practical motor BCIs [Wolpaw et al., 2002].
Mental load It is desirable for neural control to be associated with a light mental
load [Yuan and He, 2014] [Mak and Wolpaw, 2009].
Embeddability Bulky hardware components for signal acquisition, processing

or effector control are not practical for motor BCIs. Embeddable components are
therefore developed for motor BCI applications. Because of this constraint, cum-
bersome acquisition systems like MEG are currently deemed unfit for motor BCIs
[McFarland and Wolpaw, 2011].
1.2. Motor BCIs 9
These challenges drive the technical choices made when designing the components
of motor BCIs.
1.2.2 Signal acquisition

The signal acquisition system is a crucial component of a BCI system. Its invasiveness
determines the chronicity and safety of the system. Additionally, the control of
several limbs with multiple degrees of freedom requires a highly informative content
to be extracted from the acquired signals. The spatial and spectral characteristics
of the acquired signals depend on the acquisition system, and directly impact the
feasibility of accurate multi-limb, multi-DoF (Degrees of Freedom) effector control.
To date, motor BCIs mostly rely on electrophysiological sources of control [Mak and
Wolpaw, 2009], as opposed to magnetic and metabolic sources.
1.2.2.1 Cortical patterns for prosthetic control
Electrophysiologial signal acquisition systems mostly give access to the cerebral

activity generated in the cortex [Buzsáki et al., 2012], i.e. in the 2 − 4mm thick
surface of the brain (Figure 1.3.A) [Buzsáki et al., 2012].
Cortical electro-physiological activity Cortical electrophysiological activities

are generated by nerve cells called "neurons". Typical neurons are composed of
a body and of two types of extensions, namely one axon and a variable number
of dendrites [Squire et al., 2013] (Figure 1.3.C). Neurons are interconnected via
their axon and dendrites, i.e. they are organized in networks. They are specialized
in the reception, processing, and transmission of information encoded in electrical
signals [Kandel et al., 2000]. Neurons have the capacity to generate a transient
impulse called Action Potential (AP) in response to a simulus [Kandel et al., 2000].
The AP propagates inside the neuron and along the axon [Kandel et al., 2000].
At the axon’s extremity, it is converted into chemical transmitters released in the
extracellular environment, where they are captured by a cell, e.g. by the dendrites
of a neighbouring neuron. The generation of APs is an all-or-nothing phenomenon:
APs’ amplitude and shape don’t depend on the stimulus, and APs only occur if the
stimulus is superior to a specific threshold [Kandel et al., 2000]. In other words,
neurons units of communication are Boolean variables [Kandel et al., 2000]. Neurons
are said to "spike" [Homer and Nurmikko, 2013] or to "fire". Information is thus
exclusively encoded in the circulation of Boolean variables in neural networks [Kandel
et al., 2000].
The cortical neural network is composed of 6 layers of aligned and interconnected
neurons, referred to as "pyramidal neurons" (Figure 1.3.B). The pyramidal neurons
also communicate with subcortical areas via their axon [Kandel et al., 2000]. Due to
the alignment of the pyramidal neurons and to spiking synchronization phenomena,
the joint activity of cortical neurons results in modulations (rhythms) observed in
specific frequency bands. The following division of brain signals into bandwidths is
often used to describe neural population activity [Kandel et al., 2000] [Morshed and
Khan, 2014]:
• Delta: 0.5Hz < fδ < 4Hz
• Theta: 4Hz < fθ < 7Hz
• Mu: 8Hz < fµ < 13Hz (over the motor cortex)
• Beta: 13Hz < fβ < 30Hz
• Gamma: 30Hz < fγ
Figure 1.3: Cerebral cortex. A. Cerebral cortex (represented in dark purple)

[BrainMaps, 2017]. B. Pyramidal neurons [Gray, 1918]. C. Pyramidal neuron
[Kandel et al., 2000].
Functional organization Most motor BCIs exploit neural activities generated

in localized areas of the cortex, from a few mm2 [Collinger et al., 2013] [Wodlinger
et al., 2015] [Kellis et al., 2012] to few cm2 [Wolpaw and McFarland, 2004] [Baxter
et al., 2013] [Wang et al., 2013c]. The brain presents a functional organization,
i.e. spatial areas are specialized in specific functions. This functional organization
permits to select areas relevant for neural control in motor BCIs.
The anatomical division of the brain into lobes (Figure 1.4.A) is often used to
facilitate the description of its functional organization [Kandel et al., 2000]. Auditory
stimuli are mainly managed in the temporal lobe [Kandel et al., 2000]. The occipital
lobe receives and processes visual information sent by the eyes [Kandel et al., 2000].
The parietal lobe contains areas involved in space perception and processing of
sensori information [Kandel et al., 2000]. The frontal lobe is in particular in charge
of cognitive functions (thought, learning, memory, etc.) and motor skills (including
speech) [Kandel et al., 2000].
The reported motor BCIs have exploited sources of control related to different
cortex areas. Neural activity in the temporal lobe enabled users to control a cursor
1.2. Motor BCIs 11
Figure 1.4: A. Brain’s lobes [Kandel et al., 2000]). B. Cortical areas involved in
motor control (red lines) [Kandel et al., 2000]. C. Somatotopic organization of the
primary motor cortex (reproduced from [Kandel et al., 2000]).
using their speech network in [Leuthardt et al., 2011]. Users were provided with
neural control over a robotic arm by means of visual evoked potentials measured
over the occipital lobe in [Ortner et al., 2011]. A robotic device and a cursor were
controlled by subjects’ occipital activity in [Ferreira et al., 2008] and [Trejo et al.,
2006], respectively. Finally, the majority of reported motor BCIs made use of activity
elicited in the frontal and parietal lobes, and more specifically in the motor cortex
[Collinger et al., 2013] [Wodlinger et al., 2015] [Kellis et al., 2012] [Wolpaw and
McFarland, 2004] [Baxter et al., 2013] [Wang et al., 2013c].
The motor cortex is devoted to the organization of movements, and in particular
of voluntary movements [Kandel et al., 2000]. It is composed of the primary motor
area (M1) and of premotor areas, in particular the premotor cortex (PM), the
supplementary motor area (SMA) and the Posterior Parietal Cortex (PPC) [Kandel
et al., 2000] (Figure 1.4.B). Present knowledge about neuronal management of
motor control remains limited [Lebedev and Nicolelis, 2006]. Some features of
neural control have nevertheless been characterized. M1 neurons encode both low-
level (e.g., muscle force) and high-level information about desired movements (e.g.,
direction of arm movements [Kandel et al., 2000]). The amount of neurons devoted
to one limb or organ is not proportional to the organ size, but to the complexity
of its control (somatotopic organization, illustrated in Figure 1.4.C). This spatial
organization is exploited by numerous BCIs. Additionally, motor control is mainly
contralateral, i.e. cortical neurons from left hemisphere are connected to motor
neurons managing the right size of the body [Kandel et al., 2000]. Premotor areas
are involved, e.g., in motor planning and limb coordination [Kandel et al., 2000].
Although the functions managed by the SMA have not yet been clearly identified, it
has been suggested that it is involved in the control of voluntary movements and
of sequences of movements [Kandel et al., 2000]. The Posterior Parietal Cortex
manages in particular sensory information for motor planning [Kandel et al., 2000].
Whereas most BCIs indistinguishably exploit activity generated in various areas of
the motor cortex, the specific use of M1 activity has been reported in motor BCIs
[Velliste et al., 2008] [Bouton et al., 2016] [Hochberg et al., 2012]. The feasibility of
PPC-based neural control was additionally investigated in a recent study [Aflalo
et al., 2015], and bimanual movements were controlled using M1, SMA, PPC and
primary somatosensory activity in [Ifft et al., 2013].
Electro-physiological sources of control for motor BCIs Most reported

motor BCIs were based on the elicitation of internally-paced patterns [Collinger
et al., 2013] [Wodlinger et al., 2015] [Kellis et al., 2012] [Pfurtscheller et al., 2000]
[Wolpaw and McFarland, 2004] [Baxter et al., 2013] [Wang et al., 2013c]. The use of
externally-paced patterns (Evoked and Event-Related Potentials) has nevertheless
been reported in several motor studies [Palankar et al., 2008] [Bell et al., 2008]
[Ortner et al., 2011].
Cortical neurons The voluntary modulation of the firing rate of neurons in

the motor cortex has been exploited in most recent motor intracortical BCI systems
[Velliste et al., 2008] [Hochberg et al., 2012] [Collinger et al., 2013] [Wodlinger
et al., 2015] [Bouton et al., 2016]. Neural control was generally based on the spiking
activity measured in the primary motor cortex [Velliste et al., 2008] [Bouton et al.,
2016] associated to hand [Hochberg et al., 2012] or arm movements. The bimanual
movements reported in [Ifft et al., 2013] relied on a larger motor cortex area, namely
on M1, SMA, PPC and primary somatosensory activity. A preliminary study has
also demonstrated the possibility to decode users’ intentions from PPC activity
[Aflalo et al., 2015].
Sensorimotor rhythms Most semi- and non-invasive motor BCIs are based
on the exploitation of sensorimotor rhythms [Pfurtscheller et al., 2000] [Wolpaw and
McFarland, 2004] [McFarland et al., 2010] [LaFleur et al., 2013] [Baxter et al., 2013]
[Milekovic et al., 2012] [Kapeller et al., 2015] [Hotson et al., 2016]. Sensorimotor
rhythms (SMR) are patterns generated by neural populations of the sensorimotor
cortex [Leuthardt et al., 2006b]. Although SMRs are not well understood [Yuan and
He, 2014], some properties of these rhythms have been characterized.
Modulations synchronized with motor tasks have been observed in µ, β and γ
frequency bands (Figure 1.5). Two seconds before the onset of a movement, an
amplitude decrease is observed in µ and β bands (Event-Related Desynchronization).
This phenomenon is mostly contralateral to the movement, and is spatially consistent
with the somatotopic organization of the motor cortex. Simultaneously, an amplitude
increase is observed in the γ band. This rhythm is more spatially focused than µ
and β modulation [Schalk and Leuthardt, 2011]. After the movement, an amplitude
increase occurs in µ and β frequency bands (Event-Related Synchronization). These
rhythms can be observed in the contralateral sensorimotor area, with a peak 600ms
after the movement offset. Although they have been initially observed during
planning and execution of real movements, it has been shown that they also occur
during motor imagery but with a smaller amplitude. Volitional control of SMRs
1.2. Motor BCIs 13
is thus possible after training, and the simplest way to elicit it is to use motor
imagination [Leuthardt et al., 2006b] [Waldert et al., 2009].
Similarly, Movement-Related Potentials (MRPs) are synchronized with move-
ments. These low-frequency potentials appear 1 to 1.5s before a movement [Waldert
et al., 2009] [Bashashati et al., 2007a]. Their distribution is bilateral at the beginning
of the movement, and becomes contralateral when close to the movement [Bashashati
et al., 2007a]. Readiness Potential and Lateralized Readiness Potential are examples
of MRPs. Volitional control of MRPs can be achieved, e.g. via motor imagination
[Nazarpour et al., 2009].
Figure 1.5: SMRs (reproduced and modified from [Waldert et al., 2009])
SMRs are widely exploited in motor BCIs, e.g. for prosthesis [Pfurtscheller et al.,
2000] [Yanagisawa et al., 2012a] [Baxter et al., 2013] [Wang et al., 2013c] or virtual
effector control [Wolpaw and McFarland, 2004] [McFarland et al., 2010] [Doud et al.,
2011] [Kellis et al., 2012].
SCP Slow Cortical Potentials (SCPs) are slow potentials centered on the upper
surface of the brain [Leuthardt et al., 2006b] [Bashashati et al., 2007a]. SCP are
not necessarily related to motor control or to a mental specific task, but volitional
modulation of SCPs is possible after training [Leuthardt et al., 2006b] [Waldert et al.,
2009]. SCP-based neural control has been used for spelling systems [Birbaumer
et al., 1999] and can be used for motor BCIs.
Response to mental tasks Brain pattern elicitation via mental tasks has
been performed in a few cursor-control studies [Penny et al., 2000] [Vansteensel et al.,
2010]. Discrimination between responses to mental tasks has also been investigated
in feasibility studies, e.g. in [Anderson and Sijercic, 1996] [Ramsey et al., 2006], or
in [Chiappa and Bengio, 2003] where both mental and motor tasks were classified
during offline analyses. Reported mental tasks consisted in solving equations [Ramsey
et al., 2006], multiplying [Anderson and Sijercic, 1996] or substracting [Penny et al.,
2000] [Vansteensel et al., 2010] numbers, writing a letter [Anderson and Sijercic,
1996], finding words beginning with a specific letter [Chiappa and Bengio, 2003],
etc. The corresponding brain patterns were for example observed in the prefrontal
cortex [Ramsey et al., 2006] [Vansteensel et al., 2010], or in the parietal, frontal and
occipital cortex [Anderson and Sijercic, 1996].
Event-related potentials A few teams have reported P300-based neural

control over robotic effectors, namely a robotic arm [Palankar et al., 2008] and a
humanoid robot [Bell et al., 2008].
Evoked Potentials While evoked potentials have rarely been exploited for
motor BCIs, SSVEP-based hand orthosis control has for example been achieved in
[Ortner et al., 2011].
Other strategies Several motor BCIs have made use of several sources of
control [Horki et al., 2011] [Pfurtscheller et al., 2010], e.g. hybrid SSVEP- and
SMR-based hand prosthesis control in [Pfurtscheller et al., 2010] and [Horki et al.,
2011].
Measuring cortical patterns The signals measured by electrophysiological ac-

quisition systems, namely Microelectrode Arrays (MEA), Electrocorticographic
(ECoG) and Electroencephalographic (EEG) acquisition systems, mainly correspond
to the extracellular currents generated by the cortical neurons [Buzsáki et al., 2012].
Depending on the invasiveness of the acquisition system, sensors are located at
a distance which ranges from a few µm (MEA) to several cm (EEG) from the
cortical neurons generating the extracellular currents of interest [Waldert et al.,
2009] (Figure 1.6). This distance impacts the size of the neuronal population
observed by sensors. When signals are measured in the vicinity of neurons, they
mostly reflect individual neural activity. By contrast, remote electrodes acquire
the electric field generated by a large neuronal population, i.e. they measure the
superposition of extracellular currents. The individual (spatial, spectral) character-
istics of neuronal patterns are lost because of this spatial averaging. As the current
amplitude depends on the inverse of the sensor-neuron distance [Buzsáki et al.,
2012], this distance additionally impacts the Signal-to-Noise Ratio. High frequencies
are particularly hard to observe since current amplitudes are also approximately
inversely proportional to their frequency [Buzsáki et al., 2012]. They are, however,
generally more spatially focused than lower frequencies because they are mostly
generated by small neuron populations [Leuthardt et al., 2006a].
Acquisition systems thus correspond to different combinations of spectral and
spatial resolution and invasiveness. While resolution impacts the feasibility of
accurate, multiple DoF multi-limb control, the invasiveness of the acquisition system
1.2. Motor BCIs 15
Figure 1.6: Location of MEA, ECoG and EEG sensors (reproduced from [Schalk
and Leuthardt, 2011])
determines the chronicity and safety of the system. The resolution-invasiveness

trade-offs associated with MEA-, ECoG and EEG- acquisition systems are detailed
in the next sections.
1.2.2.2 Microelectrode arrays (MEA)
MEA-based motor BCIs have been deployed both in humans [Hochberg et al., 2006]
[Kim et al., 2008] [Simeral et al., 2011] [Kim et al., 2011] [Hochberg et al., 2012]
[Jarosiewicz et al., 2013] [Collinger et al., 2013] [Wodlinger et al., 2015] [Bouton
et al., 2016] and primate subjects [Taylor et al., 2002] [Carmena et al., 2003] [Velliste
et al., 2008] [Orsborn et al., 2011] [Orsborn et al., 2012] [Ifft et al., 2013] [Sadtler
et al., 2014]. MEA, also referred to as "Intracortical Electrode Array", e.g. [Maynard
et al., 1997] [Wodlinger et al., 2015], are grids of needle microelectrodes which are
implanted into the cortex [Homer and Nurmikko, 2013].
Most reported MEA-based motor BCIs relied on the Utah array [Blackrock, 2016]
(Figure 1.7) to acquire neural signals in human subjects [Simeral et al., 2011] [Kim
et al., 2011] [Hochberg et al., 2012] [Jarosiewicz et al., 2013] [Collinger et al., 2013]
[Wodlinger et al., 2015]. This commercialized array gathers 10 × 10 1.5mm-long
electrodes on a 4 × 4mm2 surface. The inter-electrode distance is 0.4mm. The size
of the electrode tips is approximately 4 µm [Simeral et al., 2011], while the diameter
of a neuron body is 50µm or more [Kandel et al., 2000]. Signal acquisition has
nevertheless been performed with other MEAs in a few motor BCIs, e.g. [Orsborn
et al., 2012] [Ifft et al., 2013]. The MEAs used in [Orsborn et al., 2012] and [Gowda
et al., 2014] for example gathered 128 electrodes, with a 35µm diameter and a 0.5mm
inter-electrode distance. Another team has reported bimodal control over virtual
upper-limb prostheses by monkeys [Ifft et al., 2013] using volumetric arrays [Schwarz
et al., 2014], i.e. arrays embedding electrodes of different lengths.
The spectral content of MEA signals stretches as far as 40kHz [Buzsáki et al.,
2012]. Preprocessing permits to extract three signals from the raw signal acquired by
an intracortical electrode, namely Single-Unit Activity (SUA), Multi-Unit Activity
(MUA) and Local Field Potentials (LFP) [Waldert et al., 2009]. MUA and SUA
signals reflect the spiking activity of the few neurons located in the immediate
Figure 1.7: Utah microelectrode array (source: US Patent # 5,215,088)
vicinity of the electrode’s tip (1 or 2 neurons typically) [Homer and Nurmikko,

2013] [Leuthardt et al., 2006a]. LFPs measure the superposed activity of a small
population of neurons located in the neighborhood of the electrode’s tip (spatial
resolution around 1mm) [Homer and Nurmikko, 2013] [Leuthardt et al., 2006a].
Signals acquired by MEAs thus exhibit a high spatial resolution along with a
large spectral content, but correspond to the neural activity generated within a
restricted cortex area [Mak and Wolpaw, 2009]. Even in the case of multiple arrays
implantation, the ability to exploit patterns associated with several brain areas
remains limited [Mak and Wolpaw, 2009].
Because of the highly informative content of the signals measured by MEAs,
MEA-based BCIs hold promise to support complex control strategies (in particular,
accurate multiple DoF and multi-limb control). Tetraplegic patients controlled a
robotic arm in [Hochberg et al., 2012]. 7D and 10D fine control over an upper-limb
prosthesis was demonstrated for MEA-based motor BCIs [Collinger et al., 2013]
[Wodlinger et al., 2015]. Several teams have additionally reported 2D cursor control
by tetraplegic patients [Hochberg et al., 2006] [Kim et al., 2008] [Kim et al., 2011]
[Simeral et al., 2011] [Jarosiewicz et al., 2013]. Finally, cursor and upper-limb
prosthesis control has been achieved by Non-Human Primates during several BCI
studies [Taylor et al., 2002] [Carmena et al., 2003] [Velliste et al., 2008] [Orsborn
et al., 2011] [Orsborn et al., 2012] [Ifft et al., 2013] [Sadtler et al., 2014].
In spite of these promising studies, the invasiveness of intracortical arrays is
for the moment a significant hindrance to their utilization for motor clinical BCI
applications. The MEA implantation surgery causes a mechanical trauma in the
brain [Polikov et al., 2005]. Despite optimizations of the implantation procedure (for
example, of the implantation speed) and of the electrode shape, cell deaths or vessel
ruptures are observed in the cortex [Polikov et al., 2005]. The MEA invasiveness also
induces immune reactions which can be problematic for MEA-based chronic BCIs.
Electrode encapsulation, i.e. the formation of a glial scar tissue around electrodes
(glial cells being supportive cells surrounding the neurons [Squire et al., 2013]),
occurs a few weeks after implantation [Polikov et al., 2005]. It causes an increase of
impedance around electrodes along with an augmentation of the electrode tip-neuron
distance [Polikov et al., 2005]. The subsequent decay of signal amplitude jeopardizes
1.2. Motor BCIs 17
signal acquisition after several months [Yuan and He, 2014], and is thus thought
to limit the life expectancy of the MEA-driven BCIs [Leuthardt et al., 2006a]. In
[Simeral et al., 2011] for example, neurons with a firing rate superior to 1Hz were
associated with only 41 out of the array’s 96 electrodes 1000 days after implantation.
These observations suggest that further improvements may be necessary before
MEAs fully support safe and chronic BCIs [Murphy et al., 2015].
1.2.2.3 Electrocorticography (ECoG)
ECoG-based motor BCIs have been investigated for both human [Leuthardt et al.,
2004] [Leuthardt et al., 2006a] [Wilson et al., 2006] [Schalk et al., 2008] [Blakely
et al., 2009] [Vansteensel et al., 2010] [Milekovic et al., 2012] [Kellis et al., 2012]
[Yanagisawa et al., 2012a] [Wang et al., 2013c] [Fifer et al., 2014] [Kapeller et al.,
2015] and primate subjects [Ashmore et al., 2012] [Sanchez et al., 2008] [Rouse and
Moran, 2009] [Rouse et al., 2013] [Williams et al., 2013] [Marathe and Taylor, 2013]
[Rouse et al., 2016]. Also referred to as "intracranial EEG" [Morshed and Khan,
2014], ECoG grids gather flat electrodes and are implanted onto the cortical surface,
more precisely under or over the dura mater (respectively, subdural and epidural
implantation) [Schalk and Leuthardt, 2011].
To date, most ECoG-based motor BCI clinical studies have been completed with
subjects who had not been not fitted with an ECoG array because they participated
in a BCI study [Vansteensel et al., 2010] [Milekovic et al., 2012] [Kellis et al., 2012]
[Yanagisawa et al., 2012a] [Kapeller et al., 2015]. They were generally enrolled
because they were undergoing a short-term ECoG implantation for the localization
of an epileptic focus before a resection surgery. They were therefore implanted with
clinical ECoG arrays which materials, configuration and size were optimized for the
requirements of this monitoring objective [Schalk and Leuthardt, 2011]. The utilized
clinical ECoG arrays generally consist in titanium, ball-shaped macro-electrodes
placed on a grid (8 electrodes) or a strip (4 or 6 electrodes) [Schalk and Leuthardt,
2011] (Figure 1.8). They are mostly implanted under the dura mater (subdural
implantation). Electrode diameters are typically of 4mm, for an inter-electrode of
1cm [Leuthardt et al., 2004] [Leuthardt et al., 2006a] [Wilson et al., 2006] [Schalk
et al., 2008] [Blakely et al., 2009] [Vansteensel et al., 2010] [Milekovic et al., 2012]
[Fifer et al., 2014]. The number of channels ranged from 15 [Yanagisawa et al.,
2012a] to 128 electrodes [Hotson et al., 2016], possibly via the joint implantation of
several arrays [Fifer et al., 2014]. Similar electrodes have been used for preliminary
studies in human subjects [Schalk et al., 2007] [Chin et al., 2007]. A few teams have
used high-density arrays, e.g. with a 0.5cm [Wilson et al., 2006] [Kapeller et al.,
2015] or 3mm inter-electrode distance in [Hotson et al., 2016] for neural control
over a cursor [Wilson et al., 2006], an humanoid robot [Kapeller et al., 2015] or a
prosthetic hand [Hotson et al., 2016]. The ECoG arrays used to develop and test
motor BCIs on primate models are generally less standardized. Custom arrays with
an inter-electrode distance of 3.5mm have been used to measure signals analyzed in
preclinical feasibility studies [Chao et al., 2010] [Shimoda et al., 2012].
Finally, the use of micro-ECoG (µECoG) arrays, i.e. arrays with an inter-
electrode distance inferior to 4mm and an electrode diameter generally inferior to
100µm [Kellis et al., 2015], has recently been investigated with both human [Kellis
et al., 2012] and primate subjects [Williams et al., 2013] [Rouse et al., 2013] [Rouse
et al., 2016]. The inter-electrode distance of the tested µECoG arrays ranged from
1 to 3mm, with electrode diameters of 30-40 µm [Kellis et al., 2012] or between
300 µm [Williams et al., 2013] [Rouse et al., 2013] [Rouse et al., 2016] and 600µm
[Williams et al., 2013]. The optimal inter-electrode distance is still an active field of
research [Wang et al., 2009b] [Slutzky et al., 2010] [Bundy et al., 2014] [Kellis et al.,
2015] [Wang et al., 2016]. Several studies have suggested that decoding performance
is higher when dense arrays rather than generic arrays with 1cm-large inter-electrode
distance are used [Bundy et al., 2014] [Wang et al., 2016].
Figure 1.8: Commercialized ECoG array [Blackrock, 2017].
ECoG spatial resolution is 1.25mm for subdural ECoG and 1.4mm for epidural
recordings [Schalk and Leuthardt, 2011]. The synaptic currents generated by
pyramidal neurons aligned in the superficial layers of cortex are thought to be the
main sources of ECoG signals [Waldert et al., 2009]. ECoG bandwidth ranges from 0
to 500Hz [Schalk and Leuthardt, 2011]. The exploitation of high frequency patterns,
which have been found useful in several motor studies [Anderson et al., 2012] [Nurse
et al., 2015], is thus possible. Additionally, ECoG signals are significantly less
sensitive to artifacts, in particular ocular ones [Ball et al., 2009a], than non-invasive
acquisition methods like EEG.
Clinical and preclinical attempts at ECoG-based neural control have yielded
encouraging results. Able-bodied epileptic patients have been reported to be able to
control 1D or 2D cursors in several studies [Leuthardt et al., 2004] [Leuthardt et al.,
2006a] [Wilson et al., 2006] [Schalk et al., 2008] [Blakely et al., 2009] [Vansteensel
et al., 2010] [Kellis et al., 2012] [Milekovic et al., 2012]. Epileptic patients controlled
a prosthetic arm, a humanoid robot and a prosthetic hand in [Fifer et al., 2014],
[Kapeller et al., 2015] and [Hotson et al., 2016] respectively. Motor-impaired and
able-bodied patients were able to manipulate a prosthetic hand in [Yanagisawa et al.,
2012a], and 3D arm prosthesis control was achieved by a patient with tetraplegia
in [Wang et al., 2013c]. 1D [Rouse et al., 2013] [Rouse et al., 2016] and 2D cursor
control [Rouse and Moran, 2009] [Ashmore et al., 2012] [Williams et al., 2013]
[Marathe and Taylor, 2013] [Rouse et al., 2016] has additionally been achieved by
NHPs.
1.2. Motor BCIs 19
The implantation of an ECoG array requires a craniotomy. As the array is

not introduced into the cortex but only positioned on it, the ECoG invasiveness is
nevertheless limited when compared to the MEA’s one. ECoG monitoring before
a resective surgery only requires a short-term implantation, namely a few days
or weeks. Most ECoG clinical arrays are therefore FDA-approved for a maximal
duration of 28 days. For this reason, ECoG long-term stability has mainly been
investigated in preclinical studies, e.g. [Chao et al., 2010] [Degenhart et al., 2016].
Chronic preclinical studies have shown that ECoG signals remain stable over months
[Chao et al., 2010] [Shimoda et al., 2012] [Degenhart et al., 2016]. Chronic ECoG
acquisition (over 7 years) has additionally been reported in a study completed with
epileptic patients [King-Stephens et al., 2015]. These findings suggest that ECoG
arrays may be profitably used to develop safe and chronic motor BCI systems.
1.2.2.4 Electroencephalography (EEG)
EEG-based control of prostheses [Pfurtscheller et al., 2000] [Onose et al., 2012]

[Baxter et al., 2013], real limbs [King et al., 2015], humanoid robots [Chae et al.,
2012] and virtual effectors [Wolpaw and McFarland, 2004] [Trejo et al., 2006] [Yuan
et al., 2007] [McFarland et al., 2010] [Royer et al., 2010] [Doud et al., 2011] [LaFleur
et al., 2013] has been achieved in the last twenty years. First designed in the
20s, EEG acquisition systems noninvasively measure electrical brain activities from
electrodes placed on the scalp [Buzsáki et al., 2012].
EEG acquisition systems are widely used in clinical contexts, e.g. for the local-
ization of epileptogenic zones or for coma diagnostics [Teplan, 2002]. For this reason,
EEG systems approved for clinical applications are available on the market. Several
EEG-based motor BCIs have made use of commercialized arrays, e.g. Electrocap In-
ternational’s [McFarland et al., 2010], Compumedics Neuroscan’s [Chae et al., 2012]
[Baxter et al., 2013] or Brain Products’ systems [Onose et al., 2012]. EEG sensors
are usually embedded in a tissue cap that maintains them on the scalp (Figure
1.9). Although motor BCI studies generally exploit montages with 2 [Pfurtscheller
et al., 2000] to 64 [McFarland et al., 2010] [LaFleur et al., 2013] recording electrodes,
high density systems with 264 embedded electrodes have been utilized in several
preliminary BCI studies [Fruitet et al., 2010] [Gwin and Ferris, 2011]. A few teams
have additionally reported the use of commercialized entertainment EEG systems
(the Emotive array [Emotiv, 2016]), for example providing users with neural control
over a robotic arm [Bhattacharyya et al., 2015] or over an upper-limb exoskeleton
[Webb et al., 2012].
EEG acquires the superposed activities of a large neuronal population (> 100000
neurons, approximatively located in a 10 cm2 -region) [Waldert et al., 2009] [Buzsáki
et al., 2012]. Similarly to ECoG, the main contributors to EEG signals are thought
to be the neurons located in the superficial layers of the cortex [Buzsáki et al., 2012].
The activity of deep sources cannot be easily observed in EEG signals because of
the sensor-source distance. The EEG spatial resolution, which amounts to a few
cm, is limited when compared to the MEA’s or ECoG’s ones. EEGs’ amplitude is
Figure 1.9: Example of a clinical EEG cap [Unicare, 2017]
additionally five times lower than ECoGs’ amplitude [Leuthardt et al., 2006a]. EEG
spectral content approximately ranges from 0 to 40Hz [Schalk and Leuthardt, 2011].
The absence of high frequencies is due to signal low-pass filtering when crossing
the skull and skin [Lebedev and Nicolelis, 2006], and to the reduced power of high
frequency components observed from a large distance to their source [Waldert et al.,
2009] (see section 1.2.2.1). EEGs are very sensitive to artefacts, in particular
ocular [Fatourechi et al., 2007a], muscular [Muthukumaraswamy, 2013] [Fatourechi
et al., 2007a], cardiac [Lebedev and Nicolelis, 2006] and mechanical ones. Artefacts
are liable to perturb neural control [Fatourechi et al., 2007a]. Strategies have been
proposed to discard them, but at the risk of losing informative content if the spectral
content of the artefacts is included in the frequency band of interest [Fatourechi
et al., 2007a].
The feasibility of EEG-driven motor BCIs has been explored by several teams.
Binary EEG-based control over a prosthetic and orthesis arm was achieved in [Onose
et al., 2012] and [Webb et al., 2012], respectively. The aperture of a robotic arm
was controlled by a tetraplegic patient in [Pfurtscheller et al., 2000]; 2D control of a
prosthetic and robotic arm was reported in [Baxter et al., 2013] and [Hortal et al.,
2015], respectively. Finally, 1D, 2D and 3D control over virtual effectors has been
accomplished by several able-bodied subjects [Yuan et al., 2007] [Trejo et al., 2006]
[McFarland et al., 2010] [Royer et al., 2010] [Doud et al., 2011] [LaFleur et al., 2013].
While these results are promising, the limited spatial resolution [Leuthardt
et al., 2006a] and high-frequency content of EEG signals [Waldert et al., 2009] may
complicate multi-DoF and multi-limb effector control. In particular, the adaptation
of users to EEG-based decoders often requires extensive training [Leuthardt et al.,
2006a], and, in the case of "BCI illiterates", may not permit users to satisfyingly
control the effector [Milan and Carmena, 2010]. EEG-based BCI studies nevertheless
constitute a significant part of motor BCI studies [Leuthardt et al., 2006a] because
of their non-invasiveness [Teplan, 2002], of their attractive cost and of the relatively
simple procedure which is required to put EEG caps into place [Milan and Carmena,
2010].
1.2. Motor BCIs 21
1.2.3 Transducers for motor BCIs

After acquisition, neural signals are processed by the transducer, which translates
them into intention estimates in real time. The user is regularly provided with
feedback on the transducer output ("closed-loop" experiments) [Héliot et al., 2010].
BCI transducers impact the capability of motor BCI systems to meet several of
the technical challenges presented in section 1.2.1. Transducers must be designed
so as to support real-time updates and to ensure a low feedback latency. These
two requirements are generally achieved by using efficient algorithmic procedures
and exploiting causal neural patterns. Importantly, the transducer also impacts
the system decoding accuracy, which is determined by the transducer’s ability to
extract reliable information from the neural signals and to correctly interpret it.
Transducer design permits to reach a high consistency, or accuracy, between true
and estimated intentions. The mental load associated with multi-limb, multi-DoF
control additionally depends on the decoding strategy the decoder relies on.
1.2.3.1 Transducer design

Before using a BCI transducer during closed-loop experiments, it is necessary to
design it, i.e. to choose its signal components (namely, the pre-processing, feature
extraction, decoder and post-processing blocks shown in Figure 1.10). Transducer
design generally relies on two processes to reach a high consistency (accuracy)
between user intention and transducer’s output: transducer identification and/or
user training [McFarland and Wolpaw, 2011].
The pre-processing, feature extraction and post-processing algorithms utilized
in BCI systems are either user-specific [Marathe and Taylor, 2013] [Kapeller et al.,
2015] or non user-specific [Yanagisawa et al., 2012b] [Wang et al., 2013c], i.e. they
don’t require specific adjustments for each BCI user. In the latter case, they are
generally chosen in preliminary studies, and subsequently utilized regardless of the
user. Their choice is mainly driven by the signal acquisition system embedded in
the BCI system. By contrast, most decoders are user-specific because of the high
variability of users’ model of dependence between brain features and intended BCI
outputs. Their structure is often chosen during a feasibility study and fixed for all
possible users, but their parameters are, with a few exceptions (e.g., [Fazli et al.,
2009]), optimized for each user.
Decoder identification is performed by analysing a data set of simultaneously

acquired neuronal signals and intended effector movements. It is carried out after a
decoder structure has been selected on the basis of preliminary studies, and consists
of tuning this decoder so that it satisfyingly reflects the dependence between neural
features and user’s intentions within this training data set. This tuning phase is also
referred to as decoder "adaptation" [McFarland and Wolpaw, 2011], "learning" (e.g.,
[Hudson and Burdick, 2007]), "training" (e.g. [Ifft et al., 2013]) or "calibration" (e.g.,
[Jarosiewicz et al., 2013]). The training data set is often collected during open-loop
acquisition sessions, i.e. sessions during which the user is not given feedback on the
Figure 1.10: Transducer components.
output of the BCI transducer. The future BCI user is generally cued to repeatedly
generate action-specific patterns, and his brain activity is recorded throughout this
process [Homer and Nurmikko, 2013]. In the case of externally-paced motor BCIs,
the user is simply exposed to the stimuli of interest. In both cases, decoders are
optimized according to their performance for open-loop data decoding.
Decoder identification on open-loop data generally suffers from two limitations.
As the acquired neural signals are noisy and only partially reflect the user brain
activity, open-loop decoders rarely reach perfect decoding performance on the open-
loop data set. Moreover, high open-loop decoding accuracy does not systematically
result in optimal intention estimation in closed-loop settings [Jackson and Fetz,
2011] [Jarosiewicz et al., 2013]. Because of a change of context between open-loop
and closed-loop neural signals, open-loop neural patterns differ from closed-loop
patterns [Leuthardt et al., 2006a] [Jackson and Fetz, 2011] [Jarosiewicz et al., 2013].
Performance drops are regularly observed when an open-loop decoder is applied dur-
ing closed-loop experiments [Tillery et al., 2003]. While it has been found that these
shortcomings of decoder identification may not impact externally-paced systems
[Duprès et al., 2014], they make user training indispensable for internally-paced
motor BCIs.
User training (or adaptation) consists of leading the user to adapt to an imperfect
decoder, i.e. to modify his brain patterns so as to compensate for the transducer’s
erroneous outputs. Thanks to the feedback, the BCI user can assess the difference
between his true intention and the transducer’s estimation of this intention, and
progressively learn how to reduce it by modifying his neural activity. User adaptation
exploits brain plasticity [McFarland and Wolpaw, 2011], i.e. the brain ability to
reorganize to learn new tasks when it is provided with neuro-feedback. While user
training is not required for externally-paced motor BCIs, it is a crucial phase of the
deployment of internally-paced motor BCIs. It is most of the time completed after
the identification of an open-loop decoder [Simeral et al., 2011].
1.2. Motor BCIs 23
Iterative decoder and user adaptation has been reported in preclinical and clinical
motor BCIs [Shenoy and Carmena, 2014] (e.g., [Wang et al., 2013c] [Wodlinger et al.,
2015]). After identification of an initial decoder on open-loop data, the user attempts
to perform tasks during closed-loop experiments. The decoder is then re-identified
so as to take into account modifications in the user’s strategy. In most studies,
several iterations of user-decoder adaptation are performed ("turn-taking" strategy)
[Hochberg et al., 2012] [Wang et al., 2013c], for example 4 to 8 recalibrations for
MEA-based prosthesis control [Hochberg et al., 2012].
1.2.3.2 Decoding approaches for motor BCIs
Two main approaches are used to provide users with control over orthoses or
prostheses. They are thought to impact the user’s ability to control multiple DoF
and multi-limb effectors as well as the mental load associated with neural control.
Kinematic decoding A first decoding strategy consists of estimating the com-

mands for effector control from the activity of neurons naturally devoted to the
control of the corresponding limb. Continuously-valued kinematic parameters of the
effector are generally directly extracted from the corresponding neural signals, for
example the position or velocity of an orthosis endpoint. Several terms have been
used in the literature to refer to such decoders, in particular "kinematic decoders"
[Yuan and He, 2014] and "direct motor Brain Machine Interfaces" [Waldert et al.,
2009]. The term "kinematic decoder" has here been arbitrarily chosen, and will be
used in the remaining of this doctoral dissertation.
Kinematic transducers exploit neural features correlated with the kinematic
parameters of the intended effector movement. Such features were first discovered in
the spiking activity of monkeys performing reaching movements [Georgopoulos et al.,
1982]. Some neurons of the monkeys’ motor cortex were found to preferentially fire
when the monkeys’ arm wass in one specific direction, i.e. they were found to be
directionnally tuned [Waldert et al., 2009] (Figure 1.11). Since then, firing rate
tuning has been extended to other trajectory characteristics. Correlations have
been observed between motor neurons’ spiking activity and hand speed, position,
velocity and acceleration, target localization, joint motion or muscle activation [Scott,
2008]. Neuron tuning has been found to persist when humans with tetraplegia where
attempting to execute arm movements [Hochberg et al., 2006]. This finding suggested
that the utilization of kinematic decoders is achievable by motor-impaired patients.
While tuned features were first thought to be exclusively present in neurons’ spiking
activity, an increasing number of studies have investigated the existence of similar
features in the activity of neuronal populations [Waldert et al., 2009], e.g., LFP
[Mehring et al., 2003] [Mehring et al., 2004], EEG [Waldert et al., 2008] [Bradberry
et al., 2010], ECoG [Gunduz, 2008] [Ball et al., 2009b] [Anderson et al., 2012] or
MEG [Waldert et al., 2008] [Bradberry et al., 2009] signals. Correlations between
neural features in the motor cortex and trajectory kinematics have been disclosed
in ECoG signals [Gunduz, 2008] [Ball et al., 2009b] [Anderson et al., 2012] [Nurse
et al., 2015] [Bundy et al., 2016] and in low-pass filtered EEG [Waldert et al., 2008]
[Bradberry et al., 2010] [Jerbi et al., 2011] signals in humans.
Figure 1.11: Directional tuning of 5 motor cortex neurons (reproduced from [Geor-
gopoulos et al., 1982]).
To date, kinematic transducers have mainly been embedded in invasive BCI

systems (e.g. [Ifft et al., 2013] [Wodlinger et al., 2015]) because tuned features
have initially be found in neurons’ spiking activity. While the feasibility or use of
kinematic transducers has mostly been investigated for upper-limb effectors, results
suggest that they may also be considered for MEA-driven lower-limb effector control
[Fitzsimmons, 2009] [Ma et al., 2017]. When tuned features can be extracted from
neural signals, they can be combined by continuous models to estimate the kinematic
parameters of the intended trajectory. Typical continuous models include regression
models [Collinger et al., 2013] [Wodlinger et al., 2015] [Bundy et al., 2016] and
Kalman Filters [Pistohl et al., 2008] [Hochberg et al., 2012]. Tuned features have
permitted accurate neural control over multiple degrees of freedom [Collinger et al.,
2013] [Wodlinger et al., 2015], or fine movement reconstruction in offline analyses
[Chao et al., 2010] [Ofner and Müller-Putz, 2012] [Shimoda et al., 2012] [Bundy
et al., 2016].
Kinematic decoders are historically biomimetic, i.e. they aim at exploiting the
mapping which related neuronal activity to limb movement before the patient began
to suffer from motor disabilities. Their goal is to provide patients with natural control
over the effector so as to minimize training. Another type of MEA-based kinematic
decoders has nevertheless been explored by a few teams, namely biofeedback decoders
[Ganguly and Carmena, 2009] [Sadtler et al., 2014]. Biofeedback decoders also focus
on the activity of motor neurons, but they rely on user training rather than on the
exploitation of the user’s natural map between neuronal activity and limb kinematic
parameters. The corresponding assumption is that BCI control can only partially
1.2. Motor BCIs 25
mimic natural motor control, in particular because MEAs only permit to acquire the
activity of a very reduced subset of the neurons involved in motor control. The limits
of brain plasticity, i.e. the extent to which a new, unnatural, initially non-intuitive
map can be learned by users, have been investigated in a few studies [Ganguly and
Carmena, 2009] [Sadtler et al., 2014]. While a monkey was able to learn how to
proficiently control a 2D cursor using a decoder tuned with random parameters in
[Ganguly and Carmena, 2009], a second closed-loop study suggests that learning
cannot be achieved if the initial decoder is too far from the user’s natural map
[Sadtler et al., 2014]. The respective relevance of biomimetic and biofeeback decoders,
in particular in terms of training duration, is still unclear, and is sometimes referred
to as the "biomimetic vs. biofeedback" or "decoding vs. learning" dilemma [Carmena,
2013] [Jackson and Fetz, 2011]. To date, most MEA-based motor BCIs nevertheless
relied on biomimetic kinematic decoders [Hochberg et al., 2012] [Wodlinger et al.,
2015].
Mental-task decoding A second approach consists of using the activity elicited

in brain areas which originally were not exclusively devoted to the control of the limb
of interest. The brain patterns used to control the prosthesis or orthosis movements
are elicited by mental tasks such as motor imageries and cognitive tasks. The term
"mental-task decoding" has here been chosen to denote BCI control based on such
mental tasks or strategies [Waldert et al., 2009]. Although it can be argued that this
term may also apply to kinematic decoders, it has nevertheless been selected because
no term has clearly arisen to refer to BCIs based on unnatural motor imageries
and cognitive tasks. Incidentally, mental-task based decoders are the basis of the
majority of internally-paced communication and navigation BCIs. Various mental
tasks have been used to elicit intention-specific and distinguishable brain patterns
for neural control in BCI systems [Waldert et al., 2009]. Motor imagery, i.e. the
fact to imagine moving a limb, is routinely used for neural control in motor BCIs
[Waldert et al., 2009] (e.g., [McFarland et al., 2010] [LaFleur et al., 2013]). Because
of the somatotopic organization of the cortex, motor imageries associated with
different limbs (e.g., tongue, foot, right arm, left arm etc.) generate patterns which
are spatially distinguishable at a macroscopic scale (see Figure 1.12) [Waldert
et al., 2009]. Mental tasks are not exclusively associated with patterns generated
in the motor cortex [Jackson and Fetz, 2011]. Several studies have focused on the
discrimination between cognitive tasks [Waldert et al., 2009], e.g. in [Curran et al.,
2004] [Penny and Roberts, 1999].
Two types of decoders are used to infer the effector kinematics from the user’s
neural signals when the transducer relies on the distinction between mental tasks.
A first strategy consists in considering discrete user’s intentions, i.e. the mental task
i is associated with a movement in the direction i. Continuously-valued commands
are subsequently inferred from the decoder discrete or probabilistic output [Pistohl
et al., 2008]. For example, movement direction can be extracted from the neural
signals by the transducer while movement velocity is fixed at a value chosen by the
Figure 1.12: Right and left hand motor imagery (EEG) (reproduced from [LaFleur
et al., 2013])
experimenter (e.g., [Yanagisawa et al., 2012a] for an ECoG-driven hand prosthesis, or

[Leeb et al., 2007] for EEG-based 1D navigation). Alternatively, the effector velocity
or position can be proportional to the class probability difference [Obermaier et al.,
2001]. Distinction between 5 classes can thus permit navigation in a 2D space (up,
down, right, left, rest). A second type of decoder directly relates the modulation
of one feature or of a set of features to the effector movement along one axis (e.g.,
[Wolpaw and McFarland, 2004] [Schalk et al., 2008] for ECoG signals). Control is
said to be proportional [Wang et al., 2013c]. The set of features is generally chosen
during a screening procedure, which aims at finding features optimally modulated
by mental tasks.
Non-invasive acquisition systems are generally associated with mental-task trans-
ducers [Waldert et al., 2009] [Milan and Carmena, 2010], which remain efficient when
the acquired signals exhibit a limited spatial resolution. 3D EEG neural control
over a quadcopter has for example been achieved in [LaFleur et al., 2013], using
volitional modulation of SMRs elicited via motor imagery. Similarly, EEG-based
neural control permitted users to perform 3D reaching movements in a virtual space
in [McFarland et al., 2010]. Mental-task decoders have also been used for cursor or
prosthesis control from SUA/MUA [Hochberg et al., 2006] or ECoG signals [Wang
et al., 2013c].
Applicability and relevance of kinematic and mental-task decoders The

relative performances of kinematic and mental-task decoders remain a matter of
discussion. The neural control provided by kinematic transducers is expected to be
more precise [Chin et al., 2007] [Nurse et al., 2015] than mental-task-based neural
control. Whereas the users of mental-task decoders are likely to encounter difficulties
to associate mental tasks with the effector movements, kinematic control is meant to
be intuitive [Pistohl et al., 2008] [Schalk et al., 2007] [Ashmore et al., 2012] [Nurse
1.2. Motor BCIs 27
et al., 2015], to involve a lesser mental load [Yuan and He, 2014] and to require
less user training [Waldert et al., 2009]. Although further studies are required to
assess the limits of kinematic approaches, kinematic control is often presented as a
feature generally desirable for BCI systems, including EEG- [Yuan and He, 2014] or
ECoG-based [Chin et al., 2007] BCI systems. The feasibility of kinematic ECoG-
and EEG-driven control is still unclear. In particular, it has been shown that the
ability of users to execute reaching movements is degraded when estimated positions
or velocities are not sufficiently correlated with the user’s intentions, e.g. correlation
is equal or inferior to 0.75 or 0.5 in the case of position and velocity decoding,
respectively [Marathe and Taylor, 2011]. This finding suggests that an accurate
extraction of kinematic parameters from the neural signals is required for kinematic
decoding to benefit patients. Feasability studies for EEG- [Waldert et al., 2008]
[Bradberry et al., 2010] [Jerbi et al., 2011] and ECoG-based kinematic decoders
[Gunduz, 2008] [Ball et al., 2009b] [Anderson et al., 2012] [Nurse et al., 2015] [Bundy
et al., 2016] have recently been completed. The interest of ECoG-driven kinematic
decoders is of particular interest because of the ECoG potential for chronic and
complex neural control.
Kinematic decoding of ECoG signals Research on ECoG-based BCI systems

historically focused on mental-task decoding strategies [Anderson et al., 2012] (e.g.,
[Lal et al., 2005] [Hill et al., 2006] [Chin et al., 2007]), and most ECoG-based closed-
loop studies have been completed with mental-task decoders [Leuthardt et al., 2004]
[Leuthardt et al., 2006a] [Schalk et al., 2008] [Rouse and Moran, 2009] [Williams
et al., 2013] [Wang et al., 2013c]. Recent studies have nevertheless suggested that
the amount of directional tuning in ECoG signals is sufficient for kinematic decoding.
The presence of upper-limb kinematic encoding, in particular position and velocity
encoding, has been disclosed in various neural features.
Position encoding has been reported in several studies. In the 2D tracking task
completed by human subjects in [Schalk et al., 2007], it permitted to reconstruct
trajectories with a correlation of 0.51 and 0.47 for the x- and y-axis, respectively
(average over 5 subjects). Position encoding was mainly observed in the Local Motor
Potential (LMP, i.e. low-pass filtered ECoG signals) (axes x and y), and in the
140-190Hz and 8-12Hz frequency bands for the x- and y-axis, respectively. In [Bundy
et al., 2016], 3D position encoding was mainly found in the LMP and high-gamma
frequency bands, and yielded trajectory reconstructions with Pearson Correlation
Coefficients (PCC) of 0.49, 0.28 and 0.38 for the x-, y- and z-axis, respectively
(average over 5 subjects). In [Hammer et al., 2013], 1D position encoding was
disclosed in the LMP features and in the phase and magnitude of frequency features
extracted between 30 and 200Hz (for a maximum Pearson Correlation Coefficient
of 0.53 at the peak of decoding accuracy, average over 3 subjects). Similarly, 1D
position encoding was mainly found in LMP features in [Hammer et al., 2016]. In
[Pistohl et al., 2008], a Kalman filter which state variable included position and
velocity exploited low-frequency and gamma neural features, and yielded position
accuracy between 0.3 and 0.4 for x-axis, and 0.45 and 0.55 for the y-axis (3 subjects
with non-compromised motor recordings). In [Hotson et al., 2014], the first Principal
Component of subjects’ wrist trajectory was decoded with a correlation coefficient
around 0.7 for each one out of 3 subjects. Depending on the subjects, various
frequency bands between 0 and 200Hz were found informative to predict the wrist
position.
Velocity encoding has been reported in several studies, often in similar features as
the position. Velocity encoding found in the LMP and in the 5-42Hz and 140-190Hz
frequency bands permitted to reconstruct 2D trajectories with a correlation of 0.43
and 0.55 for the x- and y-axis, respectively (average over 5 subjects)[Schalk et al.,
2007]. In [Anderson et al., 2012], neural features extracted from several channels were
found to be significantly tuned to velocity during the execution of 2D movements.
Average PCC of 0.45, 0.35 and 0.42 for the x-, y- and z-axis respectively were
reported between 3D velocity profiles and their ECoG-based reconstructions. In
[Hammer et al., 2013], 1D velocity encoding of LMP and various frequency features
permitted to reconstruct subjects’ wrist velocity with an accuracy comparable to
the one obtained for position decoding. It was nevertheless reported in a later study
that speed encoding was significantly more encoded than velocity [Hammer et al.,
2016], again in LMP-related features. A similar phenomenon was observed in [Bundy
et al., 2016] (average Pearson Correlation Coefficient between true and decoded
speed of 0.82). Finally, a Kalman filter yielded velocity accuracies between 0.4 and
0.5 for x-axis, and 0.48 and 0.6 for the y-axis (3 subjects with non-compromised
motor recordings) [Pistohl et al., 2008].
In [Bundy et al., 2016], speed encoding, where speed refers to the norm of the
velocity, yielded reconstructions with PCC of 0.82, thus surpassing both position and
velocity encoding (average over 5 subjects). While similar results were obtained in
[Hammer et al., 2016], where speed was found to be more encoded than velocity and
position, it was suggested in the tuning study completed by Anderson and colleagues
[Anderson et al., 2012] that speed tuning is less prominent than 2D position and
velocity.
Poor acceleration encoding has been reported in [Hammer et al., 2013] (PCC of
0.3 at the peak of decoding accuracy, average over 3 subjects) and [Hammer et al.,
2016].
Finally, 2D directional encoding has been disclosed in high-gamma signals (85-
250Hz) in 10% of the channels of arrays partly located over the motor cortex [Nurse
et al., 2015] (distinction between 8 targets, i.e. angles of 45◦ ). ECoG 2D directional
tuning has also been reported in [Schalk et al., 2007] and [Anderson et al., 2012].
These findings suggest that kinematic decoders may be utilized to interpret
ECoG signals during arm movements.
1.2.3.3 Decoders for asynchronous control
The ability of generic kinematic transducers to handle asynchronous decoding, i.e.

to limit erroneous activations of the BCI system when the user doesn’t desire to
1.2. Motor BCIs 29
control the effector, is generally poor (see, e.g. [Chao et al., 2010][Williams et al.,
2013] [Velliste et al., 2014]). Similarly, mental-task decoders are regularly designed
for synchronous decoding only [Doud et al., 2011] [Onose et al., 2012] [LaFleur
et al., 2013]. Several studies have therefore been devoted to the development of
efficient transducers for asynchronous control. The majority of these studies were
completed with simplified frameworks, namely the discrimination between intentional
control and idle states [Mason and Birch, 2000] [King et al., 2015] or the distinction
of a limited number of mental tasks from idle states [Millán and Mouriño, 2003]
[Bashashati et al., 2007b] [Bhattacharyya et al., 2015]. While a few asynchronous
kinematic decoders have been proposed by the BCI community [Srinivasan et al.,
2007] [Williams et al., 2013], their respective relevance and limits are still unclear. A
detailed presentation of the technical solutions considered for asynchronous control
is given in Chapter 2.
1.2.4 Effectors
The effectors embedded in motor BCI systems exhibit multiple degrees of freedom
and/or multiple effectors.
Several prosthetic or orthotic devices have been considered for the restoration
of upper-limb mobility. The use of Robotnik’s 6-Degrees of Freedom (DoF) JACO
robotic arm [Robotnik, 2016] has been reported in several motor BCI studies
[Bougrain et al., 2012] [Baxter et al., 2013] [Bhattacharyya et al., 2015]. The
Modular Prosthetic Limb, which has been developped at Johns Hopkins University
and includes a hand and up to 16 controllable DoF, has been controlled by tetraplegic
patients in several clinical trials [Collinger et al., 2013] [Wang et al., 2013c] [Wodlinger
et al., 2015]. The DLR (respectively, the DEKA) robotic device utilized in [Hochberg
et al., 2012] combines a 7-DoF (6-DoF, respectively) arm with a 15-DoF (4-DoF,
respectively) hand. 6-DoF and 5-DoF robotic arms were controlled by a monkey
in [Carmena et al., 2003] and [Velliste et al., 2008], respectively. Finally, a custom
upper-limb orthosis has been developed within the framework of an EEG-based BCI
project [Webb et al., 2012].
Both custom and commercialized hand orthoses and prostheses have been neurally
manipulated by BCI users [Pfurtscheller et al., 2000] [Murguialday et al., 2007]
[Chen et al., 2009] [Ortner et al., 2011].
Several teams also endeavor to restore lower-limb mobility via neural control
over a lower-limb orthesis, e.g. the MINDWALKER, the RoGO, the Rex, the H2
and the Walk Again Project’s exoskeleton in [Gancet et al., 2012], [Do et al., 2013],
[Kwak et al., 2015], [López-Larraz et al., 2016] and [Nicolelis, 2014], respectively.
Finally, early results on the utilization of Functional Electrical Stimulation (FES)
have been presented in a few studies [King et al., 2015] [Bouton et al., 2016] [Vidaurre
et al., 2016]. FES-based BCIs aim at enabling users to regain control over their own
limbs rather than over an external device. FES consists of stimulating the user’s
muscles so that they contract according to the user’s movement intentions.
The control of physical effectors (e.g., orthoses or robotic arms) is necessary
to restore limb mobility. The utilization of virtual effectors has nevertheless been
proposed in several studies, in particular because it permits to facilitate early training
phases in clinical trials [Wodlinger et al., 2015]. Users have been provided with
control over virtual effectors in many preclinical or clinical motor studies, e.g. cursors
[Taylor et al., 2002] [Leuthardt et al., 2006a] [Kim et al., 2008] [Simeral et al., 2011]
or simulated robotic arms in virtual reality environments [Wang et al., 2013c] [Ifft
et al., 2013] [Wodlinger et al., 2015].
1.2.5 Feedback for prosthesis or orthosis control

The ability of the users to interpret and use the feedback conveyed to them is crucial
for motor BCI systems. In the vast majority of motor BCIs, users are exclusively given
visual feedback about the transducer output (e.g., MEA- [Kim et al., 2011] [Hochberg
et al., 2012] [Collinger et al., 2013] [Wodlinger et al., 2015], ECoG- [Vansteensel
et al., 2010] [Milekovic et al., 2012] [Kellis et al., 2012] [Yanagisawa et al., 2012a]
[Wang et al., 2013c], and EEG-based clinical trials [Wolpaw and McFarland, 2004]
[Yuan et al., 2007] [McFarland et al., 2010] [Doud et al., 2011] [LaFleur et al.,
2013]). The addition of other types of feedback, e.g. kinesthetic feedback [Suminski
et al., 2010], has nevertheless been shown to facilitate upper-limb prosthesis control.
Haptic feedback was used to improve neural control over a hand prosthesis in
[Murguialday et al., 2007]; a few teams have completed feasibility [Cincotti et al.,
2007] or cursor control [Chatterjee et al., 2007] studies with vibrotactile feedback.
Finally, combinations of feeback modalities [Suminski et al., 2010] and intracortical
stimulation [O’Doherty et al., 2011] have been investigated.
1.2.6 Summary on current progress

To date, BCI clinical applications mostly consist of non-invasive systems for commu-
nication or environmental control [Mak and Wolpaw, 2009], e.g. EEG-driven spellers
[G.tex, 2016]. Because of the difficulties encountered to meet the requirements
presented in section 1.2.1, motor BCIs have not yet been deployed for everyday
use [Mak and Wolpaw, 2009].
While EEG-based motor BCIs have the significant advantage of being safe, a
long training process is generally necessary before the user is able to adapt to
the mental-task decoder they embed. To date, the corresponding complexity of
control additionally remains inferior to the one associated with MEA-based BCIs,
for example 2D-control over a robotic arm with 4 possible directions [Hortal et al.,
2015], 3D synchronous control [LaFleur et al., 2013], or 2D control decomposed into
sequences of 1D movements [Bhattacharyya et al., 2015].
Invasive or semi-invasive acquisition methods (MEA, ECoG) may be a promising
alternative to EEG for highly accurate, multiple DoF and multi-limb control [Lebedev
and Nicolelis, 2006]. Although invasive and semi-invasive motor BCI clinical trials
are still scarce [Mak and Wolpaw, 2009] because of the potential safety issues they
are associated with, clinical proofs of concept studies have neverthess been completed
1.3. CLINATEC’s motor BCI project 31
in laboratory environments [Hochberg et al., 2012] [Collinger et al., 2013] [Wang

et al., 2013c] [Wodlinger et al., 2015].
The feasibility of 3D [Hochberg et al., 2012], 7D [Collinger et al., 2013] and 10D
[Wodlinger et al., 2015] neural control over a robotic arm has been demonstrated in
recent MEA-based studies. The integration of multi-limb effector control into motor
BCIs has particularly been considered in a few studies [Hochberg et al., 2012] [Ifft
et al., 2013] [Wodlinger et al., 2015] [Bouton et al., 2016]. In [Hochberg et al., 2012],
both sequential or parallel MEA-based control over an upper-limb prosthesis endpoint
and a prosthetic hand were achieved by users with tetraplegia. Parallel control over
the wrist and hand of a robotic arm was additionally reported in [Wodlinger et al.,
2015]. While numerous daily life tasks require bimanual movements [Swinnen and
Wenderoth, 2004], bimanual control has only been reported over virtual effectors [Ifft
et al., 2013]. While these studies suggest the relevance of MEA acquisition systems
and of the reported decoder structure and training strategy, the issues pertaining to
the MEA invasiveness, namely safety and chronicity [Vouga et al., 2017], are to date
only partially addressed.
While ECoG arrays hold promise of chronic and stable signal acquisition, the
reported ECoG-driven motor BCIs generally relied on mental-task decoders [Schalk
et al., 2008] [Fifer et al., 2014] [Wang et al., 2013c] [Kapeller et al., 2015] and did
not permit users to achieve complex effector control. Control over a set discrete
commands was achieved using ECoG signals in [Fifer et al., 2014] and [Hotson
et al., 2016], and 3D [Wang et al., 2013c], 2D [Schalk et al., 2008] and 1D control
[Vansteensel et al., 2010] [Leuthardt et al., 2011] has been reported in a few studies.
If kinematic control has not yet been completed in human subjects, 2D kinematic
control was accomplished by monkeys in [Marathe and Taylor, 2013]. ECoG-based
multi-limb control has mainly been considered in the case of multi-finger offline
trajectory reconstruction [Wang et al., 2009b] [Wissel et al., 2013] [Saa et al., 2016]
[Liang and Bougrain, 2012] [Kubánek et al., 2009] [Acharya et al., 2010] [Flamary
and Rakotomamonjy, 2012] [Nakanishi et al., 2014b]. Individual finger ECoG-based
control was nevertheless restored in [Hotson et al., 2016]. To date, the degree of
complexity achieved with ECoG-driven control is consequently surpassed by the ones
reported for MEA-based BCIs. Synchronous protocols have mainly been considered,
and studies on ECoG-based effector chronic control are still lacking. The proof that
chronic asynchronous ECoG control over multi-limb multi-DoF effectors is feasible
thus remains to be established.
1.3 CLINATEC’s motor BCI project

CLINATEC’s BCI project is meant to contribute to the recent efforts towards motor
BCI clinical applications [Eliseyev et al., 2014]. It aims at allowing patients with
tetraplegia to chronically recover both mobility and arm control in the framework
of a 5-year clinical trial recently approved by the French competent authorities
[ClinicalTrials.gov, 2016] (see Figure 1.13).
Figure 1.13: CLINATEC’s BCI system (in particular, WIMAGINE implantr,

transducer, EMY exoskeleton).
1.3. CLINATEC’s motor BCI project 33

To date, obstacles to MEA-based chronic signal acquisition remain to be fully
removed, and the performances of EEG-based motor BCIs may be limited by EEG
spatial resolution and frequency content. An alternative approach has therefore
been selected for CLINATEC’s motor BCI system, namely an ECoG acquisition
system. The ECoG capacity to address the challenges of motor BCI systems will
be explored during CLINATEC’s clinical trial. Because of the lack of chronic
ECoG implants (see section 1.2.2.3), a wireless ECoG implant has been especially
designed for long-term BCI applications [Mestais et al., 2015]. CLINATEC’s ECoG
implant WIMAGINEr, shown in Figure 1.14, gathers 64 electrodes on a 5cm-
diameter surface. A distance of approximatively 4mm separates the electrodes’
centres [Mestais et al., 2015]. CLINATEC’s BCI system has been developed so as to
rely on internally-paced neural control, and WIMAGINEr has been designed for
an implantation on the user’s motor cortex, i.e. for the acquisition of users’ SMR.
WIMAGINEr’s safety has been approved by the competent authorities [Mestais
et al., 2015]. In particular, its implantation is performed by means of a standardized
craniotomy procedure which limits post-implantation complications.
Figure 1.14: WIMAGINEr implants.
1.3.2 Transducer
The transducers developed in CLINATEC were mainly devoted to two signal pro-
cessing challenges, namely accurate mono-limb and asynchronous control.
Kinematic decoding for accurate mono-limb control In view of the recent

studies which have suggested that the amount of directional tuning in ECoG signals
may be sufficient for kinematic decoding (see section 1.2.3.2), CLINATEC’s signal
processing team has developed a kinematic decoding approach for the restoration of
upper-limb mobility. The transducer first extracts high-dimensional time-frequency-
space features from the neural signals. A preliminary pre-processing approach
is used to discard artefacts where necessary [Eliseyev and Aksenova, 2014]. The
kinematic parameters of interest are then regressed against the neural features
using the efficient decoder training approaches designed in CLINATEC [Eliseyev
and Aksenova, 2013] [Eliseyev and Aksenova, 2014]. Offline validation of the
corresponding transducers has been completed by using ECoG signals to reconstruct
the 3D kinematic parameters of monkeys’ wrist and/or shoulder and elbow [Eliseyev
and Aksenova, 2013] [Eliseyev and Aksenova, 2014].
State detection for asynchronous decoding First steps towards asynchronous

decoding have additionally been completed by CLINATEC’s team, namely the
development of efficient binary switches for the discrimination between discrete
mental states (e.g., idle states against desire to control the effector) [Eliseyev et al.,
2011] [Eliseyev et al., 2012]. The considered binary transducers rely on the same
feature extraction procedure than CLINATEC’s kinematic transducers. While online
preclinical studies have permitted to validate the relevance of these transducers for
binary state prediction [Costecalde, 2012], their integration into kinematic decoders
had not been explored yet.
1.3.3 Effector
Clinatec BCI system aims at providing users with multiple DoF control over the
4-limb exoskeleton EMY (for Enhancing MobilitY) [Morinière et al., 2015] [Eliseyev
et al., 2014]. Developed at CEA-LIST, EMY (shown in Figure 1.13) presents up
to 14 controllable degrees of freedom (4 for each arm [Morinière et al., 2015], 3 for
each leg). The projected operating protocol consists in offering 7 degrees of freedom
(DoF) asynchronous control to patients (3 DoF for each arm, a switch to activate or
deactivate walking).
1.3.4 Feedback
Although the integration of alternative feedback types is envisaged for a future
version of Clinatec’s BCI system, neural control will initially rely on visual feedback
only.
1.4 PhD objectives

The present doctoral work has been completed within the framework of CLINATEC’s
motor Brain-Computer-Interface (BCI) project. It tackles three challenges specific
to the development of transducers for motor BCI clinical applications, namely
1.4. PhD objectives 35
asynchronous decoding, multi-limb decoding and decoding accuracy during active

states.
1.4.1 Asynchronous control

Although asynchronous control has been considered in a few motor clinical BCIs, most
motor clinical trials have been completed using a synchronous protocol [Hochberg
et al., 2006] [Wodlinger et al., 2015]. The deployment of synchronous BCI systems re-
quires the presence of an operator to switch the system on and off. The impact of the
BCI system on users’ independence is therefore limited. By contrast, asynchronous
BCIs are potential stand-alone systems. Generally, users alternate between periods
of Intentional Control (IC) and of No-Control (NC), during which they don’t intend
to use the BCI system [Mason et al., 2006]. The limitation of erroneous activations
of the BCI system during NC states is all the more important since users of motor
BCIs physically interact with the effector, in contrast with BCIs based on the control
of a virtual effector (e.g., a cursor on a computer screen). False activations are
likely to be particularly disturbing and stressful to users. If BCI outputs are not
accurate (e.g., neutral) during the NC epochs, users need to actively force them to
neutral values. A high mental load is required by such constantly-engaged control
paradigms. While NC support is highly desirable for motor BCIs [Leeb et al., 2007],
its integration into kinematic decoders has only been partially addressed in the
literature. A decoder, referred to as Switching Markov Linear Model, has been
designed in the present doctoral work to perform asynchronous kinematic decoding
with a limited number of false activations.
1.4.2 Multi-limb control

Multi-limb control is desirable for patients with tetraplegia, in particular right and
left upper- and lower-limb control and hand and/or finger control. The integration
of multi-limb effector control into motor BCIs has only been considered in a few
studies [Hochberg et al., 2012] [Ifft et al., 2013] [Wodlinger et al., 2015] [Bouton et al.,
2016]. In particular, the issue of asynchronous sequential multi-limb control has
rarely been tackled. The proposed decoder, the Switching Markov Linear Model, has
been developed so as to address the challenge of sequential multi-limb asynchronous
decoding in addition to the one of mono-limb asynchronous decoding, i.e. to prevent
parallel limb activations.
1.4.3 Accurate control

An accurate extraction of kinematic parameters from the neural signals is required
for kinematic decoding to enable patients to interact with their environment (see
section 1.2.3.2). Focus has thus been set on the improvement of decoding accuracy
during intentional control states. A specific configuration of the proposed decoder
has been utilized to improve decoding accuracy during IC periods.
Proper transducer design is necessary to achieve accurate, asynchronous multi-

limb control. The next chapter (Chapter 2) focuses on the technical solutions
which have been considered by the BCI community to address these issues. The
transducer proposed in the present doctoral work, namely the Markov Switching
Linear Model, is presented in Chapter 3. Chapter 4 consists in a description of
the data sets used to assess the performance of the Markov Switching Linear Model.
The procedure completed to measure the decoder performance and the benchmark
decoders are exposed in Chapter 5. Details on the transducer implementations are
given in Chapter 6. Results are reported in Chapter 7. Finally, perspectives are
presented in Chapter 8.
Chapter 2
Data-driven transducers for

motor BCIs
Contents
2.1 Notations and processing blocks . . . . . . . . . . . . . . . . 38
2.2 Pre-processing . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
2.3 Feature extraction . . . . . . . . . . . . . . . . . . . . . . . . . 41
2.3.1 Neural feature extraction . . . . . . . . . . . . . . . . . . . . 42
2.3.2 Features for effector control . . . . . . . . . . . . . . . . . . . 46
2.3.3 Dimensionality reduction . . . . . . . . . . . . . . . . . . . . . 47
2.4 Decoders for motor BCIs . . . . . . . . . . . . . . . . . . . . . 49
2.4.1 Generalities on decoder training . . . . . . . . . . . . . . . . 49
2.4.2 Discrete decoding - classifiers . . . . . . . . . . . . . . . . . . 50
2.4.3 Continuous decoding . . . . . . . . . . . . . . . . . . . . . . . 58
2.5 Post-processing . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
2.5.1 Discrete output . . . . . . . . . . . . . . . . . . . . . . . . . . 63
2.5.2 Continuous output . . . . . . . . . . . . . . . . . . . . . . . . . 64
2.6 Transducers for accurate asynchronous kinematic decoding 64
2.6.1 Asynchronous mono-limb decoding . . . . . . . . . . . . . . . . 64
2.6.2 Asynchronous sequential multi-limb decoding . . . . . . . . . 66
2.6.3 Accurate decoding during IC states . . . . . . . . . . . . . . . 66
2.7 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67
Accurate asynchronous neural control over multi-limb prostheses or orthoses

requires relevant online processing methods to be applied on the user’s neural signal.
Translation of the user’s brain activity into control signals is performed by the BCI
transducer [Bashashati et al., 2007a]. The present chapter presents the algorithmic
tools which have been utilized in motor BCI transducers to address the challenge of
accurate, asynchronous and/or multi-limb control. As CLINATEC’s motor BCI relies
on internally-paced neural patterns, focus has been set on data-driven transducers,
i.e. on transducers which decoder is user-specific and trained on neural data before
utilization.
38 Chapter 2. Data-driven transducers for motor BCIs
2.1 Notations and processing blocks

Let st ∈ Rnc be a multi-channel raw neural signal, where nc ∈ N∗ is the number
of channels. Let yt denote the corresponding user movement intention, yt ∈ N if
discrete-valued intentions are considered (e.g., movement towards the right or the
left) and y t ∈ Rn if continuously-valued intentions are to be decoded. When the
BCI system is turned on ("executed"), its transducer regularly issues an estimate ŷt
of the user’s intention yt from neural signals st . The computation of ŷt is usually
decomposed into several processing steps, namely signal pre-processing, feature
extraction, decoder application and post-processing (see Figure 2.1.B).
Figure 2.1: General processing block. A. Decoder identification stage. B. Transducer

execution stage.
First, an optional pre-processing transform is applied on the instantaneous raw

signal st or on an τ1 -long epoch of raw signal s(t+1−τ1 ):t . It outputs an enhanced
signal s̃t ∈ Rnc , for example a signal with an increased Signal-to-Noise Ratio. A
transform h : Rnc×τ2 → Rm is then used to extract features xt ∈ Rm from an τ2 -long
epoch of enhanced neural signal s̃(t+1−τ2 ):t . Feature extraction permits to build a
new representation of neural signals, bringing out signals’ informative attributes and
discarding redundant or irrelevant characteristics. Next, a function fˆ : Rm → Rn or
fˆ : Rm → N is applied on the neural features xt . This "decoder" permits to infer
an estimate ŷt = fˆ(xt ) of the user intention yt from neural features. The decoder
output ŷt is then optionally post-processed. The enhanced estimate ŷct is sent to
the effector controller. The latter drives the effector according to the decoded user
intention.
Before executing the BCI transducer, it is necessary to choose optimal transforms

for each of its components. In particular, the decoder fˆ is generally identified
) *T ) *T
using a training data set (X, Y), where X = xt t=1 and Y = yt t=1 . This
2.2. Pre-processing 39
preliminary modelling phase is illustrated in Figure 2.1.A. It is referred to as

"decoder identification", "training" or "calibration".
2.2 Pre-processing
Decoders are liable to issue erroneous intention estimates when they are trained
and/or applied on neural signals which are noisy or corrupted by artefacts. Pre-
processing filters aim at discarding artefacts and/or increasing the Signal-to-Noise-
Ratio (SNR). They are profitably used for various decoding tasks, e.g. kinematic or
asynchronous binary decoding.
Artefact management Artefacts are signals which are irrelevant for neural
control, e.g. signals with non-cerebral sources like ocular movements or heartbeats
[Fatourechi et al., 2007a]. The EEG sensitivity to artefacts, in particular artefacts
with ocular, cardiac or muscular origins, has been demonstrated in several studies
[Fatourechi et al., 2007a] [Whitham et al., 2007] [Muthukumaraswamy, 2013]. Albeit
to a lesser extent, ECoG signals are also liable to be corrupted by artefacts, e.g.
blink artefacts in subdural ECoG signals [Ball et al., 2009a] or chewing artefacts in
epidural ECoG signals [Shimoda et al., 2012]. Finally, the presence of common-noise
artefacts has been reported in intracortical recordings [Paralikar, 2010].
Artefact rejection and artefact removal methods have been developed to limit
artefact-induced perturbations during decoder training and/or execution. Manual
or automatic artefact rejection consists of discarding corrupted samples [Fatourechi
et al., 2007a] and is mainly performed for offline cleaning of training data before
model identification (e.g., [Kubánek et al., 2009] [López-Larraz et al., 2016]). Visual
inspection [Kubánek et al., 2009] or outlier detection [López-Larraz et al., 2016] are
example of strategies which were used to reject artefacts before decoder training.
By contrast, artefact removal methods aim at correcting neural signals corrupted by
artefacts and are suited for online application.
Main artefact removal methods are temporal filtering, linear regression and
spatial filtering [Fatourechi et al., 2007a].
Temporal filtering is used to reject frequency bands containing artefacts [Fa-
tourechi et al., 2007a], e.g. low-pass filtering to discard muscular artefacts which
are mainly observed between 20 and 300Hz [Muthukumaraswamy, 2013]. Low-pass
filtering below 30Hz has for example been reported in several EEG-based motor
BCI systems [Sadeghian and Moradi, 2007] [Herman et al., 2008]. Band-stop Notch
filters are regularly applied to remove power line artefacts at 50 or 60Hz, e.g. in
ECoG signals [Anderson et al., 2012] [Nurse et al., 2015] [Spüler et al., 2016]. The
potential rejection of frequency bands of interest is a shortcoming of temporal
filtering [Fatourechi et al., 2007a]. In EEG-based BCIs, temporal filtering-based
rejection of muscular activity results in the elimination of gamma activity, which has
been found relevant for limb kinematic reconstruction from ECoG signals [Leuthardt
et al., 2004] [Anderson et al., 2012] [Nurse et al., 2015] [Bundy et al., 2016].
Removal methods based on linear regression [Fatourechi et al., 2007a] don’t

suffer from this drawback (e.g., [Trejo et al., 2006]). They are applicable as long as
a signal containing the artefact is acquired simultaneously with the neural signals
[Urigüen and Garcia-Zapirain, 2015], e.g. the user’s ElectroOculoGraph (EOG) and
ElectroCardioGraph (ECG) in case of ocular and cardiac artefacts, respectively
[Wallstrom et al., 2004] [Waser and Garn, 2013]. The enhanced signals are computed
by subtracting the reference artefactual signal from the neural signals, i.e. they are
a linear combination of the corrupted signals and of the reference artefactual signal.
This approach is suboptimal if the reference artefactual signal contains some neural
signals of interest [Fatourechi et al., 2007a].
Spatial filtering is another strategy which use has been reported for artefact
removal [Fatourechi et al., 2007a], for example from ECoG signals in [Liu et al.,
2004] or EEG signals in [Brunner et al., 2007]. Online acquisition of pure artefactual
signals is not necessary when artefact removal is completed with spatial filters such as
Principal Component Analysis (PCA) and Independent Component Analysis (ICA)
[Urigüen and Garcia-Zapirain, 2015]. Artefact extraction relies on the projection of
the neural signals onto a new basis ("virtual channels"). The subspace A spanned
by the first k virtual channels is associated with the signals of interest (e.g., with
neural sources), and the complementary subspace with signals with artefactual
origin. Online artefact removal consists of projecting the neural signals onto the
subspace A. The corrupted virtual channels are thus discarded [Urigüen and Garcia-
Zapirain, 2015]. The subspaces associated with the signals of interest or the artefacts
are identified during preliminary offline analyses. The use of PCA, which yields
uncorrelated virtual channels of decreasing variance, has for example been considered
for subspace identification within the framework of magnetoencephalographic data
cleaning [Kelly et al., 2011]. A stronger assumption of statistical independence
between the signals of interest and the artefacts is exploited by ICA-based algorithms.
ICA-based virtual channels have for example been used in several EEG-based BCI
studies [Vigário, 1997] [Iriarte et al., 2003] [Brunner et al., 2007].
The relevance of alternative artefact removal approaches has been investigated
in several BCI studies [Eliseyev and Aksenova, 2014] [Foodeh et al., 2016]. Chewing
artefacts in ECoG signals were for example detected as outliers in time-frequency
features and replaced by neutral values in [Eliseyev and Aksenova, 2014].
Despite several demonstrations of the presence of strong artefacts in EEG signals
[Fatourechi et al., 2007a] [Muthukumaraswamy, 2013], the use of artefact removal
methods is rarely indicated in EEG-based motor BCI studies. For example, attempts
at artefact monitoring and/or management was not mentionned in the EEG-driven
motor BCI systems developed in [Chae et al., 2012] [Webb et al., 2012] [Baxter
et al., 2013] [Karin and Andres, 2014] and [Vidaurre et al., 2016]. Similarly, artefact
removal is infrequently mentioned in ECoG-based motor BCIs. In [Trejo et al.,
2006], however, ocular artefacts were subtracted from EEG signals during a virtual
effector (namely, a cursor) online control task. The use of simple approaches like
visual signal monitoring [Doud et al., 2011] [Onose et al., 2012] [LaFleur et al., 2013]
or the exclusion of channels liable to include artefactual patterns [Fifer et al., 2014]
2.3. Feature extraction 41
[Milekovic et al., 2012] [Lew, 2012] have also been reported in motor EEG- and
ECoG-based BCIs.
Signal enhancement Similarly, both spatial or temporal filters can be applied

to increase the SNR.
Most reported spatial filters are the Common Average Reference (CAR), the
Surface Laplacian (SL) and the Common-Spatial-Pattern (CSP) filters, along with
the already mentioned PCA and ICA [Bashashati et al., 2007a]. The use of CAR
filtering has been reported in numerous EEG- [Lew, 2012] [Galán et al., 2008] and
ECoG-based [Chao et al., 2010] [Shin et al., 2012] [Hammer et al., 2013] [Schalk
et al., 2007] [Hotson et al., 2014] [Fifer et al., 2014] motor BCI studies. It consists
of computing the average value of the neural signals and in subtracting this value
from all channels. SL-based pre-processing is also regularly performed in motor
EEG-based BCI systems [McFarland et al., 2010] [Chae et al., 2012]. Signals’
second spatial derivatives are extracted to discard patterns which are similar for
all electrodes, and are therefore associated with deep sources [Carvalhaes and De
Barros, 2015]. By contrast, local and superficial sources are preserved. Finally, CSP
filtering is frequently applied for neural signal preprocessing before EEG classification
[Townsend et al., 2004] [Sadeghian and Moradi, 2007] [Onose et al., 2012]. Its use
has also been reported for offline [Wei and Tu, 2008] and online [Onaran et al.,
2011] [Marathe and Taylor, 2013] [Kapeller et al., 2015] ECoG processing. The CSP
constructs virtual channels under the criterion that their variance ratio is maximized
[Blankertz et al., 2008].
In addition to their application for artefact filtering, temporal filters are used
to avoid aliasing effects. Anti-aliasing filters are low-pass filters which permit to
respect the Nyquist-Shannon sampling theorem (Nyquist frequency), which states
that sampling frequency should be at least twice superior to the maximal frequency
present in the signal. Anti-aliasing filters have been applied in numerous EEG-based
motor BCIs, e.g. [Flotzinger et al., 1994] [Sadeghian and Moradi, 2007] [Herman
et al., 2008]. They were as well utilized in motor ECoG-based studies, e.g. [Schalk
et al., 2007] [Bundy et al., 2016].
Finally, low-pass (< 300Hz) and high pass (> 300Hz) filtering of MEA signals is
performed to extract LFP and enhance MUA/SUA signals, respectively [Waldert
et al., 2009].
2.3 Feature extraction
The application of efficient feature extraction algorithms is a prerequisite for all

decoding tasks, e.g. synchronous or asynchronous kinematic reconstruction as well
as mental task discrimination.
2.3.1 Neural feature extraction

The application of several classes of transforms h : Rnc×τ2 → Rm has been reported
for the extraction of neural features.
2.3.1.1 Spike count
Spike counts in short time bins (i.e., neuron instantaneous firing rate) are often
used to characterize SUA/MUA signals in intracortical motor BCIs (e.g., [Collinger
et al., 2013] [Wodlinger et al., 2015]). Spike detection is generally performed by
thresholding the neural signal. In the case of MUA signals, an additional spike
sorting step is sometimes carried out to decouple the activity of each observed unit.
Although spike-based features have been found to encode information relevant for
trajectory decoding, their extraction cannot be performed from ECoG signals.
2.3.1.2 Temporal features
Temporal features, i.e. sequences of instantaneous neural signal characteristics, have

been exploited in several BCI systems. Reported features include signals’ amplitude
at particular instants [Palankar et al., 2008] [Flamary and Rakotomamonjy, 2012]
and signals’ latency with respect to a stimulus [Mak and Wolpaw, 2009]. Raw signals
are possibly low-pass filtered [Flamary and Rakotomamonjy, 2012] or averaged
[Palankar et al., 2008] beforehand to increase their SNR.
The amplitude of neural signals after low-pass filtering has been found to encode
useful information for the reconstruction of limb kinematic parameters from ECoG
signals (see section 1.2.3.2). It is thus regularly exploited in ECoG-based motor
BCIs [Hammer et al., 2016] [Pistohl et al., 2008] [Ball et al., 2009b] [Kellis et al.,
2012] [Milekovic et al., 2012] [Wang et al., 2012] [Hotson et al., 2014] [Schalk et al.,
2007]. Low-pass filtering is generally performed by means of a Butterworth [Hammer
et al., 2016], Savitzky-Golay [Pistohl et al., 2008] [Ball et al., 2009b] [Kellis et al.,
2012] [Milekovic et al., 2012] or Moving Average [Wang et al., 2012] [Hotson et al.,
2014] filter.
Most reported ECoG-based BCIs more precisely relied on the combination of
temporal, amplitude-based features with time-frequency features [Schalk et al., 2007]
[Pistohl et al., 2008] [Wang et al., 2012] [Hotson et al., 2014].
2.3.1.3 Time-frequency and time-scale features
Neural signals are often characterized by their spectral content in short time windows,
especially in the case of field potentials-based motor BCIs (i.e. LFP [Aggarwal et al.,
2013] [Flint et al., 2013], ECoG [Wang et al., 2013a] [Chin et al., 2007] [Yanagisawa
et al., 2012b] and EEG [Pfurtscheller et al., 2000] [Trejo et al., 2006] [LaFleur et al.,
2013] [Hortal et al., 2015]). A sine wave is fully described by its frequency, its
amplitude and its phase. Both amplitude- and phase-based features have been used
in motor BCIs. Their temporal evolution within specific frequency bands is usually
considered. Most reported internally-paced ECoG- and EEG-based motor BCIs

relied on such time-frequency features. Different methods have been utilized for the
extraction of time-frequency features from neural signals.
Amplitude Both power- and magnitude-based features have been extracted to

characterize neural signals in motor BCIs [Wang et al., 2013a] [Chin et al., 2007]
[Yanagisawa et al., 2012b] [Eliseyev and Aksenova, 2014]. The instantaneous power
- -2
of a signal xt is defined as its squared amplitude -xt - , and its magnitude by its
absolute value -x -. In most BCI studies, the power or magnitude average value is
- t-
then computed in temporal windows and used as a neural feature [Ball et al., 2009b].
As power features are extracted from some specific frequency bands, they are often
referred to as "band-power" features. Different strategies have been used to extract
amplitude information in motor BCIs.
Bank filter is one of the approaches which use has been reported for the extraction
of amplitude-based time-frequency features from EEG [Bashashati et al., 2015] and
ECoG [Wang et al., 2013b] [Wang et al., 2013a]. It consists of using a set of
real-valued band-pass filters, e.g. Butterworth filters [Shin et al., 2012], to extract
signal components included in specific frequency bands [Brodu et al., 2011]. The
instantaneous or average power or magnitude within a particular frequency band
is then estimated from the amplitude of the filtered signals. In some studies, the
amplitude of the filtered signal was post-processed [Pistohl et al., 2008].
The application of the Short-Time-Fourier-Transform (STFT) has also been
reported for feature extraction from ECoG [Chin et al., 2007] [Yanagisawa et al.,
2012b] signals. The STFT is obtained by applying a temporal window w (e.g., a
Hamming [Chin et al., 2007], Hann, Bartlett or Gaussian window) on neural signals
before computing their Fourier transform. The squared amplitude of the Fourier
descriptors |ST F T (t, w, f )|2 , where f is the frequency of interest and t characterizes
the considered instant, is used to compute the signal spectrogram.
Another approach reported in several EEG-, MEA- ECoG-based BCI studies
consists of applying a wavelet transform to compute a time-frequency representation
of neural signals [Chao et al., 2010] [Bhattacharyya et al., 2011] [Shimoda et al.,
2012] [Eliseyev and Aksenova, 2014] [Bashashati et al., 2015] [Bouton et al., 2016].
The use of different wavelets has been investigated, e.g. Daubechies [Bhattacharyya
et al., 2011] [Bouton et al., 2016], Meyer [Eliseyev et al., 2012], Haar [Kousarrizi
et al., 2009] or Morlet wavelets [Lemm et al., 2004] [Chao et al., 2010] [Eliseyev
and Aksenova, 2014] [Bashashati et al., 2015]. Wavelets are real- or complex-valued.
Both real [Chao et al., 2010] and complex Morlet [Eliseyev and Aksenova, 2014]
wavelets have for example been used for 3D wrist trajectory decoding from ECoG
signals. Complex-valued wavelets can be used to extract phase information from the
analysed signals [Fernandes et al., 2003] (see next paragraph). In contrast with the
STFT, the temporal resolution of the wavelet transform depends on the considered
frequency. Temporal resolution is lower for low-frequency bands, but is compensated
by a higher frequency resolution [Brodu et al., 2011]. Wavelet-based extraction of the

instantaneous power [Chao et al., 2010] [Shimoda et al., 2012] or magnitude [Eliseyev
et al., 2012] [Eliseyev and Aksenova, 2014] at specific instant has for example been
reported for the extraction of kinematic parameters from ECoG signals. The squared
amplitude of wavelet transform is referred to as "scalogram", and the corresponding
features as "time-scale" features.
The Hilbert transform has also been used in BCI studies, for example to analyse
the power spectral density in several frequency bands of ECoG signals [Yanagisawa
et al., 2012a].
Finally, parametric spectrum estimation is a popular approach for the charac-
terization of both ECoG [Lal et al., 2005] [Leuthardt et al., 2004] [Hill et al., 2006]
[Felton et al., 2007] [Schalk et al., 2007] [Schalk et al., 2008] [Blakely et al., 2009]
[Leuthardt et al., 2011] [Ashmore et al., 2012] [Wang et al., 2012] [Wang et al.,
2013c] [Fifer et al., 2014] and EEG signals [Schlögl et al., 2005] [Argunşah and Çetin,
2010]. The Auto-Regressive (AR) coefficients of the neural signals are estimated, for
example via the Yule-Walker [Herman et al., 2008] or Burg method [Ashmore et al.,
2012] [Fifer et al., 2014]. Spectrum estimation is inferred from the AR parameters.
When it is based on Burg AR paramers, it is referred to as Maximum-Entropy
Spectral Estimation. Maximum-Entropy Spectral Estimation has been performed in
several ECoG studies [Anderson et al., 2012] [Bundy et al., 2016] [Spüler et al., 2016].
The above-mentioned time-frequency and time-scale features have rarely been

compared on a common data set. Two extensive studies have nevertheless been
completed on EEG signals elicited by motor imageries [Herman et al., 2008] [Brodu
et al., 2011]. The periodogram and parametric power estimation approaches permit-
ted to extract features associated with the best classification accuracy in [Herman
et al., 2008], whereas Morlet wavelet transforms surpassed alternative methods in
[Brodu et al., 2011]. These inconsistent results seem to suggest that the relevance of
different time-frequency and time-scale features partially depends on the data sets
at hand.
Phase Most motor BCI systems don’t exploit the phase information of the neural
signals, which can for example be extracted using Fourier Analysis [Krusienski et al.,
2012], Hilbert transforms [Hamner et al., 2011] and complex Wavelet transforms
[Le Van Quyen et al., 2001] [Bruns, 2004]. Two classes of phase-based features have
nevertheless been explored in offline and online studies.
The first class of features is formed by the phase information associated with
each channel [Hammer et al., 2013], e.g. the instantaneous phase value of each
channel. While the use of such features has seldom been reported, they have been
shown to contain useful information for 1D kinematic offline recontruction from
ECoG signals [Hammer et al., 2013].
A second approach consists in considering the phase difference between signals
rather than the absolute phase of each signal. This phase difference is used to
characterize the coupling between two channels [Bruns, 2004]. The extraction of the
Phase-Locking-Value (PLV), which measures the degreee of phase locking between
channel k and channel l, has been reported for the classification of mental tasks
in several BCI studies, e.g. [Wang et al., 2006] [Wei et al., 2007] [Pourbakhtiar
et al., 2013] [Gysels and Celka, 2004] [Loboda et al., 2014]. The PLV of channels
k and l is defined as the average of -the instantaneous - phase difference in a T -long
- 1 qT (t)
temporal windows, i.e. P LV (f ) = - T t=1 e j∆φ kl -, where f is the frequency of
-
interest and ∆φkl (t) is the instantaneous phase difference between the two channels.
The relevance of alternative phase-based features, e.g. the instantaneous or mean
phase difference between two channels, has also been investigated [Hamner et al.,
2011].
Although phase features permitted to control a 3-class virtual effector in [Brunner
et al., 2006], they have mainly been utilized in offline studies. Coherence-, Fourier-
based and phase features were compared in [Krusienski et al., 2012] for motor
imagery offline classification from EEG signals (acquired during 1D cursor control).
Phase and coherence features did not lead to an improved classification accuracy
when compared to Fourier features. By contrast, phase features have been shown to
outperform magnitude features for 1D kinematic offline recontruction from ECoG
signals [Hammer et al., 2013]. The relevance of phase-related features thus remains
unclear.
2.3.1.4 Time-frequency-space features
The integration of spatial information into time-frequency features has been reported
for both EEG- [Onose et al., 2012] [Vidaurre et al., 2016] and ECoG-based [Marathe
and Taylor, 2013] [Kapeller et al., 2015] motor BCI transducers, for example via the
application of Common Spatial Pattern filters to neural signals filtered in specific
frequency bands [Wu et al., 2008].
2.3.1.5 Other features
The use of alternative features like fractal dimension, Horjth parameters or temporal
sequence modelling has been proposed in EEG-based studies [Boostani and Moradi,
2004] [Boostani et al., 2007] [Vidaurre et al., 2009] [Coyle et al., 2005].
Amplitude coupling between a pair of channels has also been used in BCI studies
[Wei et al., 2006] [Wei et al., 2007] [Krusienski et al., 2012]. The Magnitude-Squared
Coherence features between channels k and l used in [Krusienski et al., 2012] were
2
|Pkl (f )|
defined as Pkk (f )Pll (f ) , where Pkk (f ) is the power spectral density and Pkl (f ) is the
cross power spectral density between channels k and l. When compared to classic
band-power features for the task of motor imagery classification, they did not permit
to improve classification accuracy.
2.3.2 Features for effector control
Both discrete and continuous dependent variables can be extracted from neural
signals to control prosthesis and orthosis movements.
Continuous dependent variables Continuous variables traditionally consist

of position and/or velocity of the effector’s endpoint, e.g. the wrist kinematic
parameters in the case of an upper-limb orthosis [Li, 2014], or of the angular
characteristics of effector joints [Ajiboye et al., 2012]. Wrist speed and acceleration
[Hammer et al., 2013] [Hammer et al., 2016], force profil [Carmena et al., 2003]
[Chen et al., 2014a] and muscular activity [Carmena et al., 2003] [Koike et al., 2006]
[Choi et al., 2009] [Shin et al., 2012] have nevertheless been reconstructed in offline
preliminary studies. The principal components of the effector’s position or velocity
have also been estimated from neural signals in offline studies [Acharya et al., 2010]
[Wong et al., 2013] [Hotson et al., 2014].
Discrete dependent variables Discrete variables in particular include the di-

rection of the effector’s movement [Hortal et al., 2015] [Bhattacharyya et al., 2015],
the finger of interest [Hotson et al., 2016] or the open/closed state in the case of
hand prostheses or orthoses [Pfurtscheller et al., 2000]. Binary dependent variables
are also regularly used to characterize the state of the user during asynchronous
decoding, i.e. an Intentional Control (IC) or Non-Control (NC) state [Mason and
Birch, 2000] [Müller-Putz et al., 2010]. Binary decoding has for example been
considered for EEG-based 1D virtual navigation [Leeb et al., 2007] and for the
restoration of walking patients with tetraplegia by means of Functional Electrical
Stimulation [King et al., 2015].
It should finally be noticed that the extraction of such discrete or continuous
variables from neural signals can be complemented by a re-mapping step performed
by the user [Marathe and Taylor, 2011]. It has been shown that users are able to
learn how to control the velocity of an effector even when position parameters are
extracted from their neural signals, i.e. that position decoding is compatible with
velocity control [Marathe and Taylor, 2011]. While velocity control is generally
easier than position control with an equivalent degree of accuracy [Marathe and
Taylor, 2011], satisfying extraction of velocity parameters from neural feature is not
systematically possible (see, e.g., [Marathe and Taylor, 2013]). A combination of
position decoding, velocity control and user’s position-to-velocity remapping is liable
to result in an improved control performance. In [Marathe and Taylor, 2013] for
example, position parameters were extracted from monkeys’ neural signals, and they
were able to learn a position-to-velocity map to control the velocity of 2D cursor
[Marathe and Taylor, 2013].
2.3.3 Dimensionality reduction

High dimensional and/or correlated features are liable to disrupt decoders training.
Reduction of the independent variable dimension is mainly performed via projection
or feature selection methods.
2.3.3.1 Projection methods

Projection methods are utilized to reduce the dimension of the feature space by
projecting the features xt ∈ Rm onto a subspace of lower dimension F , F < m.
Principal Component Analysis and Partial-Least-Squares are examples of projection
methods which are popular within the BCI community.
PCA [Bishop, 2006] is widely used for unsupervised dimension reduction of neural
signals, for example MUA/SUA [Wu et al., 2003b] [Aggarwal et al., 2008] [Kao et al.,
2013] [Kao et al., 2017], ECoG [Wang et al., 2009b] and EEG signals [Ke and Li,
2009] [Suk and Lee, 2010] [Argunşah and Çetin, 2010] [Bhattacharyya et al., 2011].
PCA is an orthogonal projection of the features onto a subspace which maximizes
the variance of the projected features [Bishop, 2006]. This low-dimensional subspace
is spanned by the first F eigenvectors of the feature covariance matrix [Bishop, 2006].
The use of PCA’s variants has also been considered in BCI studies, e.g. nonlinear
PCA for motor imagery classification from EEG signals [Devulapalli, 1996].
In contrast with PCA, Partial-Least-Squares (PLS) [Höskuldsson, 1988] per-
mit to perform supervised dimensionality reduction. Data are projected onto a
low dimensional subspace which maximizes the covariance between the respective
projections of the independent and dependent variables. PLS-based regression has
for example been performed for efficient trajectory decoding from ECoG signals
[Shimoda et al., 2012] [Eliseyev and Aksenova, 2014] [Eliseyev and Aksenova, 2016]
[Bundy et al., 2016] [van Gerven et al., 2012]. The use of a combined PLS-PCA
projection strategy has also been reported for SUA/MUA decoding [Kim et al.,
2006c].
Both PCA and PLS are parametrized by the dimension F of the reduced feature
subspace. F was chosen so as to maximize decoding accuracy on a validation data
set in [Argunşah and Çetin, 2010]. Similarly, a cross-validation approach was utilized
in [Kim et al., 2006c] [Shimoda et al., 2012] [Eliseyev and Aksenova, 2016] [Bundy
et al., 2016]. Another approach consists in choosing F so as to retain a specific
percentage of variance, e.g. 95% [Aggarwal et al., 2008] or 85% [Suk and Lee, 2010].
2.3.3.2 Feature selection methods

Whereas projection methods rely on the creation a set of F new informative features,
feature selection methods permit to extract a subset of F existing features from the
full feature set (m features). Different feature selection methods have been utilized
in BCI studies.
A naive approach consists in computing a model for each possible feature
subset, and in choosing the best subset with respect to the corresponding decoding
performance on a validation data set. Probably because exhaustive search is liable to

be time-consuming, its use has not been reported for feature selection in BCI studies.
Several heuristic methods were, however, exploited for dimensionality reduction
in BCI studies [McFarland et al., 2010] [Wang et al., 2012] [Liang and Bougrain,
2012] [Hotson et al., 2014] [Wang et al., 2015] [Kelly et al., 2012] [Fazli et al., 2011]
[Fatourechi et al., 2007b] [Garrett et al., 2003] [Graimann et al., 2004].
Wrapper methods have been used to select informative features in both EEG-
and ECoG-based motor BCIs [McFarland et al., 2010] [Wang et al., 2012] [Liang
and Bougrain, 2012] [Hotson et al., 2014]. They consist in fitting a model on a
particular feature subset, and in monitoring the corresponding decoding accuracy
on an independent data set. In contrast with exhaustive search, all possible feature
subsets are not considered. Heuristics approaches, e.g. stepwise regression, are
utilized to iteratively approach the optimal subset. In the case of the stepwise
forward selection procedure used in [Wang et al., 2012] [Liang and Bougrain, 2012]
[Hotson et al., 2014], the feature subset is initially empty. The feature associated
with the highest decoding accuracy on a validation data set is chosen. At each
iteration, the feature which most increases the prediction accuracy is added to the
growing feature subset. Another stepwise selection procedure has been utilized for
EEG neural feature selection in [McFarland et al., 2010]. Forward selection was
combined with backward selection, i.e. features which had become useless after the
addition of a new feature were regularly removed from the feature set [McFarland
et al., 2010].
Feature selection embedded in the model identification procedure has been
utilized in several BCI studies. Sparse models were for example obtained by LASSO
training (Least Absolute Shrinkage and Selection Operator) in [Wang et al., 2015],
[Kelly et al., 2012] and [Fazli et al., 2011].
The use of so-called filter methods has also been reported in several closed-loop
[Schalk et al., 2007] [LaFleur et al., 2013] and offline [Spüler et al., 2016] motor BCI
studies. In contrast with wrappers methods which are computationally expensive,
filter methods do not require to train more than one model. A criterion between
feature and dependent variable, e.g. correlation, is considered. Features associated
with the highest criterion values are selected a priori.
The use of Genetic Algorithms has been reported in several EEG- [Flotzinger
et al., 1994] [Fatourechi et al., 2007b] [Garrett et al., 2003] [Graimann et al., 2004]
[Boostani et al., 2007] and ECoG-based BCIs [Wei et al., 2006]. Genetic Algorithms
are inspired from Darwin’s natural selection. Different initial features subsets are
considered. The most promising subsets (with respect to some criterion which
assesses decoding accuracy on the validation data set) are combined using random
processes so as to create a new subset generation.
Alternative feature selection approaches, e.g. Distinctive Sensitive Learning
Vector Quantization [Flotzinger et al., 1994] [Scherer et al., 2008] or selection based
on the Davies–Bouldin index [Wissel et al., 2013], have additionally been used in
BCI studies.
2.4. Decoders for motor BCIs 49
2.4 Decoders for motor BCIs

Different technical solutions have been considered to design and train BCI decoders,
for example for accurate kinematic decoding or asynchronous mental-task discrimina-
tion. Let xt ∈ Rm be an independent, input variable and yt ∈ Rn or y t ∈ Z denote a
dependent, output variable. Let us assume the existence of a model f , f : Rm → Rn
or f : Rm → Z, such that yt ≈ f (xt ). Let fˆ be an estimate of the unknown model f .
When motor BCIs rely on the decoding of continuous variables y t ∈ Rn [Hochberg
et al., 2012] [Collinger et al., 2013] [Wodlinger et al., 2015], the corresponding
fˆ : Rm → Rn is referred to as continuous decoder (a regression model for example).
Such continuous decoders are typically used to build kinematic decoders, mainly
in synchronous frameworks. In the case of discrete dependent variables, a discrete
decoder fˆ : Rm → Z (classifier) is applied on neural features [Sing et al., 2007]
[Yanagisawa et al., 2012b] [Hotson et al., 2016]. Discrete decoders are generally used
for the task of mental-task-based accurate mono-limb or multi-limb control. Their
utilization is also regularly considered to embed NC support into both mental-task
and kinematic decoders.
2.4.1 Generalities on decoder training

Most motor BCIs rely on user-specific decoders with data-driven training. Machine
learning methods are used to build a relevant decoder fˆ to model the dependence
between neural features xt and user intentions yt .
Decoder training consists in building an estimate fˆ of the unknown function f .

First, the set of admissible models G is restricted by making hypotheses about the
function f (e.g., linearity) and/or about the noise distribution. The estimation of the
optimal value of a set Θ of parameters is generally necessary to fully characterize each
admissible model fG ∈ G. The parameters of admissible models fG ∈ G are therefore
) *T ) *T
identified on "training" samples (X, Y), where X = xt t=1 and Y = yt t=1 . Each
) *T î ïT
sought-after model is trained so that the estimates Ŷ = ŷt t=1 = fˆG (xt ) are
t=1
optimal
. .with respect to a specific criterion, for example the Mean Squared Error
.Ŷ − Y. , where ë.ë2 is the ℓ2 -norm. Model selection is required when more than
. .
2
one admissible model is considered. Model selection procedures usually aim at
choosing the decoder associated with the best generalization performance, i.e. with
the most accurate estimates ŷtnew = fˆG (xtnew ) for samples xtnew , ytnew not used
) *
for training. The performance of a decoder on training samples is known to be

an optimistic estimate of its generalization performance (for example evaluated by
means of the test error) [Friedman et al., 2001]. For this reason, admissible models
are generally not directly compared on the basis of their performance on training
samples. Two main approaches are used to estimate the test error associated with
a model [Friedman et al., 2001]. Model selection criteria, for example Mallow’s
Cp or information criteria like the Bayesian Information Criterion and the Akaike
Information Criterion, estimate the test error from the training error, i.e. they
attempt to correct the optimism of the training error [Friedman et al., 2001]. The
model which minimizes the considered information criterion is selected. By contrast,
Cross-Validation (CV) is based on a direct estimation of the test error. The training
data set is typically divided into k folds. Models are trained on (k-1) folds, and
tested on the remaining fold [Friedman et al., 2001]. The procedure is repeated
with all possible combinations of (k-1) folds. The model with the best average
test performance is selected, and re-trained on the full training data set. Whereas
model selection criteria are applicable to models which are linear in their parameters
[Friedman et al., 2001], CV approaches can be used for all classes of models. Their
main shortcoming is that they involve repeated model identification procedures, and
are therefore computationally expensive.
2.4.2 Discrete decoding - classifiers

Let us consider the case of a discrete-valued dependent variable y t ∈ N. y t is
usually referred to as "class label". The associated observation xt is said to belong
to the "class" identified by its label. The classification of neural patterns, e.g. the
discrimination between different mental tasks such as motor imageries or NC (idle)
states [LaFleur et al., 2013] [Fifer et al., 2014], has been the basis of several EEG-
and ECoG-driven motor BCIs [Sing et al., 2007] [Yanagisawa et al., 2012b] [Hortal
et al., 2015] [Hotson et al., 2016]. Various classifiers have been embedded into motor
BCI transducers.
2.4.2.1 Generative and discriminative classifiers
Both generative and discriminative classifiers have been used in EEG- [Chae et al.,
2012] [Hortal et al., 2015] and ECoG-based [Yanagisawa et al., 2012b] [Fifer et al.,
2014] [Kapeller et al., 2015] motor BCIs.
Generative classifiers The use of generative classifiers has been reported in both
EEG- [Sing et al., 2007] [Pfurtscheller et al., 2000] [Chae et al., 2012] [Vidaurre et al.,
2016] and ECoG-based motor BCIs [Fifer et al., 2014] [Kapeller et al., 2015] [Hotson
et al., 2016]. The relevance of generative classifiers has additionally been investigated
in offline EEG- [Chiappa and Bengio, 2003] [Hasan and Gan, 2009] [Bhattacharyya
et al., 2011], ECoG- [Wang et al., 2016] and MUA/SUA-based [Hatsopoulos et al.,
2004] preliminary studies, e.g. for the classification of real movements [Hatsopoulos
et al., 2004] [Wang et al., 2016] or of mentals tasks [Chiappa and Bengio, 2003]
[Hasan and Gan, 2009] [Bhattacharyya et al., 2011].
Generative classifiers model the way independent variables are generated within
a class i, i.e. they model the joint probability P (xt , y t = i) [Ng and Jordan, 2002].
Once the joint probability has been fitted for each class, the classification of a
new observation sample xt is performed by computing the posterior probability
P (y t = i|xt ) with respect to each class [Ng and Jordan, 2002]. Using Bayes rule, this
posterior probability is proportional to P (xt , y t = i) for class i [Ng and Jordan, 2002].
The most likely class label is assigned to the considered observation sample [Ng and
Jordan, 2002]. The characteristics of the decision boundary (i.e., the equiprobable
hypersurface defined by P (y t = i|xt ) = P (y t = j|xt )) are not explicitly chosen, but
result from the distribution used to model data generation within each class. The
majority of generative classifiers reported in motor BCIs or preliminary studies
relied on multivariate Gaussian distributions (e.g., [Bhattacharyya et al., 2011]
[Lemm et al., 2004] [Wang et al., 2016]) or Gaussian Mixtures Models [Chiappa and
Bengio, 2003] [Hasan and Gan, 2009], i.e. P (xt |y t = i) = N (µi , Σi ) or P (xt |y t =
i) = k N (µk,i , Σk,i ). By contrast, the MUA/SUA firing rates of two NHPs were
q
modelled by Poisson distributions in [Hatsopoulos et al., 2004].

High-dimensional neural features are frequently considered in motor BCIs [Kim
et al., 2011] [Bhattacharyya et al., 2015] (up to 285 and 630 features, respectively).
In high dimension, fitting a multivariate distribution is impractical [Fan et al., 2011].
Several teams have therefore investigated the application of naive Bayes classifiers,
for left and right hand movement classification from more than 800 EEG features in
[Bhattacharyya et al., 2011], and to classify 8 wrist directions on the basis of the
firing rate of 32 to 143 neurons in [Hatsopoulos et al., 2004]. Naive Bayes classifiers
assume that features are independent conditionally to the class. The modelling
of the corresponding conditional probability is simplified, and classifier training is
consequently facilitated. Its potential advantages over traditional Bayes classifiers
have particularly been illustrated in [Bhattacharyya et al., 2011], where a naive
Bayes classifier surpassed a Gaussian-based generative classifier for the decoding
of EEG signals, both when a high-dimensional EEG feature set (871 features) was
used and when its dimensionality had been preliminary reduced (91 features).
Finally, the training of generative classifiers can be performed on partially labelled
training data sets [Sutton and Mccallum, 2012], e.g. for EEG features unsupervised
classification [Hasan and Gan, 2011].
Discriminative classifiers Discriminative classifiers have been utilized in both

EEG- and ECoG-driven motor BCIs, e.g. for EEG-based displacement of a robotic
arm’s endpoint [Hortal et al., 2015] or for ECoG-based control of a prosthetic arm
[Yanagisawa et al., 2012b]. Their performance has also been assessed in offline
preliminary studies, i.e. for the classification of motor imageries [Schlögl et al., 2005]
[Hill et al., 2006] [Chin et al., 2007].
Discriminative classifiers directly model the posterior class probability P (y t =
i|x ) [Bishop, 2006]. In contrast with generative classifiers, the intermediary proba-
t
bility P (xt |y t = i) is not modelled [Ng and Jordan, 2002]. This makes the use of
discriminative classifiers advantageous when this distribution cannot be approxi-
mated with classical distributions. Discriminative classifiers are particularly relevant
when xt is high-dimensional or includes redundant (correlated) neural features
[Sutton and Mccallum, 2012], because non-discriminant features are not considered
during model training. Support-Vector-Machine (SVM) classifiers [Lal et al., 2005]
[Hill et al., 2006] [Yanagisawa et al., 2012a][Schlögl et al., 2005] [Sadeghian and
Moradi, 2007] [Bhattacharyya et al., 2011] [Hortal et al., 2015], Logistic Regression
(LR) [Tomioka et al., 2007] [Bashashati et al., 2015] [Chen et al., 2014b] [Bundy et al.,
2016], k-Nearest Neighbors (kNN) [Chin et al., 2007] [Kayikcioglu and Aydemir,
2010] and Artificial Neural Networks (ANN) [Haselsteiner and Pfurtscheller, 2000]
[Navarro et al., 2005] [Nakayama and Inagaki, 2006] [Hatsopoulos et al., 2004] are
some of the discriminative classifiers which use has been reported in motor BCIs
and/or considered in preliminary studies.
The superiority of generative or discriminative classifiers for mental states

classification cannot be established a priori. Their relevance and performance
depend on the problem at hand, in particular on the validity of the assumed
distributions in the case of generative classifiers, on the dimension of the neural
feature representation, on the availability of labelled and/or unlabelled training
samples etc. The reported popularity of decoders within the BCI community is
consistent with this observation, as Linear Discriminant Analysis (LDA) and SVM,
which are both widely used in BCI studies [Bashashati et al., 2015] [Nicolas-Alonso
and Gomez-Gil, 2012], are a generative and discriminative classifier, respectively.
Generative and discriminative classifiers can be associated with linear or non-
linear decision boundaries.
2.4.2.2 Linear and non-linear classifiers
The comparative relevance of linear and non-linear classifiers has been frequently
investigated in preliminary studies focusing on the recognition of mental tasks from
neural signals (e.g., [Müller et al., 2003] [Garrett et al., 2003]).
Linear classifiers Different linear classifiers have been applied and tested for
online and offline mental states recognition, for example multiple Intentional Control
and No-Control states.
Linear Discriminant Analysis (LDA) classifiers are generative classifiers based on
multivariate Gaussian distributions. Their linear decision boundary [Bishop, 2006]
is obtained by using the same covariance matrix for all classes. LDA classifiers have
been embedded in several motor BCIs, e.g. in BCIs providing users with control over
hand prostheses or orthoses [Pfurtscheller et al., 2000] [Fifer et al., 2014] [Hotson
et al., 2016], lower-limb orthoses [Vidaurre et al., 2016] or humanoid robots [Kapeller
et al., 2015]. LDA has also been used for offline motor imagery classification in EEG
[Bhattacharyya et al., 2011], and for cognitive state estimation in LFP [Aggarwal
et al., 2013] and MUA/SUA [Velliste et al., 2014] signals.
Another linear classifier, namely the Support-Vector-Machine (SVM) classifier,
has frequently been applied in motor BCI studies [Schlögl et al., 2005] [Yanagisawa
et al., 2012b] [Hortal et al., 2015] The SVM’s linear decision boundary is chosen
so as to maximize its margin with the nearest training samples ("support-vectors")
[Bishop, 2006]. SVM-based classification has for example enabled users to control a
prosthetic hand [Yanagisawa et al., 2012b] and a robotic arm [Hortal et al., 2015].
SVMs are also regularly used for offline motor imagery classification in ECoG [Lal
et al., 2005] [Hill et al., 2006] [Demirer et al., 2009] [Yanagisawa et al., 2012a] and
EEG [Schlögl et al., 2005] [Sadeghian and Moradi, 2007] [Bhattacharyya et al., 2011]
signals. SVMs are attractive for neural signal decoding [Lotte et al., 2007] because
of their good generalization abilities [Schlögl et al., 2005] and of their robustness in
high-dimensional settings [Friedman et al., 2001].
The application of a threshold on the output of a linear regression model has
been reported in motor BCI studies, e.g. for ECoG-driven asynchronous 2D cursor
control [Williams et al., 2013]. This classification approach has also been considered
in offline asynchronous studies [Eliseyev et al., 2011] [Eliseyev et al., 2012].
Logistic Regression (LR) is a discriminant classifier based on Generalized Linear
Models, which extend linear models in that a non-linear link function g is applied on a
linear combination of features [Bishop, 2006]. In contrast with linear regression-based
classifiers, LR considers a discrete dependent variable and assumes that P (y t |xt )
follows a Bernoulli distribution. Although LR has rarely been used in closed-loop
studies [Penny et al., 2000], several teams have investigated its relevance for the
distinction betwen mental tasks from EEG [Tomioka et al., 2007] [Gouy-Pailler
et al., 2009] [Bashashati et al., 2015] and ECoG signals [Chen et al., 2014b] [Bundy
et al., 2016].
The previously mentioned classifiers are based on Euclidian distances, i.e. on
the ℓ2 -norm. Classifiers based on the Mahalanobis distance have been applied in
several offline BCI studies, e.g. in [Cincotti et al., 2003] [Sadeghian and Moradi,
2007] [Bai et al., 2011] [Schlögl et al., 2005]. Efficient classification of EEG patterns
has additionally been achieved by exploiting Riemannian distance [Barachant et al.,
2010] [Barachant et al., 2012].
LDA has been regularly used to provide users with neural control over prostheses,
orthoses and robotic devices, and it is particularly popular for EEG offline linear
classification [Bashashati et al., 2015]. No clear superiority of LDA decoding
performance has, however, been reported in offline comparative studies [Schlögl
et al., 2005] [Wang et al., 2009b] [Bashashati et al., 2015]. In [Bashashati et al., 2015],
a LDA classifier was slightly but not significantly surpassed by a LR-based classifier
for asynchronous and synchronous EEG decoding . In [Wang et al., 2009a], LDA and
SVM performed similarly for both motor imagery and finger movement classification
from EEG signals. In another comparative study [Schlögl et al., 2005], LDA was
significantly outperformed by a SVM for 4-class motor imagery classification in
EEG signals. By contrast, it performed better than a SVM when applied on low-
dimensional EEG features in [Bhattacharyya et al., 2011]. Its comparatively low
robustness in high dimensions was also illustrated in the same study [Bhattacharyya
et al., 2011], as LDA performance diminished when the dimension of the EEG
features had not been reduced beforehand [Bhattacharyya et al., 2011].
Non-linear classifiers Non-linear classifiers have mainly be applied in offline

preliminary studies [Schlögl et al., 2005] [Navarro et al., 2005] [Nakayama and
Inagaki, 2006] [Bhattacharyya et al., 2011], for example to discriminate between

mental tasks associated with different Intentional Control states [Schlögl et al., 2005]
[Bhattacharyya et al., 2011] or between No-Control and Intentional Control states
[Mason and Birch, 2000].
Several teams have reported EEG mental task classification by means of Quadratic
Discriminant Analysis (QDA) classifiers, i.e. Gaussian-based generative classifiers
with class-specific covariance matrices [Schlögl et al., 2005] [Bhattacharyya et al.,
2011].
The use of non-linear SVM has been investigated for EEG feature classification
in [Bhattacharyya et al., 2011] and [Bashashati et al., 2015]. Non-linear SVMs were
designed by means of non-linear kernels, typically Radial Basis Functions (RBF), in
[Bhattacharyya et al., 2011] [Bashashati et al., 2015].
Artificial Neural Networks (ANNs) have been used for the offline, non-linear
classification of mentals states [Haselsteiner and Pfurtscheller, 2000] [Mahmoudi
and Erfanian, 2002] [Navarro et al., 2005] [Nakayama and Inagaki, 2006] or real
movements [Hatsopoulos et al., 2004]. ANNs attempt to mimic information encod-
ing in biological neuron networks [Bishop, 2006] by applying cascaded non-linear
functions on weighted combinations of features xti , resulting in a highly non-linear
model [Bishop, 2006]. The flexibility of ANNs makes them attractive for the complex
problem of neural signal modelling. It has nevertheless been reported that they can
suffer from a few shortcomings, namely difficulties to select the optimal network
architecture, to avoid overfitting [Kayikcioglu and Aydemir, 2010], and to interpret
results. As a result, it has been observed in comparative studies that the accuracy
of ANN-based mental task classification is not systematically better than the one
obtained with simple non-linear models [Wang et al., 2009a] [Garrett et al., 2003].
In [Kayikcioglu and Aydemir, 2010], an ANN was outperformed by a non-linear
SVM for different training data set sizes. In [Garrett et al., 2003], where non-linear
SVMs and ANNs were compared for a 5-class discrimination task with EEG signals,
the ANN was bettered by the SVM. Similar results were obtained on two EEG data
sets in [Wang et al., 2009a].
Finally, the use of the k-Nearest Neighbors (kNN) classifier has been investigated
for offline detection of mental tasks from EEG features [Mason and Birch, 2000]
[Schlögl et al., 2005] [Wang et al., 2009b] [Kayikcioglu and Aydemir, 2010] [Bhat-
tacharyya et al., 2011], and has been applied for real movement classification from
ECoG data in [Chin et al., 2007]. In contrast with previously reported classifiers,
the kNN classifier is not parametric. A new sample is assigned with the label which
is the most represented among its k nearest training samples [Bishop, 2006]. Thus,
kNNs don’t require a time-consuming training procedure to be completed before
their application. A high computational load can, however, be associated with kNN
application, which is based on the computation of the distance between a new sample
and the training samples. The computation of these distances is computationally
expensive when large training data sets are necessary to properly model the data
structure. This shortcoming may limit its applicability for motor BCIs relying on
high-dimensional features, as online kNN-based classification may introduce a large
delay in the system. Mixed results have been reported in the comparative studies
completed on non-linear classifiers [Wang et al., 2009a] [Kayikcioglu and Aydemir,
2010]. In [Wang et al., 2009a], a kNN performed similarly to a SVM with a RBF
kernel for the discrimination between EEG motor imageries, and was only slightly
surpassed by the same SVM-based classifier for finger movement decoding from
EEG signals. This satisfying performance was obtained with low-dimensional input
features (respectively of 2 and 14). In [Kayikcioglu and Aydemir, 2010], a similar
comparison was drawn between a kNN, a RBF-based SVM and an ANN for 2-class
classification in the context of EEG-based up-down neural control of a cursor. The
kNN outperformed both the MLP and SVM for this specific classification task, and
its performance was best maintained when the researchers attempted to reduce
the training data set size. The comparison was nevertheless performed in a setting
particularly favourable to the kNN, as the input features were only of dimension
two. By contrast, in [Bhattacharyya et al., 2011], the kNN was outperformed by
a RBF-based SVM for two sizes of independent variable (namely, 871 and 91 features).
The respective advantages and limits of the above mentioned classifiers are
unclear. First, most of them have not been used for online pattern classification.
Additionally, offline comparisons have generally been completed for two or three
classifiers only, and the statistical significance of the results has seldom been estab-
lished. A few studies have nevertheless endeavoured to assess the relative interest of
linear and non-linear classifiers for offline discrimination between neural patterns
[Wang et al., 2009a] [Bhattacharyya et al., 2011] [Bashashati et al., 2015].
Linear models exhibit a lesser modelling ability. In [Bhattacharyya et al., 2011],
a RBF-based SVM was found to outperform a linear SVM as well as the other linear
classifiers implemented in this study. The same result were obtained in [Wang et al.,
2009a], where a linear SVM was bettered by a RBF-based SVM for both motor
imageries and finger movement classification.
The superiority of non-linear classifiers has not systematically been reported
in BCI preliminary studies. For example, QDA did not outperform LDA in two
comparative studies [Wang et al., 2009a] [Bhattacharyya et al., 2011]. In [Garrett
et al., 2003], a LDA classifier was compared to non-linear SVMs and to an ANN for
the classification of 5 mental tasks. The performance of the non-linear classifiers
was found to be only slightly better than the LDA’s one for this EEG classification
task. Training pitfalls were illustrated in [Schlögl et al., 2005], where a kNN was
significantly outperformed by a linear SVM and by LDA for 4-class motor imagery
discrimination from EEG signals. For this reason, it has been advocated in [Müller
et al., 2003] to use linear methods except for some specific cases with "complex,
large" data sets. Correspondingly, linear classifiers like LDA [Bashashati et al., 2015]
are regularly chosen over non-linear models despite their lesser modelling ability.
In particular, most recent clinical motor BCIs have relied on linear classifiers, e.g.
LDA [Sing et al., 2007] [Fifer et al., 2014] [Kapeller et al., 2015] [Hotson et al., 2016]
[Vidaurre et al., 2016] or SVM [Hortal et al., 2015].
Despite some trends, the relevance of a linear or non-linear classifier ultimately
depends on the problem at hand. In [Bashashati et al., 2015] for example, the two top
classifiers for self-paced data decoding were a linear and a non-linear classifier, namely
a LR classifier and an ANN. Classifier performance is in particular related to the
characteristics of the extracted neural features, for example to their type [Bashashati
et al., 2015] or dimension [Bhattacharyya et al., 2011]. In [Bashashati et al., 2015],
classifiers’ performance for synchronous data decoding was not similar when classifiers
were fed with band-pass- or with wavelet-based features. In [Bhattacharyya et al.,
2011], differences in performance ranking were observed if classifiers were applied
on a high-dimensional input variable or on the same variable after PCA-based
dimensionality reduction. Similarly, a RBF SVM and a kNN were identified as the
best classifiers for a task of motor imagery decoding in [Wang et al., 2009a], but
LDA reportedly equalled a RBF-based SVM for finger decoding in the same paper.
Finally, to the best of our knowledge, only limited comparative studies have
been completed on ECoG data [Shenoy et al., 2008]. The respective relevance of the
above-mentioned classifiers thus remains to be ascertained for ECoG data.
2.4.2.3 Static and sequential classifiers
The previously mentioned classifiers are static, i.e. they don’t take into account pos-
sible dependencies between successive independent or dependent variables. Formally,
they assume that pairs (xt , y t ) are temporally independent and identically drawn
from the distribution P (xt , y t ) [Dietterich, 2009].
This assumption is typically violated in motor BCI studies. A few teams have
therefore investigated the interest of taking into account the sequential nature of the
independent or dependent variable [Obermaier et al., 2001] [Chiappa and Bengio,
2003] [Argunşah and Çetin, 2010]. One strategy regularly utilized in BCI studies, e.g.
[Kim et al., 2011] [Flamary and Rakotomamonjy, 2012] [Eliseyev et al., 2012], con-
sists in extracting features from several time segments to build a temporal sequence
of feature vectors. This sequence is then fed to a static classifier [Lotte et al., 2007]
[Dietterich, 2009]. Another approach, namely the application of dynamic classifiers,
has been reported for neural pattern classification in SUA/MUA [Darmanjian et al.,
2003], EEG [Obermaier et al., 2001] [Argunşah and Çetin, 2010] and ECoG-based
[Onaran et al., 2011] [Saa et al., 2016] BCI studies. Dynamic classifiers directly
exploit time series temporal behaviour [Lotte et al., 2007].
Dynamical classification of neural signals has been performed by means of Hidden

Markov Models (HMMs) in EEG [Obermaier et al., 2001] [Gouy-Pailler et al., 2009]
[Argunşah and Çetin, 2010], ECoG [Onaran et al., 2011] and SUA/MUA preliminary
studies [Darmanjian et al., 2003] [Wissel et al., 2013], and in a few motor BCIs [Fifer
et al., 2014] [Hotson et al., 2016].
HMMs consider a hidden state z t ∈ Z which is generated by a first order Markov
process, i.e. such as P (z t+1 = k|z 1:t ) = P (z t+1 = k|z t ) [Rabiner, 1989]. The value
of the observation xt ∈ Zm or xt ∈ Rm depends on the corresponding hidden state
value z t via the conditional probability P (xt |z t ) [Rabiner, 1989]. Efficient recursive
algorithms permit to infer the most likely state label ẑ t by combining prior knowledge
about the previous hidden state z t−1 with the likelihood of the current observed
features xt [Rabiner, 1989].
One reported approach to HMM-based classification consists in associating one
hidden state value z t = i to each class label y t = i [Kemere et al., 2008] [Fifer et al.,
2014] [Hotson et al., 2016]. The Markovian hypothesis thus models the class label
succession. This strategy has for example been used for offline target estimation
from SUA/MUA signals in [Kemere et al., 2008]. HMM-based classifiers have also
been applied for robust online state detection in several closed-loop motor BCIs
[Fifer et al., 2014] [Hotson et al., 2016] [Kao et al., 2017]. States were for example
associated with NC and IC classes [Fifer et al., 2014] [Hotson et al., 2016].
An alternative approach has been investigated in offline preliminary studies [Ober-
maier et al., 2001] [Darmanjian et al., 2003] [Argunşah and Çetin, 2010] [Onaran
et al., 2011] [Wissel et al., 2013]. One HMM was associated with each considered
class, and several states were thus used to model feature dynamic within each class.
Classification was performed by feeding each HMM with a sequence of N consecutive
observations, and by computing the associated probability P (xt−N +1:t |y t = i). The
sequence was assigned the class i which maximized P (xt−N +1:t |y t = i). HMMs have
been used for offline modelling of the variations of neural features within NC and
IC states in SUA/MUA [Darmanjian et al., 2003] and ECoG [Onaran et al., 2011],
within finger movements in ECoG signals [Wissel et al., 2013], or within motor
imageries in EEG signals [Obermaier et al., 2001] [Argunşah and Çetin, 2010].
The use of several HMM’s variants has been proposed for the classification of
EEG and ECoG mental tasks [Chiappa and Bengio, 2003] [Hasan and Gan, 2010]
[Hasan and Gan, 2011] [Delgado Saa and Cetin, 2012] [Saa and Çetin, 2013].
Input-Output Hidden Markov Models (IOHMM) were applied on EEG signals to
discriminate between 3 mental tasks in [Chiappa and Bengio, 2003]. In contrast with
HMMs, IOHMMs are trained to distinguish between classes composed of several
hidden states, and directly map input features to the non-stationary classes [Bengio
and Frasconi, 1996].
Conditional Random Fields (CRFs) are discriminative undirected graphical
models [Sutton and Mccallum, 2012], and linear-chain CRFs are more particularly
the discriminative counterpart of HMMs [Sutton and Mccallum, 2012]. CRFs have
been used for EEG offline modelling and decoding [Hasan and Gan, 2010] [Hasan
and Gan, 2011] [Delgado Saa and Cetin, 2012] [Saa and Çetin, 2013], and for
finger movement detection in ECoG signals [Saa et al., 2016]. If they have a better
ability to model long-term time dependencies [Lafferty et al., 2001], their training is
computationally expensive [Dietterich, 2009].
Dynamic Bayesian Models (DBN) are probabilistic graphical models which permit
to take into account the dependence between several random variables [Murphy,
2002]. HMMs are a specific case of DBNs, and are therefore less flexible than DBNs.
The dynamic of EEG [Shenoy, 2005] and ECoG [Wang et al., 2012] signals has been
exploited by means of DBN [Murphy, 2002] in offline studies.
Finally, the use of a time-dependent ANN was reported for EEG dynamical
classification in [Haselsteiner and Pfurtscheller, 2000].
Most dynamic classifiers which were embedded in motor BCIs were generic
HMMs with state-class correspondence [Fifer et al., 2014] [Hotson et al., 2016]
[Kao et al., 2017]. While the respective performances of more complex dynamical
classifiers has been investigated and compared in offline studies [Chiappa and
Bengio, 2003] [Delgado Saa and Cetin, 2012] [Saa and Çetin, 2013], they have not
been assessed in closed-loop settings. In [Delgado Saa and Cetin, 2012], HMM
surpassed CRFs for the classification of EEG signals, but were outperformed by
a CRF variant, namely a hierarchical CRF. On the contrary, in [Saa and Çetin,
2013], HMM-based EEG classification accuracy was inferior to the CRF-based one.
In [Chiappa and Bengio, 2003], IOHMMs were found to outperform HMMs for
EEG dynamic classification. As both dynamic classifiers performed similarly to
their static counterparts (namely, a Gaussian Mixture Model-based Bayes classifier
and an ANN), the authors concluded on the superiority of ANN over Gaussian
Mixture Model-based generative classification for the considered EEG data set. It
has correspondingly been suggested that multi-state dynamic classifiers could be
suboptimal for asynchronous decoding [Lotte et al., 2007].
A few additional papers have investigated the respective relevance of static and
dynamic classifiers when applied to offline data, e.g. [Cincotti et al., 2003] [Saa
et al., 2016]. In [Cincotti et al., 2003], HMMs were significantly outperformed by
ANNs (and Mahalanobis Distance) for the classification of right and left hand motor
imageries in EEG signals. In [Saa et al., 2016], an extension of CRFs improved the
discrimination between finger movements from ECoG signals when compared to
LR and sparse linear regression. The interpretability of this result is nevertheless
limited, because the considered static and dynamic classifiers did not belong to
the same class of models, in contrast with the static-dynamic pairs compared in
[Chiappa and Bengio, 2003].
2.4.3 Continuous decoding
Let the dependent variable yt be continuously-valued, yt ∈ Rn . Decoding of

continuous dependent variables is mainly performed within the framework of MEA-
and ECoG-based motor BCI systems. Continuous dependent variables typically
characterize the position or velocity of the effector’s endpoint, e.g. the wrist kinetics
or kinematics in the case of an upper-limb orthosis [Li, 2014] (see section 2.3.2).
The use of different classes of models fˆ : Rm → Rn has been explored in BCI studies.
2.4.3.1 Linear and non-linear regression models
Both linear and non-linear regression models have been applied for kinematic
parameter reconstruction from neural signals.
Linear regression models Neural control over prostheses or cursors has been
achieved by means of linear models in several motor BCIs, both with human
[Hochberg et al., 2006] [Collinger et al., 2013] [Wodlinger et al., 2015] and primate
subjects [Taylor et al., 2002] [Carmena et al., 2003] [Velliste et al., 2008] [Sum-
inski et al., 2010] [Williams et al., 2013] [Willett et al., 2013]. Offline trajectory
reconstruction has also been performed by means of linear models in several EEG-,
ECoG- and MUA/SUA-based preliminary studies [Bradberry et al., 2010] [Liang
and Bougrain, 2012] [Eliseyev and Aksenova, 2014] [Bundy et al., 2016] [Koyama
et al., 2010b]. Linear models rely on the assumption that the dependent variable
is a (noisy) linear combination of the independent variable components, i.e. of the
neural features:
yt = Bxt + ǫt
where B ∈ Rn×m and ǫt ∈ Rn is the observation noise, and where the neural features
xt can embed a history of instantaneous neural features x̃t , i.e. xt = x̃(t+1−τ2 ):t .
Neural feature temporal concatenation is one of the approaches which permit to
exploit neural signal temporal characteristics [Dietterich, 2009] [Lotte et al., 2007].
A particular linear model, namely the Population Vector Algorithm (PVA), has
more specifically been used for kinematic decoding in several MEA-driven motor BCI
systems [Taylor et al., 2002] [Velliste et al., 2008] [Collinger et al., 2013] [Wodlinger
et al., 2015]. The PVA is based on the cosine directional tuning model [Georgopoulos
et al., 1986], which states that neurons of the motor cortex fire preferentially in one
specific direction. The instantaneous firing rate of each neuron is used to weight the
corresponding preferred direction.
The use of different identification algorithms has been reported in motor BCIs
and offline reconstruction studies. Linear regression models have frequently been
trained using Ordinary Least Squares (OLS) in BCI studies [Li et al., 2009]. While
OLS corresponds to the Maximum Likelihood estimator when the measurement
noise is Gaussian, the OLS estimator is unstable when the input variable xt is high
dimensional or composed of correlated explanatory features [Friedman et al., 2001].
The use of penalized approaches such as pace regression [Kubánek et al., 2009], ridge
regression [Willett et al., 2013] [Li et al., 2009] [Suminski et al., 2010] [Shanechi
et al., 2013] and sparse linear regression [Williams et al., 2013] has therefore been
proposed for model identification. Performant decoding from ECoG high dimensional
feature representations has additionally been reported using Partial Least Squares
and variants [Shimoda et al., 2012] [Eliseyev and Aksenova, 2014] [Eliseyev and
Aksenova, 2016] [Bundy et al., 2016] [van Gerven et al., 2012].
Non-linear regression models Linear regression models rely on simplistic as-

sumptions about information encoding in motor neural signals, which complexity
has been suggested in numerous studies [Scott, 2008]. Several teams have in partic-
ular investigated the use of non-linear models for neural signal decoding [Li, 2014],
assuming that yt = f (xt ) + ε with f non-linear. These studies mainly consisted of
offline trajectory reconstructions [Kim et al., 2006b] [Eliseyev and Aksenova, 2014]
[Spüler et al., 2016].
Generalized Linear Models (GLMs) constitute one class of non-linear models
which interest for offline trajectory reconstruction has been explored in BCI studies,
e.g. from ECoG signals in [Eliseyev and Aksenova, 2014]. GLMs extend linear
models by applying a non-linear function g −1 on the output of a linear filter β,
i.e. yt = g −1 (βxt ). The non-linear function g −1 , referred to as "link" function, is
chosen among standard functions (e.g., the logarithm) or fitted on training data
[Eliseyev and Aksenova, 2014]. A similar approach, namely a cascaded Wiener filter,
has been applied on SUA/MUA offline data sets in [Flint et al., 2012] and [Scheid
et al., 2013]. Generalized Additive Models (GAM) were alternatively utilized for
trajectory reconstruction from primate ECoG signals in [Eliseyev and Aksenova,
2014]. Generalized Additive Modelling consists in applying a non-linear function gi
on each component xit of the independent variable. A linear model β subsequently
combines the outputs of the non-linear functions gi (xit ). The application of non-linear
regression models such as Support Vector Machine Regression (SVR) [Kim et al.,
2006b] [Mehring et al., 2003] or ANN models [Sanchez et al., 2002] [Kim et al.,
2006c] [Kim et al., 2006b] [Hatsopoulos et al., 2004] has been additionally proposed
for SUA/MUA decoding, and tested in offline preliminary studies. SVR-based offline
trajectory reconstruction has also been reported in an ECoG-driven preliminary
BCI study [Spüler et al., 2016].
The findings of several offline preliminary studies are consistent with the idea
that non-linear regression models are likely to be more realistic than linear ones
for kinematic decoding: linear decoders were outperformed by both GLM and
GAM approaches for ECoG signal decoding in monkeys [Eliseyev and Aksenova,
2014] signals, and by SVR in simulated primate SUA/MUA signals [Kim et al.,
2006b]. Because non-linear models are more flexible than linear ones, and therefore
more prone to overfit, fine identification procedures were often required for non-
linear models proper training. The training of the GAM proposed in [Eliseyev and
Aksenova, 2014] relied on algorithms from the PLS family. Difficulties were reported
for the training of the ANN used in [Kim et al., 2006b] for trajectory decoding.
They were presented as a possible cause for the superior decoding performance
of the SVR, which is yet less flexible than ANNs [Kim et al., 2006b]. A specific
early-stopping procedure was utilized to prevent overfit during ANN training in
[Hatsopoulos et al., 2004]. The complexity of ANNs’ possible structures (e.g., number
of layers and number of neurons per layer) additionally makes their optimization
time-consuming, which is the reason why proper optimization of the ANN structure
was not performed in [Hatsopoulos et al., 2004]. Under these conditions, linear-
and ANN-based trajectory reconstructions from MUA/SUA signals yielded similar
results in this study [Hatsopoulos et al., 2004].
Thus, linear models are sometimes chosen over their non-linear counterparts
in spite of their simplistic assumptions, in particular in the case of MUA/SUA-
[Taylor et al., 2002] [Velliste et al., 2008] [Collinger et al., 2013] [Wodlinger et al.,
2015] and ECoG-driven [Schalk et al., 2008] [Wang et al., 2013c] motor clinical
BCIs. Up to 10D- and 3D-control has been achieved by means of linear filtering of
MUA/SUA and ECoG signals, respectively [Wodlinger et al., 2015] [Wang et al.,
2013c]. Linear models have additionally been shown to be reasonably efficient for
position, velocity, acceleration, speed etc. offline decoding [Wang et al., 2007] [Bundy
et al., 2016] [Hammer et al., 2016], and generally involve simpler training procedures
than non-linear models.
Since the last decade, another class of decoders, namely dynamic models, has
gained popularity in motor BCIs.
2.4.3.2 Dynamic models
Stochastic state-space models Linear or GAM-, GLM-, SVR- and ANN-based

decoders are static regression models, i.e. they assume the existence of a (parametric
or non-parametric) linear or non-linear model f so that yt ≈ f (xt ). By contrast,
the dynamic models utilized for cursor or prosthesis control in several motor BCIs
[Hochberg et al., 2012] [Ifft et al., 2013] consider stochastic state-space models, i.e.:
yt+1 = g(yt ) + wt , (2.1)

t t t
x = h(y ) + v . (2.2)
The noise processes wt and vt are generally independent and identically distributed
sequences of random variables [Krishnamurthy, 2016]. The continuous response
variable yt ∈ Rn is here composed by the trajectory coordinates and derivatives
(velocity, acceleration etc.). The transition equation (2.1) explicitly describes the
dynamic of the hidden sequence yt ∈ Rn ("movement model" [Li, 2014]). As expressed
in (2.1), movement models traditionally rely on first-order Markovian temporal
dependencies. Free or typical upper-limb movements like reaching movements
cannot be precisely characterized by a first-order model [Kim et al., 2006c]. For
this reason, movement models are generally limited to random-walk models which
permit to constrain the trajectory smoothness [Koyama et al., 2010b] [Brockwell
et al., 2004].
The dependence between measurements xt ∈ Rm and hidden state value yt ∈ Rn
is described by the emission equation (2.2), where vt is the observation noise. As
the emission equation models how neural features are generated conditionally to a
given trajectory point, state-space models are sometimes referred to as "generative
model" [Wu et al., 2002] [Gao et al., 2003] [Kim et al., 2006c].
Recursive Bayesian estimation procedures are generally used to infer the hidden
trajectory yt ∈ Rn from the sequence of noisy measurements xt ∈ Rm [Bishop, 2006].
Recursive Bayesian estimation: Kalman filter The Kalman Filter (KF) is

a recursive estimation procedure which has been frequently utilized for online and
offline trajectory reconstruction. It was first applied for 2D offline hand trajectory
decoding from SUA/MUA signals in monkeys [Wu et al., 2002] [Wu et al., 2003a],
where it was found to surpass linear filtering [Wu et al., 2003a]. It has since then
permitted to provide users with MUA/SUA-based control over prostheses [Hochberg
et al., 2012]. It has additionally been applied for trajectory decoding from ECoG
signals in online and offline studies [Pistohl et al., 2008] [Kellis et al., 2012] [Marathe
and Taylor, 2013] [Wang et al., 2013b]. KF applies to linear Gaussian state-space
models [Bishop, 2006]:
yt+1 = Ayt + wt , (2.3)

t t t
x = Cy + v . (2.4)
Here, the emission and transition models are linear, and the corresponding noises
are Gaussian: P (wt ) ∼ N (0, Γ), Γ ∈ Rn×n , P (vt ) ∼ N (0, Σ), Σ ∈ Rm×m .
After training, typically performed using Ordinary Least Squares [Wu et al.,
2002], the KF issues the estimate ŷt = E(yt |x1:t ).
Alternative recursive estimation procedures To the best of our knowledge,

dynamical modelling of ECoG data has been restricted to Gaussian state-space
models, i.e. Kalman Filtering procedures. However, further investigations have been
carried out to ascertain the interest of non-linear and/or non-Gaussian state-space
modelling of MUA/SUA data. If non-linear and/or non-Gaussian state-space repre-
sentations integrate more realistic emission and noise models (e.g., Poisson noise for
spiking counts), the associated trajectory estimation procedures are often approx-
imate and/or computationally expensive [Koyama et al., 2010a] (e.g., Unscented
Kalman Filter (UKF) in [Li et al., 2009] [Ifft et al., 2013], particle [Brockwell et al.,
2004], point-process or Laplace-Gaussian Filtering (LGF) [Velliste et al., 2014] in
the case of Poisson noise).
The relevance of non-linear emission models has in particular been studied for
MUA/SUA offline decoding in [Gao et al., 2003]. The emission model was modelled
using linear models, GLM or GAM associated with Poisson noises [Gao et al., 2003].
Non-linear models, and particularly GAM-based emission models, were found to
improve the quality of trajectory estimation. In [Koyama et al., 2010b], KF and the
LGF (i.e., procedure for non-linear emission models and Poisson noise) performed
similarly for offline trajectory reconstruction from primate SUA/MUA signals. An
additional closed-loop study suggested a slight superiority of the LGF over the KF
[Koyama et al., 2010b]. Analogously, the UKF proposed in [Li et al., 2009] sur-
passed traditional KF for a task of trajectory reconstruction from MUA/SUA signals.
These last few years, dynamical models have emerged as a promising and efficient
alternative to static (typically linear) models [Li, 2014] [Srinivasan et al., 2007]. Since
its first application in 2002, the Kalman filter and its variants have been increasingly
applied for both online and offline SUA/MUA decoding [Wu et al., 2002] [Hochberg
et al., 2012] [Aggarwal et al., 2013]. Because of the deterrent computational burden
of its variants, recursive estimation during closed-loop experiments has mainly been
2.5. Post-processing 63
achieved via Kalman Filtering [Li et al., 2009]. While the respective performance
of Wiener and Kalman filters appeared to depend on the decoding task at hand in
[Kim et al., 2006c], a few systematic comparisons of static and dynamic models have
given a steady ground to the popularity of the Kalman filter: static linear models
were outperformed by dynamical ones for open-loop [Aggarwal et al., 2009] and/or
closed-loop SUA/MUA decoding [Koyama et al., 2010b]. KF embedded approach for
smoothing has been identified as particularly efficient in both open- and closed-loop
studies [Koyama et al., 2010b].
By contrast, the respective suitability of static and dynamic continuous models
for ECoG decoding is still unclear. ECoG-based neural control has been achieved
by means of a linear decoder both in Non-Human Primates [Williams et al., 2013]
and human subjects [Wang et al., 2013c] (2D and 3D effector control, respectively).
The use of linear decoding has also been reported for ECoG offline trajectory
reconstruction [Schalk et al., 2007] [Chao et al., 2010] [Shimoda et al., 2012] [Liang
and Bougrain, 2012] [Nakanishi et al., 2013] [Williams et al., 2013] [Hammer et al.,
2013] [Wang et al., 2013c] [Hotson et al., 2014] [Bundy et al., 2016]. Up to 7DoF
have been reconstructed in offline feasibility studies led on Monkeys [Chao et al.,
2010] [Shimoda et al., 2012] and human subjects [Nakanishi et al., 2013] [Schalk
et al., 2007]. On the other hand, Kalman filtering has permitted to reconstruct 2D
kinematic parameters from ECoG signals [Pistohl et al., 2008] [Kellis et al., 2012]
[Wang et al., 2013b] [Marathe and Taylor, 2013]. In a comparative study performed
on ECoG data [Eliseyev and Aksenova, 2016], static models outperformed Kalman
Filtering for the reconstruction of kinematic parameters from high-dimensional
time-space-frequency feature representations. One reason for these findings could
be a lesser relevance of generative approaches for high-dimensional ECoG data. In
another study led on ECoG data [Marathe and Taylor, 2013], however, Kalman-based
cursor control was more precise than linear-decoder-based control.
2.5 Post-processing
Post-processing techniques rely on a priori knowledge about specific characteristics
of yt to improve the estimates ŷt . They can be used to improve the quality of both
discrete-valued and continuously-valued intention estimates, and their application
has for example been reported for asynchronous binary [King et al., 2015] and/or
kinematic decoding [Wang et al., 2011].
2.5.1 Discrete output

Post-processing is mainly used to take into account the a priori knowledge that
fast switches between classes are unlikely. Typical post-processing methods are
filtering of the classifier output [Mason and Birch, 2000][Millán and Mouriño, 2003]
[Bashashati et al., 2007b] [King et al., 2015], triggering a state transition after
successive identical state estimates only [Townsend et al., 2004] [Pfurtscheller et al.,
2010], or blocking state transitions for a predefined duration after a performed

transition [Townsend et al., 2004] [Pfurtscheller et al., 2010].
2.5.2 Continuous output

Post-processing of a continuous estimate ŷt generally aims at improving the smooth-
ness of the corresponding decoded trajectory. The use of temporal filtering has been
investigated in [Marathe and Taylor, 2015]. The improvement of other trajectory
characteristics requires the existence of an a priori trajectory model which is seldom
available. A complex a priori finger trajectory model based on a switching non-linear
dynamic model was built in [Wang et al., 2011]. In particular, this model integrated a
priori knowledge about the succession of rest (NC), flexion and extension states and
about the maximal amplitude of finger movements. The switching post-processing
model was applied on the output of a linear decoder fitted on both NC and IC
samples, and permitted both to support NC state and accurately decode multi-limb
trajectories [Wang et al., 2011].
2.6 Transducers for accurate asynchronous kinematic

decoding
Kinematic decoders are considered in the present doctoral work. Asynchronous
kinematic decoders are suited for clinical motor BCIs if they associate null-velocity
kinematic estimates with each limb during NC states. In the case of sequential
multi-limb decoding, it is additionally desirable for the estimated velocity of limbs
j, i Ó= j to be null when limb i is moving. Finally, decoding accuracy during IC
states is crucial for kinematic decoders to be profitable to patients.
Several transducers have been proposed to handle accurate asynchronous mono-
and/ multi-limb decoding. While they generally relied on generic pre-processing
and feature extraction approaches, they embedded specific decoders and/or post-
processing operators. These decoders and post-processing operators exploited and
possibly combined the algorithmic tools presented in the previous sections, for
example continuous decoders and/or discrete decoders.
2.6.1 Asynchronous mono-limb decoding

NC states were not supported in the majority of kinematic motor BCIs, which were
deployed using synchronous paradigms [Hochberg et al., 2006] [Wodlinger et al.,
2015] [Hochberg et al., 2012]. Although generic linear static and dynamic models
are favored for motor BCI kinematic decoders (see section 2.4.3), they usually fail
to output zero-velocity (neutral) estimates when they are used for asynchronous
decoding and are applied to NC states [Chao et al., 2010] [Shimoda et al., 2012]
[Velliste et al., 2014]. The development of specific algorithmic strategies is thus
required to integrate NC support into kinematic decoders, and thereby limit or
eliminate spurious movements during NC states. The use of different decoding
2.6. Transducers for accurate asynchronous kinematic decoding 65
strategies has been reported. While most of them were validated with offline
analyses, the integration of NC support into a kinematic decoder was nevertheless
utilized for asynchronous control of a robotic arm in monkeys in [Suway et al., 2013].
Decoder-based strategies A first approach to asynchronous decoding consists in

managing NC states at the level of the decoder. While the use of generic non-linear
models has been considered in a few studies (e.g., GLM or GAM [Eliseyev and
Aksenova, 2014]), the most popular decoding approach consists in switching between
continuous models. Switching models rely on a latent discrete variable to introduce
state-specific non-linearities into a generic continuous decoder [Srinivasan et al.,
2007] [Wood et al., 2005] [Bundy et al., 2016]. More specifically, the dependence
between xt and yt is conditioned on a discrete latent variable z t , i.e. the decoder
switches between models according to the value of z t . When these non-linearities are
associated with NC and IC states, i.e. z t = 0 for NC states and z t = 1 for IC states,
the resulting decoder is expected to correspond to limited spurious activations during
NC states. Both static [Williams et al., 2013] [Bundy et al., 2016] and dynamic
[Wood et al., 2005] [Srinivasan et al., 2007] switching models have been considered
for asynchronous mono-limb decoding.
Static switching models rely on both discrete and continuous algorithmic tools,
as a discrete decoder is used to infer the value of the switching state variable. In
[Bundy et al., 2016] for example, the output of a logistic regression was used to switch
between 2 linear models. One of these models was dedicated to kinematic decoding
during IC states, and the other one yielded neutral kinematic estimates during NC
states. While they were not presented as switching models, other decoders can be
formulated as switching models, for example the decoder used in [Williams et al.,
2013]. The latter applied a linear kinematic model on neural features when IC states
had been detected by a discrete decoder.
The use of dynamic switching models has also been proposed for the integration
of NC support into generic dynamical models [Wood et al., 2005] [Srinivasan et al.,
2007]. A Switching Kalman Filter and a Switching Particle Filter were implemented
and tested in [Srinivasan et al., 2007] and [Wood et al., 2005], respectively. The
value of the latent variable was used to switch between observation and/or transition
models. The transition model associated with NC states explicitly modelled the fact
that null-velocity estimates are expected during NC states.
Post-processing-based strategies A second approach to asynchronous mono-

limb decoding consists in using a single model to describe the dependence between
neural signals and kinematic parameters during both NC and IC states [Wang et al.,
2013b] [Velliste et al., 2014]. A binary discrete variable is explicitly decoded using
a discrete decoder. The output of the single kinematic model is overwritten with
null-velocity (neutral) estimates when a NC state is detected by the discrete decoder.
Post-processing-based integration of NC support into kinematic decoders has been
explored for both SUA/MUA [Velliste et al., 2014] [Aggarwal et al., 2013] and ECoG
signal (e.g., [Wang et al., 2013b]) decoding. Different pairs of discrete-continuous

decoders have been considered.
Kalman filters and (non-Gaussian) variants were gated by LDA in SUA signals
[Velliste et al., 2014] [Aggarwal et al., 2013] or Bayes classifier in ECoG signals
[Wang et al., 2013b]. A Kalman filter or one of its variants was continuously applied
on neural signals, and its output was possibly overwritten with neutral values when
an independent classifier detected a NC state [Aggarwal et al., 2013] [Wang et al.,
2013b] [Velliste et al., 2014].
Finally, a dynamic switching post-processing model permitted to integrate NC
support into a finger kinematic decoder in [Wang et al., 2011]. It was applied on
the output of a linear decoder fitted on both IC states and NC samples located near
movement onsets and offsets.
2.6.2 Asynchronous sequential multi-limb decoding

The design of continuous decoders for asynchronous sequential multi-limb neural
control has rarely been considered, and is not straightforward. Although one limb
only should be active at each time moment, the activation of one limb is liable to
result in residual movements of the other limbs. Such noisy outputs were for example
observed in [Nakanishi et al., 2014b], where the displacement of a finger resulted in
small-amplitude movements in the estimations of the other fingers’ position.
Decoder-based strategies A switching model was considered for sequential

asynchronous multi-finger decoding in [Flamary and Rakotomamonjy, 2012]. One
linear model was devoted to each finger, and applied when deemed appropriate
by a multi-class discrete decoder. The linear model associated with a finger was
exclusively trained on samples acquired during the movement of this specific finger.
Such switching models intrinsically prevent parallel activations as only one active
limb model is chosen at each instant.
Post-processing-based strategies While a post-processing approach has been

applied to improve the reconstruction of sequential multi-finger movements in [Wang
et al., 2011], the considered post-processing operator did not specifically addressed the
issue of sequential multi-limb decoding, as each finger was independently considered.
Thus, to the best of our knowledge, the use of post-processing approaches has not
yet been considered for asynchronous sequential multi-limb decoding.
2.6.3 Accurate decoding during IC states

Decoding accuracy during IC states is a typical objective for kinematic decoders.
Decoder-based strategies The improvement of decoding accuracy is often at-

tempted by using complex non-linear models, for example ANN [Kim et al., 2006c],
Support-Vector-Regression [Spüler et al., 2016] or Unscented Kalman Filters [Ifft
2.7. Conclusion 67
et al., 2013]. The use of static switching models has in particular been explored
for accurate trajectory decoding during active states [Kim et al., 2003]. In [Kim
et al., 2003] for example, a discrete latent variable permitted to switch between 10
linear models dedicated to different phases of a reaching movement, and led to an
increase of the decoding accuracy for a task of MEA-based trajectory reconstruc-
tion. Trajectory decoding from MEA signals was improved by combining reaching
movement models in [Kemere et al., 2004b] [Kemere et al., 2004a] [Yu et al., 2007].
Finally, the use of a switching state-space filter has been reported in [Wu et al.,
2004]. The objective lied in improving MEA decoding accuracy by using emission
models conditioned on an unspecified latent cognitive state.
Post-processing-based strategies Post-processing-based approach to the im-

provement of decoding accuracy are generally limited to the improvement of the
smoothness of the kinematic estimates. In [Wang et al., 2011] however, a complex
post-processing Bayes filter was used to enhance the estimation of the amplitude of
finger flexions and extensions.
2.7 Conclusion
In this chapter, algorithmic tools utilized by the BCI community to build transducers
for motor BCIs have been presented. The decoders used to extract discrete or
continuous variables from neural signals have in particular been introduced.
Asynchronous mono-limb kinematic decoding is one of the objectives of the
present doctoral work. Kinematic transducers rely on continuous decoders, e.g. static
linear models or dynamical models like the Kalman Filter. As generic continuous
decoders generally fail to issue zero-velocity estimates during No-Control (NC) states,
specific decoding and/or post-processing strategies have been proposed to integrate
NC support into kinematic decoders, i.e. to limit spurious system activations during
NC states. Two main approaches to mono-limb asynchronous decoding have emerged,
namely decoder-based and post-processing-based strategies. The introduction of NC
support within the decoder is generally performed by means of non-linear continuous
decoders. The use of switching models has particularly been reported. Switching
models rely on the estimation of a latent discrete variable to introduce state-specific
non-linearities into a generic continuous decoder. Discrete decoders are possibly
exploited to infer the value of this latent discrete variable. By contrast, a single
continuous model is trained and/or applied on both NC/IC (Intentional Control)
states when NC support is introduced at the post-processing stage. The estimation
of the value of a latent discrete variable is again used to select the post-processing
operator applied on NC and IC states.
Asynchronous sequential multi-limb decoding has rarely been considered in BCI
studies. The use of switching models has nevertheless been reported for accurate
multi-finger control. Parallel limb activations are avoided because one limb only is
detected as active at each instant.
Finally, attempts to improve decoding accuracy during IC states often focus

on the utilization of non-linear decoders, in particular generic non-linear models or
piece-wise linear models. Post-processing has also been used to improve different
characteristics of the kinematic estimates, e.g. their smoothness or spatial accuracy.
In the next chapter, the solution which has been chosen in the present doctoral
work to address the issue of accurate asynchronous mono- and multi-limb control is
presented. This solution relies on a decoder-based approach, and has been developed
within the framework of switching models.
Chapter 3
Switching models for accurate

asynchronous trajectory
decoding
Contents
3.1 Switching models for BCI control . . . . . . . . . . . . . . . . 69
3.1.1 Context-related modifications of kinematic encoding in neural
features . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
3.1.2 Switching models . . . . . . . . . . . . . . . . . . . . . . . . . 70
3.1.3 Mixture of Experts structure for discriminative switching models 72
3.2 Markov Switching Linear Models (MSLM) . . . . . . . . . . 75
3.2.1 Dynamic gating . . . . . . . . . . . . . . . . . . . . . . . . . . 75
3.2.2 Experts . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78
3.3 MSLM training . . . . . . . . . . . . . . . . . . . . . . . . . . . 79
3.3.1 Supervised training . . . . . . . . . . . . . . . . . . . . . . . . 79
3.3.2 Unsupervised training . . . . . . . . . . . . . . . . . . . . . . . 81
3.4 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86
A switching decoder, namely a Markov Switching Linear Model (MSLM), has

been proposed and designed for the task of accurate, asynchronous multi-limb ECoG
decoding.
3.1 Switching models for BCI control

The proposed MSLM decoder has been developed to ensure efficient asynchronous
multi-limb decoding in the presence of context-related modifications of a linear
dependence between neural features and kinematic parameters.
3.1.1 Context-related modifications of kinematic encoding in neu-

ral features
The existence of several motor control internal models in humans has been disclosed
in several studies [Imamizu et al., 2007] [Rouse and Schieber, 2015]. Consistently,
context-related modifications of kinematic encoding in neural features have been
Chapter 3. Switching models for accurate asynchronous trajectory
70 decoding
observed in a few BCI studies [Williams et al., 2013] [Velliste et al., 2014] [Ifft
et al., 2013]. Ifft and colleagues reported in [Ifft et al., 2013] that modifications of
the tuning properties of individual neurons were induced when NHPs performed
unimanual or bimanual reaching movements. In [Williams et al., 2013] [Velliste
et al., 2014], they were triggered when NHPs switched between NC and IC states.
It was reported in [Velliste et al., 2014] that NC samples did not fit the emission
model of a Kalman filter trained on IC samples. In [Williams et al., 2013], it was
observed that the baseline around which NHPs BCI users modulated their neural
features was different during NC and IC periods. These differences resulted in a
poor state discrimination when NC states were inferred for the estimated velocity
of the considered effector. Improvement of kinematic decoding accuracy during IC
periods has been reported when mixtures of models were applied to MUA/SUA
signals [Kim et al., 2003] [Wu et al., 2004].
These reports suggest that the framework of switching models may be relevant
to perform accurate asynchronous, multi-limb ECoG decoding for motor BCIs.
3.1.2 Switching models

The use of switching models has been reported in numerous fields. Switching Auto-
Regressive models have notably been applied for economic forecasting [Quandt,
1958] [Goldfeld and Quandt, 1973] [Quandt and Ramsey, 1978] [Kim, 1994], for
cognitive state detection from neural signals [Liehr et al., 1999] and for efficient
control of dynamical systems subject to regime changes [Alessandri et al., 2008]. In
particular, switching models have been utilized in BCI studies for accurate [Kim
et al., 2003] [Wu et al., 2003a] [Wu et al., 2004], asynchronous [Williams et al., 2013]
[Bundy et al., 2016] [Srinivasan et al., 2007] and/or multi-finger kinematic decoding
[Flamary and Rakotomamonjy, 2012]. In these studies, several continuous models
were applied on the input variable. Their relevance at time t was dependent on the
value of an unobserved (hidden, latent) discrete variable. The resulting kinematic
estimates were combined by a gate which estimated the value of the latent variable.
Two main frameworks have been used to develop these switching models, namely
switching state-space filters and switching regression models.
Generative switching models: switching state-space filters The use of

switching state-space (or Bayes) filters has been reported in several BCI studies [Wu
et al., 2004] [Wood et al., 2005] [Yu et al., 2007] [Srinivasan et al., 2007]. Bayes
filters include a model P (yt+1 |yt , Θ) of the dynamic of the state variable yt , and are
inasmuch dynamic filters. Moreover, the emission probability P (xt |yt , Θ) is utilized
to characterize the dependence between observed features and latent trajectories.
Such state-space models are thus generative [Wu et al., 2002] [Gao et al., 2003] [Kim
et al., 2006c]. The switching state-space filters applied in BCI studies extended well
known state-space (Bayes) filters, e.g. Kalman or point-process filters (see section
2.4.3.2), by conditioning their emission and/or transition models on a discrete
latent switching variable. As state-space models rely on the hypothesis that the
3.1. Switching models for BCI control 71
generation of neural features is conditioned on the continuously-valued kinematic

variable, the probability associated with each possible switching state depends on
the extent to which the corresponding couple emission/transition models can explain
neural features observed at time t (Figure 3.1), e.g. in [Wu et al., 2004] [Yu et al.,
2007] [Srinivasan et al., 2007]).
Figure 3.1: Data flow during the application of a switching state space filter, here
a Switching Kalman Filter (e.g., [Wu et al., 2003b] [Wu et al., 2004] [Srinivasan
et al., 2007]). The probability associated with each possible switching state depends
on the extent to which the corresponding couple emission/transition models can
explain neural features observed at time t.
Discriminative switching models: switching regression models Switching

modelling for kinematic motor BCIs has also been investigated in the framework of
regression models [Kim et al., 2003] [Kemere et al., 2004b] [Kemere et al., 2004a]
[Flamary and Rakotomamonjy, 2012] [Williams et al., 2013] [Bundy et al., 2016].
Regression models consider the probability P (yt |xt , Θ), i.e. they are discriminative
static models. In contrast with switching state-space (Bayes) filters, the distribution
of the neural features is assumed to be conditioned on the latent switching state
variable. The estimation of the value taken by the hidden state variable directly
relies on the current neural features (see Figure 3.2, e.g. [Bundy et al., 2016]
[Flamary and Rakotomamonjy, 2012]).
The MSLM has been developed in the framework of switching regression models.
This framework permits to introduce the key hypothesis that switching (cognitive,
internal) states are associated with different distributions of the neural features.
Such modifications are regularly observed, and are exploited in the majority of BCI
classification studies. Moreover, regression models are suitable for the decoding
72 decoding
Figure 3.2: Data flow during the application of switching regression models (e.g.,
[Kim et al., 2003] [Bundy et al., 2016]). The gate directly exploits the distribution
of the current neural features to estimate the value of the latent switching state.
of potentially high-dimensional neural features. By contrast, the applicability

of generative models like the Kalman filter may be limited in high-dimensional
settings because of deterrent computational costs (see Appendix C). A preliminary
dimensionality reduction is regularly performed when state-space models are used to
model the generation of high-dimensional data [Wu et al., 2003b] [Kao et al., 2013].
The Mixture of Experts (ME) framework [Jacobs et al., 1991] [Waterhouse, 1997]
can be used to unify the switching regression models proposed in BCI studies.
3.1.3 Mixture of Experts structure for discriminative switching

models
First introduced by [Jacobs et al., 1991], MEs combine several functions ("experts")
to model the dependence between input and output variables space [Jacobs et al.,
1991] [Waterhouse, 1997]. Experts are weighted according to the input variable
[Bishop, 2006]. Optimal weights are computed for each sample according to a given
criterion.
Let xt ∈ X ⊂ Rm be the explanatory (input, independent) variable, and
yt ∈ Y ⊂ Rn be the continuous response (output, dependent) variable. Samples
are indexed by t ∈ N. MEs assume that the input space X is partitioned into
K (possibly overlapped) regions: X = K k=1 Xk , and that a different sub-process
t
generates the output vector from the input vector in each region [Waterhouse, 1997].
Thus, Mixtures of Experts model the dependence between xt and yt by means of a
set of K local functions {fk }K
1 , fk : X → Y , referred to as "experts". Expert k is
3.1. Switching models for BCI control 73
relevant for the region Xk :
K
yt = δz t ,k fk (xt ) + ǫt
Ø
(3.1)
k=1
where δz t ,k = 1 if z t = k and δz t ,k = 0 otherwise. Here, z t = k if yt has been

generated by expert k, and ǫt is the observation noise.
The Bayes estimate ŷt = E(yt |xt ) [Bishop, 2006] of target variable yt is computed
via the decomposition of conditional expectation [Waterhouse, 1997]:
K
E(yt |xt ) = E(yt , z t = k|xt )
Ø
(3.2)
k=1
K
P (z t = k|xt ) E(yt |xt , z t = k)
Ø
= (3.3)
k=1
K
gkt ŷkt
Ø
= (3.4)
k=1
Here, ŷkt = E(yt |xt , z t = k) is the estimation issued by expert k.
The mixing coefficients gkt = P (z t = k|xt ) combine ("gate") the experts’ estimates
ŷkt .
In contrast with generic mixtures of regression models which rely on fixed prior
probabilities to combine the regression models’ outputs [Bishop, 2006], mixing
coefficients are here input-dependent. They can be interpreted as the conditional
probability of expert fk having generated the output value yt , given the value of xt
[Waterhouse, 1997]. Mixing coefficients satisfy the constraints gkt ∈ [0, 1], gkt = 1
q
k
[Bishop, 2006]. The structure which computes the mixing coefficients is often referred
to as "gating network". Linear or nonlinear experts and gating networks can be
integrated into the ME flexible structure [Yuksel et al., 2012].
ME-based approaches have been used in several BCI studies which didn’t support
NC periods or multilimb decoding, e.g. for multi-model wrist movement decoding
from monkeys’ SUA/MUA signals [Kim et al., 2003]. The switching regression
models previously implemented in asynchronous and/or multi-finger BCI studies,
e.g. [Williams et al., 2013] [Flamary and Rakotomamonjy, 2012] [Bundy et al., 2016],
can also be formulated as static mixtures of regression models. In the majority of
these ECoG-based BCI studies, the estimate yielded by a continuous decoder was
overwritten by neutral values when a classifier dedicated to state detection indicated
that the current state was most likely a NC state [Williams et al., 2013] [Flamary
and Rakotomamonjy, 2012] [Bundy et al., 2016]. More precisely, one or several
Wiener filters were weighted by the outputs of a logistic regression [Bundy et al.,
2016] or of a linear regression of state labels on neural signals [Flamary and Rako-
tomamonjy, 2012] [Williams et al., 2013]. These Wiener filters can be interpreted
as linear experts, and the classifiers used to combine them can be interpreted as
74 decoding
gates making use of soft [Chen et al., 2014b] or hard (Winner-takes-All) [Flamary
and Rakotomamonjy, 2012] [Williams et al., 2013] [Bundy et al., 2016] probabilistic
strategies to mix the experts’ outputs, where the application of hard combination
rules consists of associating a weight of 1 to the most likely current state, and a null
weight to all other possible states.
Although recurrent gates were embedded into the previously mentioned switching
state-space filters [Wu et al., 2004] [Srinivasan et al., 2007], the switching regression
models which application has been reported for BCI kinematic decoding relied on
static gates [Flamary and Rakotomamonjy, 2012] [Bundy et al., 2016]. Such static
gates are based on the assumption that pairs (xt , z t ) are temporally independent
and identically drawn from the distribution P (xt , z t ) [Dietterich, 2009]. As the
classes associated with internal states are likely to be overlapped in BCI studies, this
discrimination process is liable to issue erroneous estimates of the internal states.
More realistic hypotheses about temporal dependencies in the state sequences have
been exploited in BCI studies (see section 2.4.2.3), and in particular in a few
classification-based motor BCIs [Fifer et al., 2014] [King et al., 2015] [Hotson et al.,
2016]. When compared to static gates, dynamic (recurrent) gates are expected to
be less liable to output spurius detections (e.g., [Saa et al., 2016]). The MSLM
was thus designed as a switching regression model with recurrent gating for robust
estimation of the user’s latent switching state (Figure 3.3).
Figure 3.3: Data flow during the application of the MSLM.

3.2. Markov Switching Linear Models (MSLM) 75
3.2 Markov Switching Linear Models (MSLM)

The MLSM is built on the assumption that the continuous model which relates
neural features xt and kinematic parameters yt is a linear model conditioned on
a discrete latent variable z t ∈ N (switching state). A probabilistic rule is used to
combine the linear models respectively associated with each value z t . The discrete
latent variable z t is assumed to be generated by a first-order Markov chain. The
Mixture of Experts (ME) framework [Jacobs et al., 1991] [Waterhouse, 1997], which
was used in section 3.1.3 to formalize static mixtures of regression models, was
here extended to support dynamic mixtures of regression models.
3.2.1 Dynamic gating

The MSLM embeds a dynamic gate, which is liable to reduce the number of erroneous
state estimates.
A parallel can be drawn between efficient state classifiers and efficient ME gates
if gate optimization is momentarily considered independently of expert application.
The potential sub-optimality of static classifiers for cognitive state detection has
been taken into account by the BCI community, inasmuch as the use of several
strategies has been reported for dynamic state detection in several BCI studies.
A first approach consists in integrating the a priori knowledge that fast switches
between internal states are unlikely at the post-processing stage. Several strategies
were considered and possibly combined, namely filtering of a NC/IC classifier output
[Mason and Birch, 2000] [Millán and Mouriño, 2003] [Bashashati et al., 2007a] [King
et al., 2015], triggering a state transition after successive identical state estimates
only [Townsend et al., 2004] [Pfurtscheller et al., 2010], and/or by blocking state
transitions for a predefined duration after a performed transition [Townsend et al.,
2004] [Pfurtscheller et al., 2010]. A second approach consists in using dynamic
classifiers (see section 2.4.2.3), which explicitly integrate a model of the state
succession dynamic. In particular, dynamic classifiers such as Hidden Markov Models
(HMM) [Kemere et al., 2008] [Hotson et al., 2016] and Conditional Random Fields
(CRF) [Hasan and Gan, 2011] [Saa et al., 2016] were used for the estimation of
NC/IC states, possibly with several NC- or IC-related sub-states, in SUA/MUA
[Kemere et al., 2008], EEG [Hasan and Gan, 2011] and ECoG signals [Saa et al.,
2016].
Dynamic gating has additionally been considered in the ME framework. Several
extensions have been specifically proposed to adapt MEs for sequential data modelling
[Yuksel et al., 2012]. Recurrent gating networks [Cacciatore and Nowlan, 1994]
[Meila and Jordan, 1996] were integrated into ME and for example applied for
movement modelling [Meila and Jordan, 1996] and efficient control of switching
systems [Cacciatore and Nowlan, 1994]. Particularly, Markov Mixtures of Experts
(MME) [Meila and Jordan, 1996] [Bengio and Frasconi, 1996] (referred to as Input-
Output HMM in [Bengio and Frasconi, 1996]) model a temporal sequence of input-
output data by dynamic switching between local static models. The expert sequence
76 decoding
(z t )t∈N is assumed to be generated by a first-order Markov chain. The inputs

and outputs corrupted by noise are measured, but states are hidden and must be
estimated from measurements.
Following [Fifer et al., 2014] [Hotson et al., 2016] [Srinivasan et al., 2007], it was
here decided to explicitly model the temporal dependencies between internal states
using a first-order Markov chain. The MSLM proposed herein is a variant of MME
which relies on Hidden Markov Models (HMMs) for dynamic gating of the experts’
estimates. It relies on the underlying hypothesis that neural signals are generated by
internal (cognitive) states. Thus, the distribution of the extracted neural features xt
is conditioned on unobserved discrete states z t , P (xt |z t ). By contrast, the previously
reported MMEs and IOHMMs [Bengio and Frasconi, 1996] [Meila and Jordan, 1996]
assumed that the input variable xt conditions the probability of switching from one
state to another, P (z t+1 |z t , xt ).
3.2.1.1 HMM-based dynamic gating

Hidden Markov Models (HMMs) are a powerful tool for the modelling of stochastic
time-varying processes [Rabiner and Juang, 1986]. Two discrete-time stochastic
processes are involved in discrete-time HMMs. The first process is an unobservable
("hidden") dynamic sequence (z t )t∈N . By contrast, the second sequence (xt )t∈N
is observable, and is referred to as the "observation sequence". The MSLM state
decoder models the state (z t )t∈N and feature sequences (xt )t∈N by a HMM, where
z t ∈ N is the hidden variable and xt ∈ Rm is the observed variable.
Figure 3.4: Hidden Markov Model. The sequence (z t )t∈Z is hidden. The sequence
(xt )t∈Z is observed and permits to infer an estimate (ẑ t )t∈Z of the hidden sequence.
HMM’s hidden variable z t can take K distinct discrete values ("states"), z t ∈ Z

with Z = {z1 , z2 , ..., zK } ⊂ Z. For the sake of simplicity, {z1 , z2 , ..., zK } is here
taken to be {1, 2, ..., K}. In the typical case of a first-order HMM, the value of z t+1
depends exclusively on the current value z t : P (z t+1 = k|z 1:t ) = P (z t+1 = k|z t ),
t ∈ N. The hidden state sequence is then fully characterized by the transition
matrix A = (aij ), aij = P (z t+1 = j|z t = i), and by the probability π associated
with the value of the first hidden state: πi = P (z 1 = zi ). The probability Pi (d) to
remain in state i for exactly d consecutive time steps is therefore explicitly modelled
as Pi (d) = P (z 1 = zi , z 2 = zi , . . . , z d = zi , z d+1 = zj ) = (aii )d−1 (1 − aij ), where
3.2. Markov Switching Linear Models (MSLM) 77
i=Ó j. The probability distribution of the observed variable xt ∈ Rm is exclusively

conditioned on the current hidden state z t : P (xt |z 1:t ) = P (xt |z t ), t ∈ N, where the
conditional probability P (xt |z t ) is referred to as "state emission density". Various
parametric (Gaussian, Student) and non-parametric state emission distributions
have been integrated in HMMs [Rabiner, 1989]. Although conventional HMMs thus
directly model the emission distributions P (xt |z t = j), j, . . . , K, variants based on
discriminative modelling have been proposed by several teams (e.g., using Neural
Networks [Renals et al., 1994] [Bourlard and Morgan, 1998] [Ordóñez et al., 2013] or
Support Vector Machines [Valstar and Pantic, 2007] [Ordóñez et al., 2013]). These
variants consist in applying a discriminative model on the data, thereby obtaining
the conditional probability P (z t = j|xt ). Class priors P (z t = j) and Bayes rule are
t t)
then combined to compute HMM emission probabilities P (xt |z t = j) ∝ PP(z(z=j|x t =j)
[Bourlard and Morgan, 1998] [Valstar and Pantic, 2007] [Ordóñez et al., 2013]. Thus,
the MSLM can embed both discriminative and generative state detection into its
HMM-based gate.
Let {b1 , . . . , bK } gather the parameters necessary to characterize the distributions
P (xtî|z t = j), j = 1, . .ï. , K. Then the HMM is fully described by the parameter set
λ = A, {bj }K j=1 , π . HMM training is performed using supervised or unsupervised
approaches, and typically relies on Maximum-Likelihood estimation [Ghahramani,
2001].
3.2.1.2 MSLM online gating

As online estimation of users’ intentions is required for the utilization of BCI systems,
the MSLM’s input-dependent mixing coefficients are exclusively conditioned on past
and current neural observations: gkt = P (z t = k|x1:t ). Mixing coefficients are first
decomposed using Bayes rule:
P (z t = k, x1:t )
P (z t = k|x1:t ) = . (3.5)
P (x1:t )
The Viterbi algorithm or the forward-backward algorithm are often used to
infer the hidden state sequence z = (z t )Tt=1 associated with an observation sequence
X = (xt )Tt=1 . Both Viterbi and forward-backward algorithms cannot be applied
for online processing because they rely on a back-tracking procedure from the last
estimated state ẑ T = argmax P (z t = k | x1:T ) [Rabiner, 1989]. By contrast, the
k
forward algorithm [Rabiner, 1989] permits to compute estimates P (z t = k|x1:t )
which are only conditioned on past and current observations. The denominator
P (x1:t ) and numerator P (z t = k, x1:t ) of (3.5) are thus computed using the HMM’s
forward algorithm for online application of the MSLM.
3.2.1.3 Soft and hard gating

Both discrete-valued [Flamary and Rakotomamonjy, 2012] [Suway et al., 2013] [Bundy
et al., 2016] and continuous-valued mixing coefficients [Chen et al., 2014b] were used
78 decoding
in the switching regression models previously implemented for neural signal decoding.
When the mixing coefficients are discrete-valued, i.e. gkt ∈ {0, 1} ∀t, k = 1, . . . , K,
the model which is associated with the most likely class at time t is applied on the
neural features xt ("hard" combination). This Winner-Takes-All strategy (or "hard"
switching) has been used in the majority of switching kinematic decoders [Flamary
and Rakotomamonjy, 2012] [Suway et al., 2013] [Bundy et al., 2016]. By contrast,
the utilization of continuously-valued mixing coefficients, i.e. gkt ∈ [0, 1], has rarely
been reported in BCI studies. It was for example used in [Chen et al., 2014b] to
weight IC models for asynchronous finger trajectory reconstruction. In contrast
with hard model combination which results in abrupt transitions between decoding
regimes, this probabilistic, "soft" combination results in smooth transitions between
experts.
Although abrupt transitions may not be absolutely deterrent because users may,
to a certain extent, learn to correct high-frequency errors [Stavisky et al., 2015]
[Marathe and Taylor, 2015], we expect them to be disturbing to BCI users and
therefore undesirable. Without loss of generality, the MSLM thus makes use of a
soft, probabilistic strategy to combine models.
3.2.2 Experts
Any expert fk : Rm → Rn can be integrated into the ME switching structure. The

extraction of kinematic parameters from ECoG features during IC periods has mainly
been performed by means of linear models, in particular in the case of wrist [Schalk
et al., 2007] [Hammer et al., 2016] [Bundy et al., 2016] and finger [Kubánek et al.,
2009] [Liang and Bougrain, 2009] [Flamary and Rakotomamonjy, 2012] trajectory
estimation. Linear experts were therefore integrated into the proposed switching
model. If a Winner-takes-all strategy is used to combine its experts’ estimates, the
MSLM can be seen as a piecewise linear model:
K
yt = δz t ,k Bk xt + ǫt
Ø
k = 1, . . . , K, (3.6)
k=1
where Bk ∈ Rn×m and ǫt ∈ Rn is the observation noise.

The use of other types of nonlinear models, e.g. Generalized Linear Models
[Eliseyev and Aksenova, 2014], has been explored for (possibly asynchronous) ECoG
decoding. When compared to generic non-linear models, the MSLM makes use
of the prior knowledge that discrete internal states are the sources of the model
non-linearity. It benefits from a simple application procedure, a straightforward
interpretation, and can be easily extended by adding a new state (e.g., progressive
introduction of new limbs during training). Moreover, as GLMs apply a non-linear
link function on an intermediary trajectory estimate, i.e. state and trajectory
decoding are not independent, they can be seen as a sophisticated post-processing
approach to the integration of NC support into regression models.
3.3. MSLM training 79
3.3 MSLM training

î ï
Let Θ = {Θg , Θe } gather the HMM-based gating network Θg = A, {bj }K
j=1 , π
and the experts’ Θe = {Bk }K k=1 parameters, where, depending on context, Bk refers
either to the full set of parameters associated with expert i (i.e., parameters of
the linear model and of the noise distribution) or only to the parameters of the
linear model. Prior to any MSLM application, estimation of both expert and gating
networks’ parameters is required. When the expert responsibility is known, training
is supervised with respect to the expert sequence. Training is otherwise unsupervised
with respect to the expert sequence.
The Maximum-Likelihood (ML) criterion is routinely used for both EM [Yuksel
et al., 2012] and HMM training [Ghahramani, 2001]. ML training procedures were
derived for supervised and unsupervised training of the MSLM.
3.3.1 Supervised training

Fully supervised training of the MSLM relies on a training data set {X, Y, z}, where
X ∈ RT ×m , Y ∈ RT ×n , and z ∈ NT ×K gather the observed sequences (xt )Tt=1 ,
(yt )Tt=1 and (z t )Tt=1 respectively.
3.3.1.1 Independent training for gate and experts

Let us consider a MSLM composed of K linear experts gated by a HMM-based
sequential decoder. Each linear expert i is parametrized by Bi . Let A = (aij ) be
the transition matrix aij = P (z t+1 = j|z t = i), i, j = 1, ..., K and π be the initial
state distribution πi = P (z 1 = zi ), i = 1, ..., K associated with the first-order
Markovian hidden state variable z t . Finally, let {b1 , . . . , bK } gather the parameters
necessary to characterize the distributions P (xt |z t = j), j = 1, . . . , K, e.g. mean and
variance in the case of Gaussian emissions. Let {X, Y, z} be a complete training
data îset available for Maximum-Likelihood
ï estimation of the MSLM parameters
Θ = {Bi }K i=1 , π, A, {b }K
i i=1 .
T 1 2
Assuming the temporal independence P (x1:T , y1:T |z 1:T , Θ) = P xt , yt |z 1:T , Θ ,
r
t=1
the data complete log-likelihood Lc (Θ, X, Y, z) = ln P (X, Y, z|Θ) can be expressed
as (see Appendix A)
K −1
K TØ
K Ø
t
Ø Ø
Lc (Θ, X, Y, z) = δz 1 ,i ln(πi ) + ωi,j ln(aij )
i=1 i=1 j=1 t=1
T
K Ø T
K Ø
δz t ,i ln(P (yt |xt , Bi )) + δz t ,i ln P (xt |bi ), (3.7)
Ø Ø
+
i=1 t=1 i=1 t=1
where δz t ,k = 1 if z t = k and δz t ,k = 0 otherwise, and ωi,j

t = 1 if z t = i and z t+1 = j,
and ωi,j = 0 otherwise.

t
80 decoding
Maximizing (3.7) amounts to solving the following maximization problems (see

Appendix A):
T
δz t ,i ln P (yt |xt , Bi ) (3.8)
Ø
maximize i = 1, . . . , K
Bi
t=1
K
Ø
maximize δz 1 ,j ln πj
π
j=1
(3.9)
K
Ø
subject to πj = 1
j=1
−1
K TØ
K Ø
t
Ø
maximize ωij ln aij
A
i=1 j=1 t=1
(3.10)
K
Ø
subject to aij = 1 ∀i ∈ {1, . . . , K} .
j=1
T
δz t ,i ln P (xt |bi ) (3.11)
Ø
maximize i = 1, . . . , K.
bi
t=1
The maximization problems (3.8), (3.9), (3.10), and (3.11) correspond to ML esti-
mation of the expert, initial state, transition and emission parameters, respectively.
3.3.1.2 Expert parameter identification
ML estimates of the experts’ parameters B̂k , k = 1, . . . , K are found by solving the

problems (3.8). As expressed in (3.8), the training of expert k is based on sample
observations {Xk , Yk }, where Xk and Yk gather training samples observed at times
t such that z t = k. The resolution of (3.8) depends on the expert noise distribution.
Let for example assume that the observation noise ǫt is Gaussian, i.e. P (ǫt ) =
P (yt −xt Bk ) ∼ N (0, σI), k = 1, . . . , K. It can then be shown that the maximization
problem (3.8) amounts to the minimization of the squared error ëYk − Xk Bk ë2 . The
corresponding system is generally undetermined in BCI studies, which input variables
are regularly highly correlated and/or high-dimensional, i.e. m >> T . The minimum
of ëYk − Xk Bk ë2 is reached at several points, i.e. the solution is not unique. One
solution consists in computing the Minimal Length Least Squares (MLLS) estimator
B̂M LLS = (XTk Xk )† XTk Yk , where X† refers to Moore–Penrose pseudoinverse. The
MLLS estimator may be unstable when high-dimensional features are considered.
Several approximate solutions can be used to deal with high dimensional variables,
e.g. penalized training (ridge, LASSO), Principal Component Regression or Partial
Least Squares. The estimators which use has been considered for the MSLMs
implemented in the present doctoral work are presented in Appendix B. The
solutions which were finally utilized for efficient training of the implemented MSLMs
are exposed in Chapter 6.
3.3.1.3 Gate parameter identification

Training of the HMM-based gating network is performed on the data set {X, z}.
(i) Transition matrix The estimate Â of the transition matrix is found by

counting transition frequencies in the training sequence (z t )Tt=1 [Dietterich, 2009]:
qT t
t=1 ωij
âij = qK qT t
. (3.12)
k=1 t=1 ωik
(ii) Initial probabilities The initial probability associated with the constrained
optimization problem (3.9) is π̂i = δz 1 ,i ∀i ∈ {1, . . . , K}.
(iii) Emission probabilities The MSLM gate either relies on generic emission
models, or it exploits discriminative classifiers and Bayes rule for indirect emission
modelling (see section 3.2.1). The solution to (3.11) depends on the distribution
chosen to model P (xt |z t = i). In the case of Gaussian distributions for example,
analytical solutions to (3.11) are obtained by applying well-known ML estimators on
1 qT
each data set Xk , e.g. µk = card(t s.t.z t =i) t=1,s.t.z t =i x for the mean of a Gaussian
t
distribution. Details on the resolution of (3.11) for the distributions which were here
considered for the MSLM implementation are given in Chapter 6.
3.3.2 Unsupervised training

Supervised training is performed when the expert sequence (z t )Tt=1 is not hidden
in the training data set. This may not always be the case. For example, several
experts may be devoted to trajectory encoding during IC states. In that case, no
prior information is available about the responsibility of each expert. In this more
general case where the expert sequence (z t )Tt=1 is hidden in the training data set,
ME training is an unsupervised learning problem [Weigend et al., 1995].
The estimator Θ̂M L (X, Y) = argmax L(Θ, X, Y), where L(.) refers to the
Θ
incomplete data log-likelihood, is considered. The use of gradient ascend methods
has been reported for direct likelihood maximization of both HMM [Cappé et al.,
2006] and ME [Jacobs et al., 1991] [Moerland, 1997] parameters. These methods
don’t utilize the fact that knowledge on the HMM or ME switching latent variable
permits to decouple the parameters’ contribution on data likelihood, and therefore
involve complex derivation procedures [Moerland, 1997]. On the other hand, data
augmentation-based methods, e.g. the well-known Expectation-Maximization (EM)
algorithm [Dempster et al., 1977], are particularly suited for parameter estimation
in the case of missing data.
The EM algorithm facilitates ML or Maximum A Posteriori (MAP) parame-
ter estimation [Dempster et al., 1977] when the training observations are condi-
tioned on an unobserved (latent, missing) variable, here the switching state z t .
It exploits the fact that the maximization of the complete data log-likelihood
82 decoding
Lc (Θ, X, Y, z) = ln P (X, Y, z|Θ) is simpler than the maximization of L(Θ, X, Y)

because the impact of each parameter is decoupled. The unknown latent variable
vector z, z = (z 1 , . . . , z T )′ is artificially re-introduced in the maximization problem,
and permits to consider the complete data log-likelihood Lc (Θ, X, Y, z). An alter-
nating estimation procedure is used to infer both the parameters Θ and the vector
z from the training data. The two alternating steps are referred to as the E- and
M-step, for Expectation- and Maximization-step, respectively.
Application of the EM algorithm is regularly reported for both HMM [Rabiner,
1989] [Cappé et al., 2006] and ME [Yuksel et al., 2012] unsupervised training. An
EM-based training was thus derived for the MSLM unsupervised training. An
insight on the application of the EM algorithm for MSLM training is given in
section 3.3.2.1. MSLM-specific E- and M-step are exposed in sections 3.3.2.2
and 3.3.2.3, respectively.
3.3.2.1 The Expectation-Maximization algorithm
Let Θ gather the unknown MSLM parameters. Let {X, Y} be the incomplete
training data set available for Maximum-Likelihood identification of the parameters
Θ. After initialization of the MSLM parameters, the Expectation-step (E-step) and
the Maximization-step (M-step) are performed iteratively. Successive iterations
of the E- and M-step lead to the convergence of the data set log-likelihood to a
local optimum [Dempster et al., 1977]. Let us consider the ith iteration of the EM
algorithm.
E-step The unknown latent variable vector z depends on the parameter Θ. During
the ith E-step, the value of Θ is fixed and is assumed to be equal to its current
estimate Θ̂i . The expectation of the data log likelihood with respect to the unknown
latent
î variable vector ïz is computed for this estimate Θ̂i . Let first define the function
Θ Θ̂
Q : D , D , R , R → R as follows:
m n
è é
Q(Θ+ , Θ̂, X, Y) = E Lc (Θ+ , X, Y, z)|Θ = Θ̂, X, Y . (3.13)
The E-step considers the function Q evaluated at Θ̂i and X, Y:
è é
Θ+ Ô→ Q(Θ+ , Θ̂i , X, Y) = E Lc (Θ+ , X, Y, z)|Θ = Θ̂i , X, Y , (3.14)
i.e. the expectation of Lc (Θ, X, Y, z) with respect to z, given the current parameter
estimate Θ̂i :
è é
E Lc (Θ+ , X, Y, z)|Θ = Θ̂i , X, Y = Lc (Θ+ , X, Y, z) P (z|Θ = Θ̂i , X, Y)
Ø
z
(3.15)
M-step The M-step consists of maximizing Q(Θ+ , Θ̂i , X, Y) with respect to Θ+ ,

i.e. of optimizing the parameter value when the latent state is replaced by its
expectation. The following updated ML estimate Θ̂i+1 is obtained:
Θ̂i+1 = argmax Q(Θ+ , Θ̂i , X, Y) (3.16)

Θ+
Depending on the function Q, analytical solutions or iterative approaches (gradient
methods) are used to compute Q(Θ+ , Θ̂i , X, Y) with respect to Θ+ . The utilization
of iterative maximization approaches results in the presence of an "inner-loop" within
each M-step, and significantly increases the computational cost of the EM algorithm.
The E- and M-steps were here derived for MSLM unsupervised training.
3.3.2.2 E-step for MSLMs

Execution
è of the MSLM-specific E-step
é consists of expressing the function Q(Θ+ , Θ̂, X, Y) =
E Lc (Θ+ , X, Y, Z)|Θ = Θ̂, X, Y (see equation (3.13)). Taking the expectation of
the complete data log-likelihood presented in equation (3.7), and thus exploiting the
hypotheses presented for the MSLM supervised training, we get:
è é K K TØ
K Ø −1
E Lc (Θ+ , X, Y, z)|Θ = Θ̂, X, Y = γi1 ln(πi+ ) + t
ln(a+
Ø Ø
ξi,j ij )
i=1 i=1 j=1 t=1
T
K Ø T
K Ø
γit ln P (yt |xt , B+ γit ln P (xt |b+
Ø Ø
+ i )+ i ), (3.17)
i=1 t=1 i=1 t=1
where
è é
γi1 = E δz 1 ,i |Θ = Θ̂, X, Y = P (z 1 = i|Θ = Θ̂, x1:T , y1:T ), (3.18)
è é
γit = E δz t ,i |Θ = Θ̂, X, Y = P (z t = i|Θ = Θ̂, x1:T , y1:T ), (3.19)
è é
t
ξi,j t
= E ωi,j |Θ = Θ̂, X, Y = P (z t = i, z t+1 = j|Θ = Θ̂, x1:T , y1:T ). (3.20)
Estimation of the state and transition probabilities The state γit and tran-
sition probabilities ξi,j
t can be expressed as follows (see derivation of the formula in
Appendix A):
γit = P (z t = i|x1:T , y1:T , Θ)

P (yt+1:T , xt+1:T |z t = i, Θ)P (z t = i, x1:t , y1:t |Θ)
= ,
P (x1:T , y1:T |Θ)
84 decoding
and
t
ξi,j = P (z t = i, z t+1 = j|x1:T , y1:T , Θ)
P (xt+2:T , yt+2:T |z t+1 = j, Θ)P (xt+1 , yt+1 |z t+1 = j, Θ)
=
P (x1:T , y1:T |Θ)
× P (z t+1 = j|z t = i, Θ)P (z t = i, x1:t , y1:t |Θ),
where
P (z t+1 = j|z t = i, Θ) = aij ,

P (xt , yt |z t = j, Θ) = P (yt |xt , Bj )P (xt |bj ).
Finally, the intermediary probabilities P (z t = k, x1:t , y1:t |Θ), P (xt+1:T , yt+1:T |z t =

j, Θ) and P (x1:T , y1:T |Θ) are computed using extensions of the forward and backward
algorithms originally developed for HMMs [Rabiner, 1989]. The HMM forward and
backward algorithms permit to compute the probabilities P (z t = i, x1:t |Θ) and
P (xt+1:T |z t = i, Θ) (see their derivation in [Bengio and Frasconi, 1996]) . MSLM-
specific forward and backward algorithms are derived in the present doctoral work.
They extend the HMM-specific ones by taking into account the likelihood of each
continuous expert when estimating the probability associated with each possible
hidden state.
MSLM forward algorithm The MSLM forward algorithm permits to compute

the probabilities P (z t = k, x1:t , y1:t |Θ), t = 1, . . . , T . The recurrence is initialized by
P (z 1 = j, x1 , y1 |Θ) = P (y1 |x1 , z 1 = j, Θ)P (x1 |z 1 = j, Θ)P (z 1 = j, Θ). The follow-
ing recurrence formula is then applied (see derivation of the formula in Appendix
A):
P (z t+1 = j, x1:t+1 , y1:t+1 |Θ) = P (z t+1 = j, x1:t , y1:t , xt+1 , yt+1 |Θ)
= P (xt+1 , yt+1 |z t+1 = j, Θ)
P (z t+1 = j|z t = i, Θ)P (z t = i, x1:t , y1:t |Θ),
Ø
×
i
where P (xt , yt |z t = j, Θ) = P (yt |xt , Bj )P (xt |bj ). Finally, a termination step yields
N
P (x1:T , y1:T |Θ) = P (z T = i, x1:T , y1:T |Θ).
q
i=1
MSLM backward algorithm Similarly, a modified backward algorithm permits

to compute P (xt+1:T , yt+1:T |z t = k, Θ). Following the generic HMM-specific back-
ward algorithm [Fink, 2014], the recurrence is initialized with P (xT , yT |z T = j, Θ) =
(1, . . . , 1)′ . The following iterations are then completed (see derivation of the formula
in Appendix A):
P (xt+1:T , yt+1:T |z t = i, Θ) = P (xt+1 , yt+1 |z t+1 = jΘ)

Ø
× P (xt+2:T , yt+2:T |z t+1 = j, Θ)P (z t+1 = j|z t = i, Θ).
Again, the likelihood of each continuous expert is taken into account during state
estimation.
3.3.2.3 M-step for MSLMs
The introduction of the latent switching variable z t permits to decouple parameters

in (A.12). Optimization problems similar to the ones associated with supervised
MSLM training are obtained, namely:
T
γit ln P (yt |xt , Bi ) (3.21)
Ø
Bi
t=1
K
γj1 ln πj
Ø
maximize
π
j=1
(3.22)
K
Ø
subject to πj = 1
j=1
K Ø −1
K TØ
t
Ø
maximize ξi,j ln aij
A
i=1 j=1 t=1
(3.23)
K
Ø
subject to aij = 1 ∀i ∈ {1, . . . , K} .
j=1
T
γit ln P (xt |bi ) (3.24)
Ø
maximize i = 1, . . . , K.
bi
t=1
where (3.21), (3.22), (3.23), and (3.24) correspond to ML estimation of the expert,
initial state, transition and emission parameters, respectively. In contrast with
the MSLM supervised maximization problems, the impact of observation t on
a parameter set associated with state i is weighted by the corresponding state
probability γjt = P (z t = i|x1:T , y1:T , Θ).
Experts Identification of the experts’ parameters consists in maximizing (3.21).

Specific training procedures are required because training samples are weighted
by state probabilities rather than pooled by state labels. Again, they depend on
the noise distribution. The approximate solution used to solve this maximization
problem for the MSLMs implemented in the present doctoral work (case of Gaussian
noise with high-dimensional features) is presented in Chapter 6.
86 decoding
Gate Similarly, weighted training is required to identify the transition, initial state
and emission probabilities.
(i) Transition matrix The application of Lagrange Multipliers permits to

obtain the updated estimate of the transition matrix [Bilmes, 1998]:
qT t
t=1 ξij
âij = qK qT t
i = 1, . . . , K, j = 1, . . . , K. (3.25)
k=1 t=1 ξik
Intuitively, this expression corresponds to a weighted frequency of each possible

transition.
(ii) Initial state probability The constrained optimization problem (3.22)

is again solved by applying Lagrange Multipliers, which yield [Bilmes, 1998]:
πz 1 = γi1 i = 1, . . . , K.
(iii) Emission probability The solution of the maximization problem (3.24)

depends on the distribution used to model feature emission. The solution is for
example well known in the case of Gaussian emission distributions (i.e., weighted
Bayes classifier) [Rabiner, 1989]. Implementation details for the MSLMs considered
in the present doctoral work are given in Chapter 6.
3.4 Conclusion
In this chapter, a decoder has been introduced for the task of accurate, asynchronous
mono- and multi-limb ECoG kinematic decoding. This decoder, referred to as
Markov Switching Linear Model (MSLM), presents three key features expected to
be associated with efficient asynchronous kinematic decoding. First, it has been
developed as a switching model. Thus, as opposed to post-processed models, the
MSLM combines several models. It is therefore able to adjust to task-dependent
changes in the model between ECoG signals and limb kinematics. This property is
expected to be useful to both introduce No-Control (NC) support into kinematic
decoders and to improve Intentional Control (IC) decoding accuracy during complex
movements. If one active limb only is associated with each model, the MSLM
intrinsically limits parallel limb activations. Second, it has been designed as a
discriminative rather than generative switching model. The discriminative framework
is expected to facilitate state and kinematic extraction from high-dimensional neural
features. The Mixture of Experts (ME) framework has been used to describe the
discriminative switching models reportedly applied for the integration of NC support
into synchronous ECoG decoders. Finally, in contrast with the discriminative
switching models proposed in the literature, the MSLM performs dynamic state
estimation. Dynamic state detection is liable to improve state detection, and thus to
further reduce spurious activations during NC periods. It relies on an extension of
3.4. Conclusion 87
the ME framework to the case of dynamical state detection. A supervised approach

has been proposed for the training of MSLMs. It is applicable when the values
taken by the switching variable are known in the training data set. An EM-based
training algorithm, which can be applied when the switching variable is hidden in
the training data set, has been derived for unsupervised training of the MSLM.
The implementation of the supervised and unsupervised MSLM training approaches
depends on the distributions chosen to model the neural feature distribution and
the experts’ measurement noise. Implementation details for the MSLMs considered
in the present doctoral work are presented in Chapter 6.
Chapter 4
Data set presentation
Contents
4.1 Asynchronous preclinical data set . . . . . . . . . . . . . . . 89
4.2 Asynchronous multi-limb (multi-finger) clinical data set . . 95
4.2.2 Data acquisition . . . . . . . . . . . . . . . . . . . . . . . . . 95
4.3 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99
The performance of the proposed Markov Switching Linear Model (MSLM)

was evaluated on two publicly available ECoG data sets, namely a preclinical
and a multi-limb (multi-fingers) clinical data set. These data sets permitted to
complete a first offline investigation on the ability of the MSLM to perform accurate
asynchronous multi-DoF (Degrees of Freedom) or multi-limb ECoG decoding, and
paved the way for the MSLM closed-loop evaluation which will be accomplished
during CLINATEC’s forthcoming clinical trial.
4.1 Asynchronous preclinical data set

The preclinical data set was composed of subdural and epidural ECoG data sets
acquired and distributed by the Laboratory for Adaptive Intelligence, RIKEN
Brain Science Institut, Saitama, Japan. The data sets are publicly available at
http://neurotycho.org/. Full description of the experimental set-ups can be
found in [Chao et al., 2010] and [Shimoda et al., 2012]. The cortical activity of 4
Non-Human Primates (Monkeys A, K, B and C) was recorded by chronic ECoG
arrays during a food reaching task.
4.1.1 Behavioural task

The Non-Human Primates (NHPs) were sitting in front of an experimenter during
the data acquisition sessions. The experimenter presented food to the NHPs at
random intervals. The NHPs performed reaching movements to grab the food and to
90 Chapter 4. Data set presentation
bring it to their mouth (Figure 4.1). An example of reaching trajectory is shown

in Figure 4.2.
Figure 4.1: Behavioral task (reproduced from [Chao et al., 2010]). The NHPs
performed reaching movements to grab food presented at random intervals.
Figure 4.2: Reaching movement example. After having grabbed the food presented
by the experimenter and put it in its mouth, the NHP brought its arm back to an
idle position.

Monkeys A and K were implanted with a 32- and 64- channel subdural ECoG
array, respectively (Figure 4.3). Both Monkeys B and C were epidurally implanted
with a 64-channel ECoG array (Figure 4.3). The customized arrays embedded
2.1mm-diameter electrodes with 3.5mm inter-electrode distances.
4.1. Asynchronous preclinical data set 91
Figure 4.3: ECoG array implantation for Monkeys A, K, B and C (reproduced

from [Chao et al., 2010] [Shimoda et al., 2012]). Solid grey circles indicate reference
electrodes. Ps: principal sulcus; As: arcuate sulcus; Cs: central sulcus; IPs:
intraparietal sulcus.
Table 4.1: State sequence characteristics, preclinical data set. The intra-session
mean number of movements and mean duration of NC and IC periods was averaged
over 5, 3, 6 and 10 sessions for Monkeys A, K, B and C, respectively.
Data set NHP Movement number NC period IC period

duration (s) duration (s)
subdural A 104 ± 24 5.2 ± 1.7 5.7 ± 2.0
K 69 ± 8 5.7 ± 2.0 3.6 ± 0.1
epidural B 70.3 ± 14.3 7.7 ± 3.1 7.5 ± 1.2
C 127.4 ± 20.6 4.4 ± 0.9 3.7 ± 0.5
ECoG signals were acquired at a sampling rate of 1kHz. A motion tracking system
tracked the subjects’ wrist coordinates at a sampling frequency of 120Hz. Body-
centred 3D wrist trajectories downsampled at 20Hz were extracted from the outputs
of the motion tracking system.
4.1.3 Data set characteristics

4 sessions were excluded from the epidural data set because of difficulties to assess
NC/IC labels. The final data set was thus composed of 8 subdural sessions and 16
epidural sessions. The corresponding tracked trajectories exhibit essential differences
in terms of temporal characteristics of NC/IC state succession (i), variability of the
wrist position during relaxed states (ii) and trajectory complexity (iii) (see trajectory
examples in Figure 4.4).
(i) Duration of NC and IC periods Manual labelling of NC and IC states was

used to compute the average duration of NC and IC periods and the number of
reaching movement executed during each acquisition session (Table 4.1). NC and
IC classes were found to be relatively well balanced for most sessions.
(ii) Neutral position variability The variability of monkeys’ wrist position

during NC periods was established for each axis. Results are displayed in Table
4.2. The y1 - and y2 -axes are associated with unstable NC wrist positions, i.e. NHPs
relaxed their wrist on variable points on the horizontal plan. Wrist positions along
the y3 -axis (vertical), on the other hand, are relatively stable during NC periods.
(iii) Trajectory complexity The complexity of the IC trajectories was measured

as followed. First, the trajectory derivative was computed by convolving it with the
derivative of a Gaussian kernel. The derivatives’ zero-crossings were then detected
(see Figure 4.5). Their average number was obtained by dividing the number
of detected zero-crossings by the number of IC movements. 3 sessions from the
subdural data set were found to have an average complexity superior to 2.
4.1. Asynchronous preclinical data set 93
A
y 1 (m)
0.2
0
y 2 (m)
0.2
0
y 3 (m)
0.2
0
B
y 1 (m)
0.2
0
y 2 (m)
0.2
0
y 3 (m)
0.2
0
C
y 1 (m)
0.2
0
y 2 (m)
0.2
0
y 3 (m)
0.2
0
200 220 240 260 280 300 320 340 360 380 400
time (s)
Figure 4.4: Trajectory examples. Trajectories along each axis were centred around
0. An identical scale was used for all NHPs and all axes. A. Monkey A, session 1.
B. Monkey K, session 2. C. Monkey C, session 6.
Table 4.2: NC position variability during NC states. The intra-session standard

deviation of the wrist position was averaged over 5, 3, 6 and 10 sessions for Monkey
A, K, B and C, respectively.
Data set NHP y1 -axis (mm) y2 -axis (mm) y3 -axis (mm)

subdural A 25.9 ± 12.3 29.6 ± 14.2 8.9 ± 2.8
K 15.8 ± 3.1 21.4 ± 10.9 2.8 ± 0.6
epidural B 12.2 ± 4.6 16.3 ± 7.5 28.2 ± 10.6
C 38.7 ± 4.2 36.4 ± 8.7 11.1 ± 4.5
300
y 3 (mm)
200
100
700 720 740 760 780 800 820 840 860 880
time (s)
observed trajectory detected derivative zero-crossings
Figure 4.5: Measuring the movement complexity, example (subdural data set,
Monkey K). Red dots indicate detected zero-crossings of the trajectory’s smoothed
derivative.
4.2. Asynchronous multi-limb (multi-finger) clinical data set 95
The preclinical data set was analysed session per session (24 sessions, 4 monkeys).
Each recording was split into training and test set: the first 70% of each session
were used for model training and validation, and the last 30% were used for testing.
4.2 Asynchronous multi-limb (multi-finger) clinical data

set
The ECoG clinical data set was acquired while human subjects were executing
sequential finger movements. It gathers data made publicly available for the IVth
BCI competition [Schalk et al., 2007] [Miller and Schalk, 2008] [Tangermann et al.,
2012] (http://www.bbci.de/competition/iv/) and data distributed by Stanford
University (https://purl.stanford.edu/zk881ps0522). The BCI competition
IV data set comprises data measured in three subjects. The Stanford data set
is composed of data acquired in nine subjects. All patients participated in a
purely voluntary manner, after providing written consent, under experimental
protocols approved by the Institutional Review Board of the University of Washington
(#12193). All patient data was anonymized according to IRB protocol, in accordance
with HIPAA mandate. These data originally appeared in the manuscript "Human
Motor Cortical Activity Is Selectively Phase-Entrained on Underlying Rhythms"
published in PLoS Computational Biology in 2012 [Miller et al., 2012]. As a result,
the full data set was composed of data acquired from 12 different subjects. A similar
protocol was utilized to acquire the IVth BCI competition and the Stanford data
sets.
4.2.1 Behavioural task

All subjects participated in one 10min-long acquisition session. Subjects were
instructed to perform finger extensions and flexions in response to visual cues. Each
cue was presented during 2 seconds, and indicated which finger had to be moved
(sequential movements). The subjects were asked to execute successive extensions
and flexions of the requested finger as long as the cue was on display. A blank screen
was displayed for 2 seconds after disappearance of the visual cue. Subjects were
instructed to relax during this inter-trial period. 30 cues were devoted to each finger.
The resulting 150 cues were randomly presented to the subjects.
4.2.2 Data acquisition

The 12 subjects were epileptic patients undergoing an ECoG-array implantation for
the localization of their epileptic focus. ECoG arrays (Ad-Tech, Racine, WI) were
implanted under the subjects’ dura mater (subdural implantation) [Miller and Schalk,
2008] [Miller et al., 2012]. They gathered 38 to 64 4mm-diameter electrodes spaced
out with a 1cm-inter-electrode distance. The location of the ECoG array was chosen
with respect to the monitoring purpose of the implantation. No information has
been given as to the precise localization of the arrays implanted into the subjects of
Figure 4.6: Behavioral task (reproduced from [Miller and Schalk, 2008]).
the BCI competition IV data set [Miller and Schalk, 2008]. The arrays of Stanford’s
data base covered various brain areas, namely frontal, parietal, occipital and/or
temporal areas (see details in the data base description file), i.e. both motor and
non-motor activity was measured by the electrodes. ECoG signals were acquired at
a sampling rate of 1kHz, and band-pass filtered between 0.15 and 200Hz.
Subjects were asked to move the fingers contralateral to the array location. A
data glove tracked the finger coordinates at a 25Hz sampling frequency. The tracked
coordinates were normalized and centred so that each finger trajectory exhibited
similar spatial amplitudes within an acquisition session. A trajectory example is
presented in Figure 4.7. Fingers are ordered as follows: thumb, index finger, middle
finger, ring finger and little finger.
4.2.3 Data set characteristics

One session of the Stanford data set was discarded because of the absence of NC
states (subject "jp"), and two more sessions were excluded because they were shorter
and therefore unsuited for the training and test of complex models (subjects "mv"
and "wm"). Details on the number of channels available in the remaining data sets
are shown in Table 4.3. The 9 remaining finger trajectories were analysed from two
points of view: duration of the NC and IC states (i) and stability of finger position
during NC states (ii).
(i) Duration of NC and ICi periods Both IC and NC states were theoretically
2s-long, as IC cues were displayed for 2 seconds and were followed by 2s-long blank
screens. Manual segmentation of the finger trajectories was used to measure the
empirical duration of NC and IC states (Table 4.4). NC and IC classes, where the
IC class is composed of the ICi class, i = 1, . . . , 5, were relatively well balanced.
(ii) Neutral position variability The stability of the NC positions was measured
for each finger i by averaging finger positions during NC periods and during ICj , j =
Ó
i, j ∈ {1, . . . , 5} periods (Table 4.5).
Similarly to the preclinical data set, the clinical data set was analysed session
4.2. Asynchronous multi-limb (multi-finger) clinical data set 97
finger 1 (a.u.)
8
6
4
2
0
-2
finger 2 (a.u.)
8
6
4
2
0
-2
finger 3 (a.u.)
8
6
4
2
0
-2
finger 4 (a.u.)
8
6
4
2
0
-2
finger 5 (a.u.)
8
6
4
2
0
-2
140 150 160 170 180 190 200
time (s)
Figure 4.7: Finger trajectories example (subject 1). The trajectory of each finger was
centred and normalized. Movement amplitudes are thus indicated with an arbitrary
unit (a.u.). Finger 1: thumb, Finger 2: index finger, Finger 3: middle finger, Finger
4: ring finger, Finger 5: little finger.
Table 4.3: Number of ECoG electrodes, clinical data set.
Data source Subjects Electrode number

1 62
BCI competition IV 2 48
3 64
bp 46
cc 63
ht 64
Stanford data set jc 47
wc 64
zt 61
Table 4.4: State sequence characteristics, clinical data set. The number of movements
and the intra-session average duration of NC and IC periods was measured for each
session of the clinical data set.
Data set Subject NC period duration (s) IC period duration (s) Movement number
1 1.10 ± 0.66 2.95 ± 0.94 148
BCI IV 2 1.85 ± 0.62 2.22 ± 0.73 147
3 1.12 ± 0.50 3.26 ± 2.42 137
bp 1.93 ± 0.78 2.21 ± 0.64 147
cc 1.32 ± 0.82 2.79 ± 1.05 148
ht 1.60 ± 0.79 2.62 ± 0.86 144
Stanford jc 1.74 ± 0.67 2.27 ± 0.31 132
wc 1.29 ± 0.94 2.85 ± 1.40 147
zt 1.21 ± 0.87 2.93 ± 0.81 147
Table 4.5: Variability of normalized finger i position during NC and ICj , j Ó= i, j ∈

{1, . . . , 5} states, clinical data set. The intra-session standard deviation of the finger
positions was measured for each session of the clinical data set. Deviations are
indicated in arbitrary units because normalized finger trajectories are considered.
Data set Subject Finger 1 Finger 2 Finger 3 Finger 4 Finger 5

1 0.22 0.43 0.33 0.47 0.33
BCI IV 2 0.43 0.54 0.36 0.59 0.49
3 0.35 0.58 0.70 0.75 0.33
bp 0.44 0.54 0.34 0.57 0.49
cc 0.33 0.58 0.72 0.76 0.34
ht 0.33 0.56 0.41 0.49 0.56
Stanford jc 0.19 0.27 0.17 0.42 0.43
wc 0.28 0.39 0.33 0.31 0.67
zt 0.21 0.43 0.36 0.48 0.34
4.3. Conclusion 99
per session (9 subjects, one session per subject). Likewise, each recording was split
into training and test set (70% and 30% of each session, respectively).
4.3 Conclusion
The data sets used to assess the ability of the Markov Switching Linear Model
(MSLM) to perform accurate asynchronous multi-DoF or multi-limb ECoG decoding
were presented in the present chapter. Both data sets are publicly available. The
first data set is composed by preclinical ECoG data acquired while monkeys were
executing asynchronous unimanual reaching movements. The second data set gathers
clinical ECoG data collected while human subjects were performing sequential multi-
fingers flexions and extensions. Next chapter consists of a presentation of the
methodology used to measure the MSLM performance when applied on the two
considered data sets.
Chapter 5
Evaluation and comparison
Contents
5.1 Comparative approaches for supervised MSLMs . . . . . . . 101
5.1.1 Markovian post-processed Wiener Filter . . . . . . . . . . . . 102
5.1.2 Switching Kalman Filter . . . . . . . . . . . . . . . . . . . . . 103
5.1.3 Static decoders . . . . . . . . . . . . . . . . . . . . . . . . . . . 104
5.2 Performance indicators . . . . . . . . . . . . . . . . . . . . . . 107
5.2.1 NC support for asynchronous mono- and multi-limb decoding . 107
5.2.2 Multiple IC state discrimination for sequential multi-limb asyn-
chronous decoding . . . . . . . . . . . . . . . . . . . . . . . . 108
5.2.3 Performance indicators for continuous dependent variable . . 108
5.2.4 Trajectory synchronization . . . . . . . . . . . . . . . . . . . 108
5.3 Statistical tests . . . . . . . . . . . . . . . . . . . . . . . . . . . 112
5.4 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 112
The Markov Switching Linear Model (MSLM) was compared to a post-processed

decoder and to a switching state-space filter. These two decoders were chosen so as
to represent two alternative strategies previously exploited for asynchronous and/or
multi-limb kinematic decoding. Performance indicators specific to discrete and
continuous variables were selected to measure the respective ability of each decoder
to handle (possibly multi-limb) asynchronous and accurate decoding, respectively. A
statistical procedure was chosen to assess the statistical significance of the observed
performance differences.
5.1 Comparative approaches for supervised MSLMs

The performance of the MSLM was first compared to the decoding performance of
a Markovian post-processed Wiener Filter (MpWF). This comparison permitted to
assess the interest of the first key feature of the MSLM, namely switching modelling,
when compared to the post-processing approach utilized in earlier studies. It was
then compared to another switching model, namely the Switching Kalman Filter
(SKF). This comparison allowed for the evaluation of the second key feature of the
MSLM, i.e. the discriminative framework, as opposed to the generative framework
utilized in a few BCI studies. Both the MpWF and the SKF performed dynamic state
detection. Variants of the MSLM and of the MpWF with static detection were finally
102 Chapter 5. Evaluation and comparison
considered to evaluate the interest of the MSLM’s Markovian hypothesis. They

are referred to as SLM (for Switching Linear Model) and pWF (for post-processed
Wiener Filter), respectively.
5.1.1 Markovian post-processed Wiener Filter

MSLM decoders combine several models to reckon with context-dependent modifica-
tions in the dependence between trajectory kinematics and neural features.
By contrast, the use of a single continuous decoder has been reported for
asynchronous [Wang et al., 2013b] [Velliste et al., 2014] and/or multilimb [Wang
et al., 2011] decoding in earlier BCI studies. The output of the single decoder was
post-processed so as to integrate NC and/or sequential multi-limb NC support into
regression models [Wang et al., 2011] or into state-space filters [Wang et al., 2013b]
[Velliste et al., 2014].
Formally, a first static or dynamic model f : Rm → Rn output an intermediary
estimate ŷtmp
t = f (xt ). A post-processing model v : Rn → Rn was then applied on
t
ŷtmp , and yielded the final estimate ŷt = v(ŷtmp
t ) (Figure 5.1). In some studies, the
post-processing operator v was additionally fed with the input signal (see Figure
5.1, blue arrows (ii)), e.g. in [Wang et al., 2013b] [Velliste et al., 2014]. State
detection was therefore performed using neural signals rather than intermediary
position or velocity estimates.
Figure 5.1: Post-processed kinematic decoder, general data flow. A regression model
or a state-space filter (additional orange arrows (i)) is fed with the neural features,
and, in the case of state-space filters, with statistical characteristics of the previous
estimate. A post-processing operator is applied on the output of this decoder.
The respective efficiency of MSLMs and of post-processed models depends on

the existence or absence of context-driven modifications of the dependence between
neural signals and kinematic parameters. If such modifications do exist, then the
training of a single continuous model may yield a suboptimal decoder. Kinematic
decoding accuracy may be degraded, for example during IC states. Moreover, if the
post-processing operator is not conditioned on neural signals, i.e. if it is exclusively
fed with the intermediary trajectory estimates ŷtmp
t , then state detection accuracy
depends on the quality of the trajectory estimates. Consequently, state detection is
also liable to be corrupted by the application of a suboptimal regression model.
5.1. Comparative approaches for supervised MSLMs 103
The respective relevance of switching and post-processed models for asynchronous

and multi-limb ECoG decoding is unclear. The comparison between switching models
and post-processed models has rarely been drawn. In [Williams et al., 2013], a
simple post-processing operator fed with velocity estimates and a switching model
which gate was fed with ECoG neural signals were compared for the integration of
NC support into the kinematic estimates issued by a linear regression model. The
authors concluded to the superiority of the second decoding strategy. This study
suggested the interest of switching models. The use of post-processing strategies has
nonetheless been reported for asynchronous [Wang et al., 2013b] and/or multi-limb
[Wang et al., 2011] ECoG decoding in several BCI studies.
To clarify the relevance of switching modelling, a Markovian post-processed
Wiener Filter (MpWF) was implemented for comparison purposes. The performance
of the MpWF was evaluated for two decoding tasks, namely mono-limb and multi-
limb (multi-finger) asynchronous trajectory reconstruction. The continuous model
was fitted on both NC and (potentially multiple) IC samples, and the post-processing
operator was exclusively fed with kinematic estimates (see Figure 5.2). This config-
uration was chosen to investigate the interest of the MSLM switching hypothesis for
both kinematic decoding during (possibly multiple) IC states and for state detection.
Similarly to the MSLM which embeds a dynamic gate, a Markovian hypothesis was
used by the post-processing operator, i.e. the post-processed operator performed dy-
namic state detection from the kinematic estimates yielded by the single continuous
model. This Markovian post-processing strategy permitted to discard the effect of
dynamic or static state detection when investigating the comparative efficiency of
the post-processing and switching approaches. The comparative quality of dynamic
and static post-processed-based state detection was nevertheless established using a
variant of the considered post-processed WF with static rather than dynamic state
detection (see section 5.1.3).
Figure 5.2: Markovian post-processed Wiener Filter decoder, data flow. The post-
processed operator performs dynamic state detection from the kinematic estimates
yielded by a single Wiener filter.
5.1.2 Switching Kalman Filter

The second competing decoder was chosen to assess whether switching regression
models outperform or are outperformed by switching state-space filters. The MSLM
combines regression models. Regression models are based on a discriminative-like

interpretation of the relationship between neural signals and kinematic parameters,
i.e. they model the probabilities P (yt |xt , z t = i, Θ). By contrast, the switching
state-space models considered in BCI studies are generative, i.e. they model the
probability P (xt , yt , z t = i, Θ) [Wu et al., 2004] [Srinivasan et al., 2007].
An analogue of the MSLM in the framework of switching state-space models
was here considered, namely the Switching Kalman Filter (SKF). The SKF (see
Appendix C) has been applied in several BCI studies [Wu et al., 2004] [Wu
et al., 2003a], and its use has been advocated for the integration of NC support
into state-space models [Srinivasan et al., 2007]. Similarly to the MSLM, it relies
on the assumption that its observation and transition models are conditioned on
a Markovian hidden state [Murphy, 1998]. Trajectory estimation is based on a
probabilistic combination of state-specific models [Murphy, 1998]. Because of this
hypothesis about neural signal generation, the value of the mixing coefficients
depends indirectly on the observed neural signals. In contrast with the MSLM,
SKF state estimation is based on the consistency of both its emission and transition
models with the observed neural signals (see Figure 5.3 and section C.2.3 in
Appendix C).
5.1.3 Static decoders

The interest of the Markovian hypothesis used to perform dynamic state detection
was investigated by comparing the MSLM and the MpWF with the SLM and
pWF, i.e. variants of the MSLM and MpWF with static rather than dynamic state
detection. The corresponding data flow are shown in Figures 5.4 and 5.5.
5.1. Comparative approaches for supervised MSLMs 105
Figure 5.3: SKF (A) and MSLM (B) data flow. A. The weights used to combine the
KF experts depend on the adequacy of each KF with the current and past observed
neural signals. B. The weights used to combine the linear experts directly depend
on the value of current and past observed neural signals.
Figure 5.4: Post-processed Wiener Filter decoder, data flow. The post-processed
operator performs static state detection from the kinematic estimates yielded by a
single Wiener filter.
Figure 5.5: Switching Linear Model decoder, data flow. The gate performs static
state detection. The SLM thus corresponds to a generic Mixture of Experts model.
5.2. Performance indicators 107
5.2 Performance indicators

Performance assessment was performed on test sets independent from the training
data sets. Indicators associated with binary decoding were first computed to monitor
the ability of each decoder to support NC states when performing asynchronous
mono-limb or multi-limb decoding. When multiple IC experts were considered, the
multiple IC states ICj , j = 1, . . . , K were pooled into a general IC state, and an
additional indicator was introduced to measure the ability of the state decoder to
distinguish between the multiple ICj , j = 1, . . . , K states when detecting an IC state.
The accuracy of the kinematic estimates during IC periods was finally established
via a set of indicators suited for continuously-valued variables.
5.2.1 NC support for asynchronous mono- and multi-limb decod-

ing
Two groups of performance indicators were used to measure the quality of the
NC support associated with asynchronous mono- and multi-limb decoding. A first
indicator set characterized NC/IC classification accuracy (i.e., number of correctly
classified samples). In the case of multi-limb decoding, ICi , i = 1, . . . , K states were
pooled into a global IC state. A second indicator set focused on the number of false
activations/deactivations and on their durations.
Confusion matrix-based performance indicators The indicators used to as-

sess the performance of binary decoders [Mason et al., 2006] mainly rely on the
confusion matrix. The confusion matrix gathers the number of NC states which
are correctly (True Negatives, T N ) or wrongly (False Positives, F P ) labelled, and
the number of IC states which are correctly (True Positives, T P ) or wrongly (False
Negatives, F N ) labelled by the decoder. Several indicators are derived from the
P +F N
confusion matrix. The classification error ERR = T P +TFN +F P +F N and the ac-
T P +T N
curacy ACC = T P +T N +F P +F N are relevant measures of the global classification
quality when classes are well balanced [Mason et al., 2006] [Thomas et al., 2013].
The True Positive Rate T P R = T PT+F P
N (sensitivity) and the False Positive Rate
FP
F P R = F P +T N additionally permit to reflect the classifier ability to correctly
decode IC and NC states, respectively.
The NC and IC classes were relatively well balanced in the preclinical and
clinical data sets (see Chapter 4). The accuracy ACC, the TPR and the FPR were
therefore computed. In the case of mono-limb decoding, the Area Under the Curve
(AUC) [Davis and Goadrich, 2006] was additionally computed to reflect the decoding
performances which would be obtained by using a different prior probability for
each state, i.e. by using a probability threshold different from 0.5 to assign labels to
observation samples.
Block-wise indicators: frequency and duration of false detections The

TPR, FPR, ACC and AUC are sample-based indicators. They don’t take into account
the dynamic of misclassified samples, even if it is liable that several consecutive

misclassified samples are less disturbing to users that the same number of isolated
misclassified samples. State decoding performance was thus considered in terms of
false activations/deactivations’ duration and frequency, where a "false activations"
(respectively, a "false deactivation") refers to a block of consecutive NC samples
misclassified as IC samples (respectively, a block of IC samples mistaken for NC
samples). A block of misclassified samples was counted as one false activation or
deactivation. Its duration was computed as the number of samples divided by the
decision rate. These block-wise criteria are expected to reflect the users’ evaluation
of the quality of NC support during asynchronous closed-loop control sessions.
5.2.2 Multiple IC state discrimination for sequential multi-limb

asynchronous decoding
When multiple IC states are considered, it may additionally be needed to assess the
decoder ability to discriminate between the K different IC states, ICi , i = 1, . . . , K.
In the case of multi-limb movements for example, it is desirable to check whether
the activation of one limb can be distinguished from the movement of the other
limbs. To this aim, the ratio of correctly classified TP samples was established
where appropriate. A K × K confusion matrix was computed on TP samples, and
the sum of its diagonal terms was divided by the total number of IC samples (see,
e.g., [Gouy-Pailler et al., 2009]).
5.2.3 Performance indicators for continuous dependent variable

The accuracy of continuous variable estimates is typically assessed via the Pear-
son Correlation Coefficient (PCC) and the Normalized Root-Mean-Squared Error
(NRMSE) [Spuler et al., 2015]. The PCC measures the amount of linear depen-
dence between observed y and predicted ŷ variables: PCC(y, ŷ) = cov(y,ŷ) σy σŷ , where
cov(y, ŷ) indicates the covariance between y and ŷ and σy refers to the standard
deviation of y. The NRMSE measures the ℓ2 -error between the vectors of observa-
tions y = (y 1 , y 2 , . . . , y T )′ and ŷ =ñ
(ŷ 1 , ŷ 2 , . . . , ŷ T )′ , where (.)′ denotes the transpose:
ëy−ŷë2 qT 1 qT
NRMSE = ëy−ȳë2 , where ëyë2 = t=1 (y
t )2 and ȳ = T t=1 y .
t
ëy−ŷë1
The Normalized Mean Absolute Error NMAE = ëy−ȳë1 , where ë.ë1 is the ℓ1 -
qT
norm (ëyë1 = t=1 y ), was additionally computed. It issues a measure of the
- t-
- -
ℓ1 -error between y and ŷ and is less sensitive to outliers than the NRMSE [Hyndman
and Koehler, 2006].
5.2.4 Trajectory synchronization

Temporal uncertainties are often observed in BCI estimated trajectories, for example
because of the temporal instability of neural signals [Tranquillo, 2013] or of the
limited temporal resolution of the BCI decoders. Generic distance-based performance
indicators are highly sensitive to temporal uncertainties, e.g. shift or time warping
[Cassisi et al., 2012] [Wang et al., 2013d]. Slight positive and negative delays between
estimated and target time series, for example, substantially degrade distance-based
performance indicators, especially the ℓ2 -distance between both time series. Such
performance indicators have consistently be shown to be suboptimal when they are
used to estimate the similarity between time series, and when the resulting similarity
measures are utilized to train a time series classifier [Wang et al., 2013d]. Although
large delays between intended and estimated trajectories, e.g. 300ms pure delay
[Willett et al., 2013], are known to significantly degrade BCI control performances
[Willett et al., 2013] [Marathe and Taylor, 2015], it is unclear whether slight temporal
uncertainties are more or less disturbing to users than spatial errors, which have
been shown to degrade control quality [Marathe and Taylor, 2015]. Performance
indicators insensitive to temporal uncertainties were thus added to the set of generic
indicators presented in section 5.2.3.
Various indicators have been proposed to measure the similarity between time
series [Wang et al., 2013d], e.g. the Dynamic-Time-Warping or the Longest Common
Subsequence similarity measures [Cassisi et al., 2012]. In particular, Dynamic-Time-
Warping (DTW) criteria have been used alonside traditional ("lock-step" [Wang
et al., 2013d]) indicators to assess the performance of several BCI decoders in
[Eliseyev and Aksenova, 2014]. Let us consider two times series x ∈ RT and y ∈ RT
of similar length T . Dynamic Time Warping finds the T̃ -long path P̃ ∈ NT̃ ×2
which minimizes the cumulative distance T̃k=1 d(xp̃k,1 , yp̃k,2 ), where d is a "local cost
q
measure" [Muller, 2007], e.g. the ℓ2 - or ℓ1 -distance. The admissible paths can be
limited by constraining the maximum delay between two paired samples, e.g. using
the Sakoe-Chiba band [Cassisi et al., 2012]. An illustration of the errors reflected by
lock-step and DTW-based indicators is shown in Figure 5.6.
Figure 5.6: Generic lock-step- (A) and DTW-matching (B).
Synchronized signals are defined on a longer, distorted temporal scale defined

by the DTW optimal path (see Figure 5.7.A). Although the cumulative distance
DT W (x, y) = minP̃ T̃k=1 d(xp̃k,1 , yp̃k,2 ) associated with the optimal DTW path can
q
be used as a criterion, it does not reflect the impact of desynchronization on tradi-

tional performance indicators, e.g. the PCC, NRMSE or NMAE. A method is here
proposed for desynchronization-robust direct computation of generic performance
indicators between two trajectories x ∈ RT and y ∈ RT . The indicators of interest
are computed after having obtained signals synchronized in the original time scale
of one of the trajectories, taken as the reference signal. This approach makes direct
comparisons of performance indicators between raw and synchronized trajectories
possible.
First, synchronized signals x̃ ∈ RT̃ and ỹ ∈ RT̃ are obtained along the DTW-
distorted time scaled defined by the DTW-optimal path P̃ = argminP̃ T̃k=1 d(xp̃k,1 , yp̃k,2 ).
q
The DTW path cannot be directly interpreted as a generic function f : NT → NT

because several outputs can be associated with an input value, i.e. the DTW
path can be seen as a multi-valued map [Kuratowski, 2014] (Figure 5.7.B). For
this reason, the DTW path P̃ ∈ NT̃ ×2 was approximated by a function-like path
P ∈ NT ×2 . This path is obtained by computing the average of the signal y over the
(possibly multiple) instants associated with an original instant (see Figure 5.7.C).
A synchronization example is shown in Figure 5.8. The ℓ2 -distance was used
to synchronize the estimated trajectory before computing the PCC and NRMSE
because these two performance indicators are based on the ℓ2 -distance. Similarly,
ℓ1 -based synchronization was performed to assess the NMAE between synchronized
signals. Maximum lags of 100ms or 200ms (1 and 2 samples, respectively) were
utilized to constrain the considered DTW paths (see Chapter 7).
Figure 5.7: A. Example of a DTW path, clinical data set. DTW was applied
on a true and estimated finger trajectory. The initial part of the DTW path
(corresponding to the first 18 samples of the reference (true) trajectory) is here
displayed. Signals are synchronized when they are evaluated on the distorted time
scales represented by the DTW path. B. Correspondence between instants of the
time scale 1 and 2. Most instants defined on the first (reference) time scale are
associated with a single instant of the second time scale (yellow, blue and green
arrows). In some cases, however, several instants of the second time scale correspond
to the same original instant (red circles). C. Approximation of the DTW path by a
function-like path.
300
250
200
y 3 (mm)
150
100
50
0
40 50 60 70 80 90 100
time (s)
observed estimated estimated after synchronization
Figure 5.8: Synchronization example, MSLM decoder, subdural data set. The
proposed synchronization method was used to synchronize the MSLM trajectory
estimate (blue line) with the observed trajectory (black line). The ℓ1 -distance and a
maximum lag of 300ms (3 samples) were used to compute the synchronized MSLM
estimate (red line).
5.3 Statistical tests

A statistical procedure for the comparison of several decoders on multiple data sets is
here necessary. Following the recommandations of [Demšar, 2006], the significance of
performance differences was established using the Friedman test with the significance
level α = 0.05.
The Friedman test permits to analyse repeated measures performed over the same
samples, here for the analysis of the performance of D ≥ 2 different decoders applied
on the same acquisition sessions. It ranks the decoders on each acquisition session
according to their respective performance [Demšar, 2006]. It then tests whether the
decoders’ respective averageèranks are significantly different. To this aim, the test
D(D+1)2
q 1 q j 2 é
12N
statistic tF riedman = D(D+1) j(N i ri ) − 4 , where rij is the rank of the
decoder j for the acquisition session i, is computed. Here, D = 3 is the number of
compared decoders and N is the number of acquisition sessions (N = 8, N = 16 and
N = 9 for the subdural, epidural and finger data sets, respectively). The tF riedman
statistic follows a χ2F law with D − 1 degrees of freedom for large enough N and
D (typically, N > 10, D > 5) [Demšar, 2006]. Adjusted critical values have been
proposed for small N and D, and have been used for the subdural data set (N = 7).
The implementation made publicly available by [Cardillo, 2009] was used.
Various procedure have been proposed for post-hoc testing, e.g. the Nemenyi
Post-hoc test [Demšar, 2006] [Pereira et al., 2015], Bonferroni-Dunn, or step-wise
procedures such as Holm’s, Hochberg’s or Hommel’s procedures. In [Benavoli et al.,
2015], it was advocated to avoid post-hoc tests based on the mean-ranks issues by
the Friedman test. Post-hoc pairwise decoder comparisons were thus performed
using the sign test with Bonferroni correction [Benavoli et al., 2015].
5.4 Conclusion
In this chapter, it is proposed to compare the Markov Switching Linear Model
(MSLM) with a Wiener Filter with Markovian post-processing (MpWF) and with
a Switching Kalman Filter (SKF), which combines K Kalman filters. The MpWF
and SKF permit to investigate the relevance of key features of the MSLM, namely
switching and discriminative modelling. It is finally proposed to specifically evaluate
the impact on state detection of the third MSLM’s key feature, i.e. dynamic gating,
by considering a Switching Linear Model (SLM) and a post-processed Wiener Filter
(pWF). The SLM and pWF are variants of the MpWF and the MSLM with static
state detection.
A set of performance indicators has been chosen to assess the ability of the
decoders to perform asynchronous mono-limb and sequential multi-limb trajectory
decoding and to accurately reconstruct kinematic parameters during IC periods. The
quality of NC support is traditionally measured by means of confusion-matrix-based
performance indicators. It is here proposed to complement these indicators by block-
wise performance indicators. Block-wise indicators take into account the dynamic
of false activations. They are therefore expected to reflect the users’ evaluation of
5.4. Conclusion 113
the quality of NC support during asynchronous closed-loop control sessions. The

performance indicators associated with kinematic reconstruction during Intentional
Control (IC) periods generally measure both temporal and spatial accuracy. Because
desynchronizations between observed and estimated trajectories are likely to impact
performance indicators, it is proposed to compute generic performance indicators
between both raw (desynchronized) and synchronized versions of the observed and
estimated trajectories. An approach has been designed to compute synchronized
trajectories which are defined on the original time scale of the observed trajectory.
Finally, a statistical procedure has been chosen to assess the significance of the
differences between the decoders’ respective performances.
Chapter 6
Implementation
Contents
6.1 Feature extraction and pre-processing . . . . . . . . . . . . . 115
6.1.1 Kinematic parameter extraction . . . . . . . . . . . . . . . . 115
6.1.2 Neural signal feature extraction . . . . . . . . . . . . . . . . . 116
6.2 Decoder implementation . . . . . . . . . . . . . . . . . . . . . 117
6.2.1 Preclinical data set . . . . . . . . . . . . . . . . . . . . . . . . . 117
6.2.2 Clinical data set . . . . . . . . . . . . . . . . . . . . . . . . . 123
6.3 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 125
Transducers embedding a Markov Switching Linear Model (MSLM) were imple-

mented for two decoding tasks, namely asynchronous wrist trajectory reconstruction
(preclinical data set) and multi-limb (multi-finger) trajectory reconstruction (clinical
data set). Wiener Filters with Markovian post-processing (MpWF) and Switching
Kalman Filter decoders (SKF) were additionally implemented for comparison pur-
poses (see section 5.1). Application of the MSLM, MpWF and SKF decoders was
preceded by the extraction of neural and kinematic features.
6.1 Feature extraction and pre-processing

Features were extracted from the subjects’ ECoG neural signals and from their
wrist’s or fingers’ kinematics.
6.1.1 Kinematic parameter extraction

Kinematic encoding in ECoG features has mainly been disclosed for direction,
position, velocity and speed (see section 1.2.3.2). Although speed decoding has
yielded promising results [Bundy et al., 2016] [Hammer et al., 2016], speed alone
cannot enable users to control a BCI effector. While ECoG directional tuning
has been explored in several studies [Schalk et al., 2007] [Anderson et al., 2012]
[Nurse et al., 2015], kinematic reconstruction from ECoG signals has mainly been
investigated for position and/or velocity [Chao et al., 2010] [Shimoda et al., 2012]
[Marathe and Taylor, 2013] [Pistohl et al., 2008] [Spüler et al., 2016] [Liang and
Bougrain, 2009] [Chen et al., 2014b] [Flamary and Rakotomamonjy, 2012]. In
particular, the studies previously led on the preclinical [Chao et al., 2010] [Shimoda
116 Chapter 6. Implementation
et al., 2012] and clinical data sets [Liang and Bougrain, 2009] [Chen et al., 2014b]
[Flamary and Rakotomamonjy, 2012] considered in the present dissertation focused
on position reconstruction with Wiener filtering or variants. Position decoding was
therefore utilized to assess the performance of the MSLM.
A state variable composed of both position and velocity was used to implement
the SKF, as it had been reported as optimal for Kalman filtering of ECoG signals
in an earlier study [Pistohl et al., 2008].
Preclinical data set The NHPs’ wrist position yt ∈ R3 was issued by the motion
tracking system. Velocity was derived from position using a forward-difference
approximation [Eberly, 2014]. NC and IC states were manually labelled.
Clinical data set Finger positions yt ∈ R5 were acquired by the data glove worn
by the subjects. Similarly to the preclinical data set, velocity was derived from
position using a forward-difference approximation. NC and ICi , i = 1, . . . , IC5 states
were manually labelled.
6.1.2 Neural signal feature extraction

A high-dimensional time-frequency-space representation was chosen to exploit the
position- and velocity-tuned features disclosed in ECoG signals.
Time-frequency features were extracted for each channel from ∆t-long ECoG
sliding epochs (∆t = 1s, sliding step: 100ms) following [Eliseyev and Aksenova, 2014].
A Complex Continuous Wavelet Transform (CCWT) was applied on neural features.
The complex Morlet wavelet (Figure 6.1), which had been found efficient for EEG
and ECoG neural signal analysis in earlier studies [Lemm et al., 2004] [Eliseyev and
Aksenova, 2014], was chosen in the present study. The ECoG frequency content was
analysed between 1 and 250Hz to exploit the frequency bands in which kinematic
tuning had been disclosed (see section 1.2.3.2). Redundant sampling of this
frequency domain was achieved via 38 daughter wavelets chosen with a logarithmic
scale. A logarithmic transform was applied on the CCWT transformed signals’
absolute value. The average value of the log-transformed signals was computed in
100ms sliding windows (100ms step), resulting in a 10-points description of ECoG 1s-
long time bins for each frequency band and each channel. Low frequency components
have been shown to be particularly informative for kinematic reconstruction from
ECoG signals (see section 1.2.3.2). They were extracted by means of a Savitzky-
Golay filter [Schafer, 2011] (window length of 200ms, order 2), and added to the
CCWT-based frequency features. Thus the ECoG epoch [t − ∆t : t] was described by
the temporal-frequency-spatial feature vector xt ∈ Rm , where m = nc × 10 × (38 + 1).
Regarding the preclinical data set, nc = 32 for Monkey A and nc = 64 for Monkeys
K, B and C. Because of differences in the number of electrodes implanted in each one
of the subjects (see section 4.2.3), 46 6 nc 6 64 for the clinical data set. Finally,
the pre-processing approach proposed in [Eliseyev and Aksenova, 2014] has been
6.2. Decoder implementation 117
used to detect potential artefacts in the neural data (e.g., chewing artefacts in the
epidural data set), and to replace them with neutral values.
0.5
-0.5
-8 -6 -4 -2 0 2 4 6 8
Real part Imaginary part
Figure 6.1: Real and imaginary parts of a Morlet wavelet.
6.2 Decoder implementation

The three decoders (MSLM, MpWF and SKF) were trained on the first 70% of each
session of the preclinical and clinical data set.
6.2.1 Preclinical data set

MSLMs with K = 2 experts were first implemented to explore the integration of
NC support into mono-limb kinematic decoders. As NC and IC were labelled in the
training data set, a supervised training approach was used. A MpWF and a SKF
with K = 2 experts were trained for comparison purposes. The training of MSLMs
with K > 2 experts was additionally performed for complex reaching movements.
As the labels associated with each IC experts were unknown, unsupervised training
of the MSLMs was carried out.
6.2.1.1 NC/IC MSLM (supervised training)

MSLM decoders with K = 2 experts, namely a NC and an IC expert, were imple-
mented for the integration of NC support into a continuous mono-limb trajectory
decoder. Supervised Maximum Likelihood training of the MSLM decoders was
performed using complete training data sets {X, Y, z}.
Experts The NC expert yielded the NC neutral position value ȳN C , which was
estimated as the average value of yt computed over NC states. The IC expert was
dedicated to wrist position decoding during IC periods. ML training of the IC expert
T
parametrized by BIC consisted in maximizing δz t ,IC ln P (yt |xt , BIC ), where T
q
t=1
refers to the number of training samples, δz t ,IC = 1 if the observation t corresponds
to an IC period, and δz t ,IC = 0 otherwise. A Gaussian noise was associated with
the IC expert. The corresponding Ordinary Least Squares (OLS) solution to this
maximization problem is unstable because of the high dimension of the input variable
xt . Several approximate solutions have been proposed to identify high-dimensional
linear models, in particular Principal Component Regression, Partial Least Squares
and LASSO training. A pilot study was therefore performed to select a relevant
approximate ML estimator of the IC expert (see Appendix B). Partial Least
Squares (PLS) [Höskuldsson, 1988] regression was found to provide accurate IC
experts, and was therefore chosen to identify the IC experts. The optimization
of the subspace dimension relied on a 6-fold cross-validation procedure completed
on the training data set, namely on the application of Wold’s R criterion on the
cross-validated PRESS statistic [Li et al., 2002].
Gate Fully supervised training of the MSLM involves the identification of HMM-
based dynamic state decoder (gate) (see section 3.3.1). In the considered case of
mono-limb asynchronous decoding, the gate has to distinguish between 2 states,
namely NC and IC states.
The state-specific emission distributions associated with the HMM-based gate
of the MSLM can be modelled by means of generic distributions or by using an
t t)
alternative approach based on discriminative modelling: P (xt |z t = j) ∝ PP(z(z=j|x
t =j) ,
where P (z t = j|xt ) is yielded by a discriminative state decoder (e.g., using Neural

Networks [Renals et al., 1994] [Bourlard and Morgan, 1998] [Ordóñez et al., 2013]
or Support Vector Machines [Valstar and Pantic, 2007] [Ordóñez et al., 2013], see
section 3.2.1). Thus, both generative and discriminative state decoders can be
embedded into the dynamic HMM-based gate of the MSLM. Both types of model
have been applied for state detection in BCI studies (see section 2.4.2). Their
respective performances has often been reported as dependent on the problem
at hand. In particular, decoding efficiency may depend on the distribution of
the neural features within each state. If for example this distribution cannot be
satisfyingly modelled with a generic distribution like a Gaussian or Mixture of
Gaussians distribution, using a discriminative approach may permit to improve
state detection. A preliminary study was therefore completed to assess the relevance
of different modelling approaches and select the strategy which was optimal for
NC/IC discrimination on the preclinical data set (see Appendix B). For the sake
of simplicity, this study was carried out on static decoders. The selected modelling
approach was subsequently embedded into the HMM-based dynamic gate of the
MSLM.
The set of investigated classifiers included Linear Discriminant Analysis (LDA)
and Support-Vector-Machine (SVM), which are regularly applied to distinguish
between neural states in motor BCIs [Fifer et al., 2014] [Kapeller et al., 2015]
[Hotson et al., 2016] [Hortal et al., 2015]. It additionally comprised nonlinear
counterparts of the LDA and of the SVM, namely Quadratic Discriminant Analysis
(QDA) and a nonlinear SVM. Although it has been comparatively unfrequently
applied in BCI studies, high performance of Logistic Regression-based (LR) state
classification has been reported in recent EEG- and ECoG-based studies [Bashashati
et al., 2015] [Bundy et al., 2016]. The relevance of LR was therefore additionally
assessed for the considered binary classification task. LDA and QDA are generative
classifiers relying on Gaussian distributions (see Chapter 2). They are thus static
analogues of generic HMMs with Gaussian emission distribution. By contrast,
LR and SVMs are discriminative classifiers (see Chapter 2). They can be seen
as static analogues of HMMs exploiting a discriminative approach to model the
emission probabilities. Both linear (LDA, SVM and LR) and nonlinear (QDA and
nonlinear SVM) classifiers were considered because linear state detection is liable to
be suboptimal if neural signals are not linearly separable.
The neural input feature variable was here high dimensional. As classifier per-
formances have been shown to depend on the characteristics of the input variable
[Bashashati et al., 2015], in particular on its dimensionality [Bhattacharyya et al.,
2011], the classifiers were tested after reduction of the feature dimension. Different
projection-based dimensionality reduction procedures were compared. Both unsu-
pervised and supervised projections were considered, namely PCA- and PLS-based
dimensionality reduction. Computation of the PLS-based projector was found by
fitting a PLS model between the high-dimensional input variable xt and the state
variable z t ∈ {0, 1}. A detailed presentation of the preliminary study can be found
in Appendix B.
PLS-based dimensionality reduction followed by the application of a logistic
model was found to be efficient for NC/IC detection (see Tables B.6 and B.7).
It was consequently integrated into the HMM-based gating network for dynamic
state detection. 6-fold cross-validation was used to find the optimal dimension of
the latent subspace yielded by a PLS regression between xt and z t . Optimization
was completed by applying Wold’s R criterion on the 6-fold cross-validated PRESS
statistic [Li et al., 2002]. The logit model parameters were fitted using the Iteratively
Reweighted Least Squares algorithm [Bishop, 2006].
A Switching Linear Model with static state detection was additionally inferred
from the MSLM gate and experts. It was obtained by discarding the Markovian
hypothesis used by the MSLM’s gate, i.e. by exploiting the static LR embedded into
the MSLM’s HMM-based gate to combine the experts identified during the training
of the MSLM.
6.2.1.2 Post-processed Wiener Filters
A Markovian post-processed Wiener Filter was fitted via PLS regression on the
training data set {X, Y}. The optimal number of PLS factors was estimated by
applying Wold’s R criterion on the 6-fold cross-validated PRESS statistic. Trajectory
post-processing was used to integrate NC support into the decoder. A logistic
regression was trained to infer NC and IC states from the kinematic estimates
ŷt . Similarly to the MSLM, a Markovian hypothesis was used to limit spurious
detections during both NC and IC states. The corresponding transition matrix was
found using the procedure presented in Chapter 3 in the case of the MSLM. The
state probabilities P (z t |ŷ1:t ) yielded by the dynamic LR state decoder were used to
weight the Wiener estimates and the neutral values associated withîNC states.ï The
dynamic logistic regression was identified on the training data set Ŷ, z .
The parameters of a Wiener Filter with static post-processing (pWF) were
additionally inferred from the parameters of the MpWF. The pWF combined the
LR and Wiener models identified during the training of the MpWF, i.e. the outputs
of the LR embedded into the dynamic post-processing operator of the MpWF were
directly used to weight the Wiener estimates and the neutral values associated with
NC states.
6.2.1.3 SKF
A SKF (see Appendix C) was implemented for dynamic combination of K = 2

Kalman filters, one specialized in NC periods and the other in IC periods. The SKF
state variable (i.e., response variable) ySKF
t was chosen as the monkey’s wrist position
and velocity because it had been reported as optimal for ECoG decoding [Pistohl
et al., 2008]. The neural features xt fed to MSLM decoders embedded a 1s-long
history of ECoG signal, i.e. they consisted of frequency-space features observed in
ten 100ms-long time bins. By contrast, the neural features xSKFt considered by SKFs
were only observed within an 100ms-long time bin, because SKFs rely on an explicit
approach to perform dynamic modelling. Similarly to the MSLM state decoder, the
dimension of the neural features was reduced before application of the SKF. PCA-
and PLS-based dimensionality reduction procedures were compared in a pilot study
presented in Appendix C, which additionally permitted to find the optimal lag
between the neural feature time bin and the kinematic parameters of the monkeys’
wrist. PLS-based dimensionality reduction of the neural features associated with the
last 100ms before the instant t considered for trajectory estimation (10th time bin of
the neural features xt ) was found to correspond to the best KF decoding accuracy,
and was therefore used to limit the SKF computational cost. PLS regression between
t
xSKF and ySKF
t was thus used to identify an informative low-dimensional subspace.
While neural features were normalized after feature extraction, a constant term
was nevertheless added to the neural features to account for modifications of the
baseline activity during NC and IC states. Subspace dimension was chosen by 6-fold
cross-validation on the training data set. OLS estimates of the transition Ak and
emission Ck matrices were computed on the training data sets {XSKF,k , YSKF,k }.
ML estimates of the variance matrices Γk and Σk were found following [Aggarwal
et al., 2013].
6.2.1.4 MSLM with multiple IC experts (unsupervised training)
Some reaching movements of the preclinical data set exhibit complex behaviours,
where complexity is measured by the number of zero-crossing of the trajectory
first derivative (see Chapter 4). Subdural sessions with an average complexity
superior to 2 were selected for an additional analysis (three sessions of the Monkey
A), namely for the training of a MSLM with 2 IC experts. Unsupervised Maximum
Likelihood training of the MSLM decoders was performed using the incomplete
training data sets {X, Y}. Both NC and ICi , i = 1, 2 states were considered as
hidden during training. The decision of using unconstrained NC/IC state labels was
taken because of labelling uncertainties, for example around transitions or for very
short null-velocity periods during manually identified movements.
Structure A MSLM decoder with K = 3 experts was implemented to improve

kinematic modelling during reaching movements in the context of asynchronous
mono-limb kinematic decoding. One expert was specialized in NC periods, and
the K − 1 = 2 other experts were dedicated to complex wrist kinematic decoding
during IC periods. The experts were assumed to be associated with a Gaussian
noise. Following the observation that penalized LR was optimal for multi-class
classification (see Appendix B, clinical data set), a 3-class LR was here considered
to compute the emission probabilities of the MSLM HMM-based gate.
Initialization The EM-algorithm converges to a local solution [Roche, 2011]. Pa-

rameter initialization is therefore crucial. While clustering algorithms are frequently
used for parameter initialization, generic clustering algorithms such as Mixtures of
Gaussian and k-means are generally not suitable for high-dimensional data [Par-
sons et al., 2004]. High-dimensional clustering involves complex and potentially
time-consuming procedures [Parsons et al., 2004] [Bouveyron and Brunet, 2013].
For this reason, initialization was here performed by dividing movements into K-1
= 2 parts (see Figure 6.2), and by using the corresponding state sequence to
perform supervised training of the MSLM gate and experts. This initialization
approach permitted to exploit the a priori knowledge that models are likely devoted
to different phases of the movement. Similarly to the supervised NC/IC MSLM
presented above, an approximate solution to the initial ML training of the MSLM
was considered because of the high-dimension of the neural subspace. PLSR was
used to fit the experts, and LASSO training was used to identify the parameters
of the LR which permitted to model the emission probabilities. Both gate and
expert training depend on the value of a hyperparameter, namely the amount λ of
penalization for LASSO training of the gate and the number F of PLS factors for the
experts. Both hyperparameters were selected via 6-fold cross-validation performed
during the gate and expert initialization. The amount λ of penalization was selected
so as to minimize the cross-validated error. Similarly, the number F of expert factors
was found by applying Wold’s R criterion on the cross-validated PRESS statistic.
The considered hyperparameters were not refitted during the M-step iterations.
250
Axis 1 (mm)
200
150
100
50
200
Axis 2 (mm)
-200
400
Axis 3 (mm)
200
0
3000 3100 3200 3300 3400 3500 3600 3700 3800
Time (s)
Figure 6.2: Initialization of the EM algorithm, example (subdural data set). Each
color refers to a different state.
E-step and M-step The E-step and M-step presented in Chapter 3 were itera-
tively performed to train the MSLM. Specific optimization methods are generally
necessary to solve the M-step maximization problems because the impact of the ob-
servations t on a parameter associated with state i is weighted by the corresponding
state probability γit = P (z t = i|x1:T , y1:T , Θ) (see section 3.3.2.3).
Weighted expert training While the Weighted Linear Least Squares (WLLS)
estimator B̂k = (X′ Γk X)−1 X′ Γk Y, where the weight matrix Γk ∈ RT ×T is a
diagonal matrix with the diagonal elements γkt , t = 1, . . . , T [Moerland, 1997], is
regularly used to train Mixtures of Experts, it is not defined when m ≤ T . For
this reason, a weighted variant of the PLS-based approximate solution presented in
Appendix B was here utilized, namely weighted PLS. A PLS-based approximate
1 1
ML estimator of expert k was found by fitting a PLSR between Γk2 X and Γk2 Y
[Souza and Araújo, 2014]. Similarly to the supervised training case, the NC expert
yielded the NC neutral position value ȳN C .
Weighted gate training Similarly, weighted LR training was required to

update the gate parameters at each M-step. Procedures for weighted training of
LR models are for example used for EM-based training of Mixtures of Experts
with a LR gate (e.g., [Waterhouse, 1997] [Moerland, 2000]). As generic estimators
of LR models are liable to be unstable [Bishop, 2006], and may not exist when
high-dimensional features are considered, a LASSO weighted training procedure
was utilized to update the parameters of the LR embedded into the HMM’s gate.
Similarly to the comparative study presented in Appendix B, the efficient training
approach [Friedman et al., 2010] implemented in the open-source glmnet toolbox
[Qian et al., 2013] was used to perform the LR weighted LASSO training. It should
be noted that the utilization of such iterative optimization approaches results in
the presence of an "inner-loop" within each M-step, and significantly increases the
computational cost of the EM algorithm (hence the LDA-based gating network
presented in [Xu et al., 1994] or the approximate approach proposed in [Moerland,
1997]).
Convergence A stopping criterion was used to terminate the EM training

when the data log-likelihood was not improved by additional iterations.
6.2.2 Clinical data set

MSLMs with K = 6 experts were implemented to explore the ability of the MSLM
to handle asynchronous sequential multi-limb kinematic decoding. Similarly to
the study led on the preclinical data set with binary MSLM, a supervised training
approach was used because NC and ICi , i = 1, . . . , 6 states were labelled in the
training data set. A MpWF and a SKF with K = 6 experts were trained for
comparison purposes.
6.2.2.1 MSLM
A MSLM decoder with K = 6 was implemented for asynchronous finger trajectory

decoding. The IC expert ICi was associated with periods during which finger i was
moving, i = 1, . . . , IC5 .
Experts Similarly to the preclinical data set, the NC expert yielded the NC neutral
position value ȳN C , estimated as the average value of yt computed over NC states.
The ith IC state referred to periods during which movements were executed by the ith
finger. A linear expert was trained for each IC state, IC1 , . . . , IC5 . LASSO training
was found optimal to identify the parameters of the experts in the preliminary study
presented in Appendix B. 6-fold cross-validation was used to choose the amount λ
of ℓ1 -penalization. Similarly to the preclinical data set, the parameters of the SLM
were inferred from the MSLM parameters.
Gate A preliminary study was again completed to select the best state decoder
for the considered gating task. Multi-class decoders were considered, as opposed to
the binary decoders considered for the preclinical data set. LR fitted on the whole
feature set using a LASSO procedure was found to surpass the other multi-class de-
coders (see Appendix B). It was therefore integrated into the MSLM recurrent gate.
Supervised training of the MSLM decoders was performed using complete training
data sets {X, Y, z}. As a prior was set on the gate and expert parameters, it
corresponded to Maximum A Posteriori rather to Maximum Likelihood training.
Similarly to the preclinical data set, the parameters of the SLM decoders were
inferred from the MSLM parameters.
6.2.2.2 Post-processed WF
A Wiener filter was fitted via PLS regression on the training data set {X, Y}. Wold’s
R criterion was applied on the 6-fold cross-validated PRESS statistic to estimate
the optimal number of PLS factors.
Trajectory post-processing was used to integrate NC support into the decoder.
A logistic regression was trained to infer NC and IC states from the kinematic
estimates ŷt . Similarly to the MSLM, a Markovian hypothesis was used to limit
erroneous state detections. The corresponding transition matrix was found by
using the procedure presented in the case of the MSLM. The state probabilities
P (z t = i|ŷ1:t ), i = 1, . . . , 6 yielded by the dynamic LR state decoder were used
to weight the Wiener estimates associated with each finger and the neural values
corresponding to NC î states.
ï The dynamic logistic regression was identified on the
training data set Ŷ, z . Similarly to the preclinical data set, the parameters of
the pWF decoders were inferred from the MpWF parameters.
6.3. Conclusion 125
6.2.2.3 SKF
A SKF was implemented for dynamic combination of K = 6 Kalman filters, one
specialized in NC periods and the 5 other ones in IC periods. Similarly to the
preclinical data set, the SKF state variable ySKF
t (i.e., response variable) was chosen
as the finger position and velocity. The dimension of the neural features xSKF t
corresponding to the last 100ms before the instant t was reduced before application
of the SKF (see preliminary study in Appendix C). PLS regression between xSKF t
and the ySKF was used to identify an informative low-dimensional subspace. The
t
subspace’s dimension was chosen by 6-fold cross-validation on the training data

set. A constant term was integrated into the SKF’s neural features to account for
modifications of the baseline activity during NC and ICi , i = 1, . . . , 5 states. OLS
estimates of the transition Ak and emission Ck matrices were computed on the
training data sets {Xk,SKF , Yk,SKF }. ML estimates of the variance matrices Γk
and Σk were estimated following [Aggarwal et al., 2013].
6.3 Conclusion
This chapter includes a description of the procedure used to extract features from
both ECoG signals and limb trajectories. Details on the implementation of the
Markov Switching Linear Model (MSLM) have been exposed for the two evaluation
data sets, namely the publicly available preclinical and clinical data sets. Precisions
on the unsupervised procedure used to identify several IC experts have been given for
the considered expert and gate structures. The procedure completed to implement
Markovian post-processed Wiener Filter (MpWF) and Switching Kalman Filter
(SKF) decoders on the preclinical and clinical data set has been detailed. An insight
on the approaches used to optimize the MSLM’s gate and experts as well as to perform
efficient dimensionality reduction before applying the SKF has additionally been
presented. In particular, the respective relevance of generative and discriminative
strategies for the modelling of the MSLM’s gate emission distribution has been
investigated. Next chapter consists in the presentation of the decoding performance
of the MSLM for the considered pre-clinical and clinical data sets, along with of a
comparison of its decoding efficiency with the ones obtained with the alternative
decoders.
Chapter 7
Results
Contents
7.1 Preclinical data set . . . . . . . . . . . . . . . . . . . . . . . . 127
7.1.1 Integration of NC support into kinematic decoders . . . . . . . 127
movements . . . . . . . . . . . . . . . . . . . . . . . . . . . . 142
7.2 Clinical data set . . . . . . . . . . . . . . . . . . . . . . . . . . 145
7.3 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 159
7.3.1 Mono-limb decoding . . . . . . . . . . . . . . . . . . . . . . . 160
7.3.2 Multi-limb decoding . . . . . . . . . . . . . . . . . . . . . . . 163
7.3.3 Absolute decoding performance . . . . . . . . . . . . . . . . . . 164
7.4 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 165
The performance of the Markov Switching Linear Model (MSLM) was assessed for
two tasks, namely asynchronous mono-limb and asynchronous multi-limb decoding
of preclinical and clinical ECoG data, respectively.
7.1 Preclinical data set

A first study focused on the utilization of the MSLM, Wiener Filter with Markovian
post-processing (MpWF) and Switching Kalman Filter (SKF) decoders for the intro-
duction of NC support into kinematic decoders. The MSLM decoding performance
was compared to the MpWF and SKF respective performances on the test subsets
of the preclinical data set’s acquisition sessions.
A second study focused on complex movements. It was attempted to improve
kinematic estimates during IC states by combining several IC experts. As three
data sets only exhibited complex movements, this study was completed on a reduced
subdural data set.
7.1.1 Integration of NC support into kinematic decoders

The MSLM decoding performance was compared to SKF and MpWF respective
performances on the test subset of the preclinical data set’s 24 acquisition sessions
(2 and 2 monkeys, 8 and 16 sessions for the subdural and epidural data sets,
respectively). The state detection performances of variants of the MSLM and
128 Chapter 7. Results
MpWF with static state detection (referred to as SLM and pWF, respectively) were
additionally considered to assess the interest of the Markovian hypothesis to limit
spurious activations. Post-hoc pairwise decoder comparisons [Benavoli et al., 2015]
associated with the Friedman test with the significance level 0.05 [Demšar, 2006]
were performed using the sign test with Bonferroni correction (see section 5.3).
7.1.1.1 State decoding performance
Table 7.1 shows the respective performances of the MSLM, SLM, MpWF, pWF and
SKF decoders in terms of NC/IC classification accuracy. Both raw and synchronized
state estimates were considered (maximum delay between synchronized samples
τmax = 200 ms, ℓ1 -based synchronization). τmax = 200 ms was chosen because it
had been used in an earlier study completed on the epidural data set [Eliseyev and
Aksenova, 2014]. The corresponding p-values are gathered in Table 7.2 for the
decoders with dynamic state detection (MSLM, MpWF and SKF) and in Table
7.3 for the MSLM and the alternative decoders with static state detection (SLM
and pWF).
The AUCs corresponding to the MSLM were found to be significantly better
than the AUCs obtained with all alternative dynamic decoders (i.e., the MpWF
and the SKF) (see Table 7.2). The MSLM additionally permitted to significantly
improve the ACC when compared to the SKF on the subdural and epidural data
sets. The small number of sessions composing the subdural and epidural data sets
may have limited the power of the statistical test used to assess the significance
of performance differences. While the observed differences were not found to be
significant, the MSLM was additionally associated with generally better TPR, FPR
and ACC than the MpWF (median improvements of 1.2%, 26.6% and 1.8% for the
subdural data set, respectively, and of 3.1%, 14.4%, 2.5% for the epidural data set,
respectively). A significant improvement of the ACC by the MSLM when compared
to the MpWF was also observed on synchronized state sequences of the subdural
data set. When compared to the SKF, the MSLM improved the TPR, FPR and
ACC by 1.3%, 43.2% and 3.4% on the subdural data set (median value over 8
sessions), respectively, and by 19.8%, 29.8% and 17.0% on the epidural data set
(median value over 16 sessions), respectively.
No difference in the confusion-matrix-based performance indicators was observed
between the MSLM and the SLM, i.e. its variant with static state detection (see
Table 7.3).
Table 7.4 shows false activations/deactivations in terms of frequency and

duration. P-values are gathered in Table 7.5 and Table 7.6.
The decoders with dynamic state detection generally corresponded to false
activations and deactivations with similar (not-significantly different) characteristics,
except for the SKF which may have suffered from the lower state detection ability it
exhibited in Table 7.1. The frequency of the false activations and false deactivations
associated with the MpWF (SKF) was nevertheless higher that the one obtained
7.1. Preclinical data set 129
TPR FPR ACC AUC

MSLM 0.90 0.06 0.91 0.97
SLM 0.91 0.07 0.91 0.97
Subdural raw MpWF 0.88 0.10 0.89 0.96
pWF 0.87 0.10 0.89 0.95
(8 sessions) SKF 0.88 0.10 0.88 0.94
MSLM 0.92 0.03 0.94 0.98
SLM 0.93 0.06 0.94 0.98
synchronized MpWF 0.92 0.07 0.92 0.98
pWF 0.92 0.09 0.93 0.97
SKF 0.93 0.08 0.91 0.96
MSLM 0.77 0.14 0.79 0.89
SLM 0.76 0.15 0.79 0.89
Epidural raw MpWF 0.72 0.17 0.80 0.88
pWF 0.69 0.16 0.78 0.86
(16 sessions) SKF 0.67 0.21 0.66 0.76
MSLM 0.80 0.11 0.83 0.91
SLM 0.80 0.11 0.82 0.91
pWF 0.74 0.14 0.81 0.89
SKF 0.71 0.18 0.70 0.79
Table 7.1: Preclinical data set, state classification performance, raw and synchronized
state estimates (τmax = 200 ms, ℓ1 -based synchronization). Median False Positive
Rate (FPR), True Positive Rate (TPR), accuracy (ACC) and the Area Under the
Curve (AUC) are displayed for the MSLM, SLM, MpWF, pWF and SKF decoders.
They were computed over 8 and 16 acquisition sessions for the subdural and epidural
ECoG data sets, respectively.
TPR FPR ACC AUC

raw MSLM/MpWF 0.07 0.07 0.07 0.008*
Subdural MSLM/SKF 1 0.07 0.008* 0.008*
(8 sessions) synchronized MSLM/MpWF 0.73 0.07 0.008* 0.008*
MSLM/SKF 0.29 0.07 0.008* 0.008*
raw MSLM/MpWF 0.45 0.21 0.02 < 0.001*
Epidural MSLM/SKF 0.61 0.08 < 0.001* < 0.001*
(16 sessions) synchronized MSLM/MpWF 0.45 0.45 0.08 0.004*
MSLM/SKF 0.45 0.08 < 0.001* < 0.001*
Table 7.2: Preclinical data set, p-values for state classification performance, MSLM,
MpWF and SKF decoders, raw and synchronized state estimates (τmax = 200 ms,
ℓ1 -based synchronization). The significance of the differences between the decoders’
respective performances was assessed using the Friedman test with the significance
level α = 0.05. Post-hoc comparisons were performed where appropriate using the
sign test with Bonferroni correction, i.e. αBonf erroni = 0.0167. Significant differences
are indicated by a star (*).
TPR FPR ACC AUC

raw MSLM/SLM 0.45 0.29 0.07 0.29
Subdural MSLM/pWF 0.29 0.29 0.008* 0.008*
(8 sessions) synchronized MSLM/SLM 0.29 1 0.73 0.07
MSLM/pWF 0.73 0.07 0.008* 0.008*
raw MSLM/SLM 0.80 0.80 1 0.21
Epidural MSLM/pWF 0.80 0.45 0.45 0.004*
(16 sessions) synchronized MSLM/SLM 1 0.30 1 0.21
MSLM/pWF 0.80 0.45 0.21 0.004*
Table 7.3: Preclinical data set, p-values for state classification performance, MSLM,
SLM and pWF decoders, raw and synchronized state estimates (τmax = 200 ms,
ℓ1 -based synchronization). The significance of the differences between the decoders’
respective performances was assessed using the Friedman test with the significance
level α = 0.05. Post-hoc comparisons were performed where appropriate using the
sign test with Bonferroni correction, i.e. αBonf erroni = 0.0167. Significant differences
are indicated by a star (*).
with the MSLM by a median of 26.3% and 6.9% (62.3% and 32.4%), respectively,
on the subdural data set, and by a median of 12.1% and 5.3% (43.8% and 16.5%),
respectively, on the epidural data set. The improvement of the false activation
frequencies was found to be significant for both data sets when the MSLM and the
SKF were considered (Table 7.5).
The decoders with static gating were generally associated with more frequent
false activations and deactivations than the MSLM. In particular, the MSLM’s
dynamic gate resulted in significantly fewer false activations/deactivations than the
SLM on the epidural data set (median improvement of the false activations and
deactivations of 32.7% and 16.7%, respectively, on the subdural data set, and of
31.9% and 26.0%, respectively, on the epidural data set). It additionally significantly
surpassed the pWF in terms of false activation frequencies on both the subdural and
epidural data sets (median improvement of the false activations and deactivations of
50.0% and 21.8%, respectively, on the subdural data set, and of 40.0% and 26.6%,
respectively, on the epidural data set). The duration of the false activations and
deactivations were, on the other hand, higher for the MSLM than for the SLM
(median increase: 30.1% and 19.2% respectively, on the subdural data set, and 26.3%
and 28.9% respectively, on the epidural data set).
False activations False deactivations

Frequency Duration Frequency Duration
(1/s) (s) (1/s) (s)
MSLM 0.05 0.30 0.08 0.31
subdural SLM 0.10 0.26 0.09 0.24
MpWF 0.09 0.36 0.07 0.27
pWF 0.15 0.28 0.10 0.27
SKF 0.14 0.28 0.08 0.32
MSLM 0.14 0.35 0.19 0.39
epidural SLM 0.20 0.28 0.26 0.30
MpWF 0.14 0.37 0.22 0.40
pWF 0.20 0.29 0.27 0.32
SKF 0.23 0.35 0.25 0.36
Table 7.4: Preclinical data set, false activation/deactivation median frequency and
duration for MSLM, SLM, MpWF, pWF and SKF. False activations (respectively,
false deactivations) are blocks of consecutive NC samples misclassified as IC samples
(respectively, a block of IC samples mistaken for NC samples).

Subdural MSLM/MpWF 0.07 0.008* 0.73 0.29
(8 sessions) MSLM/SKF 0.07 0.29 0.29 0.29
Epidural MSLM/MpWF 0.45 0.08 0.61 0.45
(16 sessions) MSLM/SKF < 0.001* 0.21 0.45 0.45
Table 7.5: Preclinical data set, p-values for the false activation/deactivation frequency
and duration associated with the MSLM, MpWF and SKF decoders. The significance
of the differences between the decoders’ respective performances was assessed using
the Friedman test with the significance level α = 0.05. Post-hoc comparisons were
performed where appropriate using the sign test with Bonferroni correction, i.e.
αBonf erroni = 0.0167. Significant differences are indicated by a star (*).

Subdural MSLM/SLM 0.07 0.008* 0.07 0.008*
(8 sessions) MSLM/pWF 0.008* 0.07 0.22 0.29
Epidural MSLM/SLM < 0.001* < 0.001* < 0.001* < 0.001*
(16 sessions) MSLM/pWF < 0.001* < 0.001* 0.021 < 0.001*
Table 7.6: Preclinical data set, p-values for the false activation/deactivation frequency
and duration associated with the MSLM, SLM and pWF decoders. The significance
7.1.1.2 IC decoding performance, known states

The relevance of the continuous IC kinematic models embedded in the hybrid
decoders was then assessed. As state labels are traditionally hidden during decoder
application, the performance indicators measured over IC states reflect both the
quality of the IC continuous decoder and of the state decoder (a false negative
for example results in a high ℓ2 -error). To decouple the impact of the state and
continuous decoders, the MSLM, MpWF and SKF were applied on complete test data
sets {X, z}, i.e. the state sequence was not hidden during application. Continuous
performance indicators were then computed using exclusively the true IC samples.
Table 7.7 gathers the IC continuous models’ decoding performance.
While no significant differences were observed between the MSLM’s and MpWF’s
IC decoding ability on the subdural data set, the MSLM corresponded to significantly
better NRMSE and NMAE than the MpWF on the epidural data set. A preliminary
synchronization of the considered trajectories additionally disclosed a significantly
higher PCC for the MSLM than for the MpWF on the epidural data set.
The MSLM and the SKF generally exhibited similar IC decoding abilities on
both the subdural and epidural data sets. When raw (un-synchronized) subdural
trajectories were considered, however, the MSLM significantly outperformed the
SKF in terms of NRMSE and NMAE on the subdural data set.
An illustration of the decoded trajectories is shown in Figure 7.1.
PCC NRMSE NMAE

MSLM 0.60 0.77 0.77
raw MpWF 0.61 0.79 0.74
subdural SKF 0.57 0.82 0.79
MSLM 0.66 0.71 0.67
synchronized MpWF 0.67 0.74 0.66
SKF 0.64 0.77 0.72
MSLM 0.27 0.99 0.96
raw MpWF 0.24 1.03 0.99
epidural SKF 0.33 0.98 0.94
MSLM 0.41 0.92 0.87
SKF 0.42 0.92 0.85
Table 7.7: Preclinical data set, known state sequence. Decoding performance during
IC states, with and without synchronization. The PCC, RMSE and NMAE associated
with the MSLM, MpWF and SKF decoders were averaged over the 3 available axes
within each session. Their median values were then computed over 8 and 16 sessions
for the subdural and epidural data sets, respectively. Synchronization was completed
with τmax = 200ms (ℓ2 -based synchronization for the PCC and NRMSE, ℓ1 -based
synchronization for the NMAE).
MSLM MpWF SKF

350
y 1 (mm)
300
250
200
0
y 2 (mm)
-100
-200
-300
200
y 3 (mm)
100
0 20 40 0 20 40 0 20 40
time (s) time (s) time (s)
observed estimated
Figure 7.1: Preclinical data set, known state sequence. Example of observed and
estimated trajectories (epidural data set, Monkey C). The projections of the monkey’s
wrist trajectory onto the horizontal axes (y1 and y2 ) and the vertical axis (y3 ) are
indicated in solid black lines. Red trajectories represent the estimates yielded by
the MSLM, SKF and MpWF decoders.
PCC NMRSE NMAE

raw MSLM/MpWF 0.29 1 0.73
subdural MSLM/SKF 0.07 0.07 0.29
synchronized MSLM/MpWF 0.29 0.29 1
MSLM/SKF 0.29 0.008* 0.008*
raw MSLM/MpWF 1 < 0.001* < 0.001*
epidural MSLM/SKF 0.08 0.80 0.80
synchronized MSLM/MpWF < 0.001* < 0.001* < 0.001*
MSLM/SKF 0.80 1 0.08
Table 7.8: Preclinical data set, known state sequence. P-values for IC decoding
performance, raw and synchronized signals (τmax = 200ms, ℓ2 -based synchronization
for the PCC and NRMSE, ℓ1 -based synchronization for the NMAE). The significance
performed using the sign test with Bonferroni correction, i.e. αBonf erroni = 0.0167.
Significant differences are indicated by a star (*).
7.1.1.3 IC decoding performance, hidden states
The performance of the three decoders was then assessed over IC periods when
state labels were hidden in the test data set. Results are gathered in Table 7.9,
and corresponding p-values are displayed in Table 7.10. An example of decoded
trajectories is presented in Figure 7.2.
The MSLM, MpWF and SKF generally presented similar decoding performances
on the subdural data set. The MSLM nevertheless corresponded to PCC, NRMSE
and NMAE higher by a median of 10%, 7% and 10% than the SKF’s PCC, NMRSE
and NMAE, respectively. This trend was confirmed by the significant difference
between the MSLM’s and SKF’s NRMSE and NMAE when temporal errors were
partly discarded, i.e. when synchronized trajectories were compared.
On the epidural data set, however, the MSLM significantly surpassed the MpWF
in terms of NRMSE and NMAE. Synchronized performance indicators addition-
ally permitted to disclose a significant improvement of the PCC when the MSLM
rather than the MpWF was applied on the epidural data set. While the MSLM
corresponded to NRMSE and NMAE higher by a median of 4% and 5% than the
SKF, these differences were not found to be significant.
An additional comparison was completed to assess the interest of training the

MSLM on IC samples exclusively in the case of hidden state sequences. An alternative
MSLM was built by replacing the IC experts trained on IC samples by models trained
on both NC and IC samples, i.e. by the linear models applied by the MpWF. The
ability of each switching model to correctly deal with IC periods was compared (see
Table 7.11 for the median results, and Table 7.12 for the corresponding p-values).
PCC NRMSE NMAE

MSLM 0.60 0.78 0.76
raw MpWF 0.61 0.80 0.72
subdural SKF 0.56 0.86 0.81
MSLM 0.66 0.72 0.67
SKF 0.61 0.81 0.74
MSLM 0.23 1.04 1.00
raw MpWF 0.23 1.07 1.03
epidural SKF 0.21 1.04 1.02
MSLM 0.32 0.98 0.93
SKF 0.28 1.00 0.97
Table 7.9: Preclinical data set, hidden state sequence. Decoding performance
during IC states, with and without synchronization. The PCC, RMSE and NMAE
associated with the MSLM, MpWF and SKF decoders were averaged over the 3
available axes for each session. Their median values were then computed over 8 and
16 sessions for the subdural and epidural data sets, respectively. Synchronization was
completed with τmax = 200ms (ℓ2 -based synchronization for the PCC and NRMSE,
ℓ1 -based synchronization for the NMAE).
PCC NMRSE NMAE

raw MSLM/MpWF 0.29 0.73 0.73
Subdural MSLM/SKF 0.07 0.07 0.07
synchronized MSLM/MpWF 1 0.29 0.73
MSLM/SKF 0.07 0.008* 0.008*
raw MSLM/MpWF 0.08 0.004* < 0.001*
Epidural MSLM/SKF 0.21 0.45 0.02
synchronized MSLM/MpWF < 0.001* < 0.001* < 0.001*
MSLM/SKF 0.21 0.45 0.08
Table 7.10: Preclinical data set, hidden state sequence. P-values for IC decoding
performance, raw and synchronized trajectories (τmax = 100ms, ℓ2 -based synchro-
nization for the PCC and NRMSE, ℓ1 -based synchronization for the NMAE). The
significance of the differences between the decoders’ respective performances was
assessed using the Friedman test with the significance level α = 0.05. Post-hoc
comparisons were performed using the sign test with Bonferroni correction, i.e.
7.1. Preclinical data set
Figure 7.2: Preclinical data set, hidden state sequence. Example of observed and estimated trajectories for the epidural data set
(Monkey C). The projections of the monkey’s wrist trajectory onto the horizontal axes (y1 and y2 ) and the vertical axis (y3 ) are
indicated in solid black lines. Red trajectories represent the estimates yielded by the MSLM, SLM, Wiener Filter (WF), MpWF and
SKF decoders.
137
In contrast with the indicators computed on IC samples when state sequences were
known (see Table 7.7 above), differences between IC experts and single models of
the MpWF trained on both NC and IC samples were thus monitored on IC samples
correctly classified by the MSLM gate. As a result, the single kinematic decoder
associated with the MpWF was not penalized if its TPR was lower than the MSLM’s
TPR.
Consistently with the previously reported results, no difference was observed
on the subdural data set between the MSLM’s IC experts exclusively fitted on IC
samples and the models trained on both NC and IC samples (i.e., the single models
utilized by the MpWFs). On the epidural data set, however, the NRMSE and NMAE
associated with models trained on both NC and IC samples were significantly higher
than the ones corresponding to the MSLM’ IC experts, i.e. on models exclusively
trained on IC samples. A non significant median improvement of 6% was additionally
found for the PCC, which was significantly improved when synchronized trajectories
were considered.
MSLM PCC NRMSE NMAE

IC expert training 0.60 0.78 0.76
subdural raw NC/IC expert training 0.61 0.80 0.73
synchronized NC/IC expert training 0.66 0.75 0.66
epidural raw NC/IC expert training 0.22 1.07 1.03
synchronized NC/IC expert training 0.30 1.02 0.97
Table 7.11: Preclinical data set, hidden state sequence. Decoding performance of
MSLMs over IC states with either a IC- or a NC/IC- trained IC expert, hidden
state sequence, with and without synchronization. The PCC, RMSE and NMAE
associated with both variants of the MSLM were averaged over the 3 available axes
for each session. Their median values were then computed over 8 and 16 sessions for
the subdural and epidural data sets, respectively. Synchronization was completed
with τmax = 200ms (ℓ2 -based synchronization for the PCC and NRMSE, ℓ1 -based
synchronization for the NMAE).
MSLM IC expert training MSLM PCC NMRSE NMAE

Subdural raw IC / NC-IC 0.07 0.73 0.73
synchronized IC / NC-IC 0.73 0.07 0.29
Epidural raw IC / NC-IC 0.08 < 0.001* < 0.001*
synchronized IC / NC-IC < 0.001* < 0.001* < 0.001*
Table 7.12: Preclinical data set, hidden state sequence. P-values for the decoding
performance of MSLMs over IC states with either a IC- or a NC/IC- trained IC
expert, hidden state sequence, raw and synchronized trajectories (τmax = 200ms,
ℓ2 -based synchronization for the PCC and NRMSE, ℓ1 -based synchronization for
the NMAE). The significance of the differences between the decoders’ respective
performances was assessed using the sign test with the significance level α = 0.05.
Significant differences are indicated by a star (*).
7.1.1.4 Modality influence

An example of the influence of frequency, temporal and spatial features of the
MSLM’s IC expert and gate is shown in Figure 7.3 (subdural data set, Monkey A,
session 1). Contributions were assessed as the normalized summation of absolute
values of models’ coefficients along each modality. The WF corresponding modality
influence is presented in Figure 7.4 (subdural data set, Monkey A, session 1).
y1 y2 y3 w gate
Spectral influence
(a.u)
0 100 200 0 100 200 0 100 200 0 100 200

Frequency (Hz) Frequency (Hz) Frequency (Hz) Frequency (Hz)
Temporal influence
(a.u)
-1 -0.5 -1 -0.5 -1 -0.5 -1 -0.5

Time (s) Time (s) Time (s) Time (s)
Spatial influence
0.06
0.06 0.04
(a.u)
0.04 0.04
0.04
0.02 0.02 0.02
0.02
Figure 7.3: Preclinical data set, example of the average contribution of frequency,
temporal and spatial modalities to the MSLM’s IC expert and to the MSLM’s
gate (subdural data set, Monkey A, session 1). The vector wgate parametrizes the
1
PLS-based followed by the logistic regression such that P (z t = 1|xt ) = −w′ xt
.
1+e gate
Contributions were assessed as the normalized summation of absolute values of
models’ coefficients along each modality. "a.u." refers to the utilization of arbitrary
units.
y1 y2 y3
Spectral influence
(a.u)
50 100 150 200 250 50 100 150 200 250 50 100 150 200 250
Frequency (Hz) Frequency (Hz) Frequency (Hz)
Temporal influence
(a.u)
-0.8 -0.6 -0.4 -0.2 -0.8 -0.6 -0.4 -0.2 -0.8 -0.6 -0.4 -0.2
Time (s) Time (s) Time (s)
Spatial influence
0.04 0.06 0.06

(a.u)
0.04
0.04
0.02 0.02
0.02
Figure 7.4: Preclinical data set, example of the average contribution of frequency,
temporal and spatial modalities to the MpWF continuous decoder (subdural data
set, Monkey A, session 1). Contributions were assessed as the normalized summation
of absolute values of models’ coefficients along each modality. "a.u." refers to the
utilization of arbitrary units.

movements
3 sessions of the subdural data set exhibit complex reaching movements (see Chapter
4). MSLM with k = 2 IC experts were trained to assess whether complex movements
are better modelled using several IC experts. Because of this limited number of
sessions, 10-fold cross-validation was performed on each of the 3 data sets. The sign
test was applied on the 10 values obtained for each data set, and permitted to assess
whether MSLMs with 1 or 2 IC experts performed similarly on each data set.

Considering several IC states may result in a degradation of the state detection. A
first analysis thus focused on state detection. 10-fold cross-validated state-related
performance indicators are gathered in Table 7.13.
NC/IC state labels, where IC states gather IC1 and IC2 states, were not con-
strained during training because of labelling uncertainties, for example around
transitions or for very short null-velocity periods during movements (see Chapter
6). Using several IC states resulted in a dramatic diminution of the FPR for the
3 considered sessions. By contrast, the TPR (and, consequently, the ACC) was
diminished when compared with the supervised binary MSLM. As continuous perfor-
mance indicators computed over IC samples were improved when a 3-state MSLM
was used (see next paragraph), the TPR nevertheless seems irrelevant to assess
decoding accuracy during IC periods. These results suggest that the MSLM may
have advantageously applied the NC rather than IC expert on samples manually
labelled as IC, for example at the beginning of IC periods if the IC expert was not
able to properly model movement initiations, or for samples corresponding to short
null-velocity periods within movements.
7.1.2.2 IC decoding performance

The IC decoding ability of single and multiple IC experts was then assessed. 10-
fold cross-validated performance indicators between raw and synchronized observed
and estimated trajectories are gathered in Table 7.14. Using multiple experts
generally permitted to improve the modelling of wrist kinematic during IC states. A
significant improvement of the MAE was observed for the three data sets. The PCC
was additionally significantly higher for two out of three data sets, and the NRMSE
for one data set. An illustration of the trajectory estimates yielded by single and
multiple (k = 2) IC experts is shown in Figure 7.5.
Session decoder TPR FPR ACC

1 MSLM, 1 IC state 0.90 0.09 0.10
MSLM, 2 IC states 0.83 0.05 0.12
raw 4 MSLM, 1 IC state 0.88 0.13 0.88
MSLM, 2 IC states 0.76 0.01 0.82
5 MSLM, 1 IC state 0.85 0.08 0.89
MSLM, 2 IC states 0.67 0.03 0.88
1 MSLM, 1 IC state 0.92 0.05 0.07
MSLM, 2 IC states 0.87 0.04 0.10
synchronized 4 MSLM, 1 IC state 0.91 0.08 0.91
MSLM, 2 IC states 0.79 0.01 0.85
5 MSLM, 1 IC state 0.86 0.07 0.91
MSLM, 2 IC states 0.70 0.02 0.90
Table 7.13: Preclinical data set, sessions exhibiting complex movements. Cross-
validated state classification performance, raw and synchronized state estimates
(τmax = 200 ms, ℓ1 -based synchronization). 10-fold cross-validated median False
Positive Rate (FPR), True Positive Rate (TPR) and the accuracy (ACC) are
displayed for the MSLM with one and two IC states.
session decoder PCC RMSE MAE

1 1 IC state 0.58 0.83 0.84
2 IC states 0.62 0.80 0.76
raw 4 1 IC state 0.56 0.85 0.84
2 IC states 0.60 0.82 0.78
5 1 IC state 0.69 0.75 0.70
2 IC states 0.70 0.73 0.70
1 1 IC state 0.64 0.79 0.78
2 IC states 0.70 0.72 0.66
synchronized 4 1 IC state 0.59 0.80 0.77
2 IC states 0.65 0.75 0.70
5 1 IC state 0.74 0.70 0.64
2 IC states 0.76 0.67 0.59
Table 7.14: Preclinical data set, hidden state sequence, sessions exhibiting complex
movements. The IC cross-validated decoding performance of MSLMs with 1 or 2 IC
states was measured on the three subdural data sets composed of complex movements.
The PCC, NRMSE and NMAE were averaged over the 3 considered axes. Their
10-fold cross-validated median value is displayed for MSLMs with 1 or 2 IC states.
Performance indicators were computed for both raw and synchronized observed and
estimated trajectories. Synchronization was completed with τmax = 200ms (ℓ2 -based
synchronization for the PCC and NRMSE, ℓ1 -based synchronization for the NMAE).
session MSLM decoder PCC NMRSE NMAE

raw 1 1 IC state / 2 IC states 0.02* 0.1 0.02*
4 1 IC state / 2 IC states 0.02* 0.02* 0.002*
5 1 IC state / 2 IC states 0.3 0.8 0.02*
synchronized 1 1 IC state / 2 IC states 0.02* 0.02* 0.02*
4 1 IC state / 2 IC states 0.02* 0.02* 0.002*
5 1 IC state / 2 IC states 0.1 0.8 0.02*
Table 7.15: Preclinical data set, sessions exhibiting complex movements, hidden
state sequence. P-value for IC cross-validated decoding performance of MSLMs with
1 or 2 IC states(τmax = 100ms, ℓ2 -based synchronization for the PCC and NRMSE,
ℓ1 -based synchronization for the NMAE). The significance of the differences between
the decoders’ respective performances was assessed using the sign test with the
significance level α = 0.05. Significant differences are indicated by a star (*).
MSLM, 1 IC state MSLM, 2 IC states

600
y 1 (mm)
400
200
400
y 2 (mm)
200
0
400
y 3 (mm)
200
observed estimated
Detected states Detected states

400
y 3 (mm)
200
0
0 10 20 30 40 0 10 20 30 40
time (s) time (s)
Figure 7.5: Preclinical data set, sessions exhibiting complex movements, example of
observed and estimated trajectories (Monkey A). The projections of the monkey’s
wrist trajectory onto the horizontal axes (y1 and y2 ) and the vertical axis (y3 ) are
indicated in solid black lines. Red trajectories represent the estimates yielded by
MSLM decoders with 1 and 2 IC experts, respectively. Detected NC states are
represented with green dots, IC states with blue and yellow dots.
7.2. Clinical data set 145
7.2 Clinical data set

The clinical data set was used to assess the respective performances of the MSLM,
MpWF and SKF decoders for asynchronous multi-limb decoding. The MSLM
decoding efficiency was compared to the SKF and MpWF respective efficiencies on the
test data set of the 9 acquisition sessions of the clinical data set (9 subjects, one session
per subject). The state detection performance of a SLM and a pWF was additionally
investigated to evaluate the interest of the dynamic state detection performed by
the MSLM. A maximal lag τmax = 100 ms was chosen for synchronization because
the frequency content of the clinical (finger) trajectories was higher than the one
associated with preclinical (wrist) trajectories. Using τmax = 200 ms may thus have
artificially improved the quality of the kinematic estimates because of a too flexible
synchronization.

Table 7.16 shows the respective performances of the decoders in terms of confusion-
matrix-based indicators. The corresponding p-values are gathered in Table 7.17
for the decoders with dynamic state detection (MSLM, MpWF and SKF) and in
Table 7.18 for the MSLM and the alternative decoders with static state detection
(SLM and pWF).
The ACCs corresponding to the MSLM were found to be significantly better than
the ACCs obtained with the alternative dynamic decoders. MSLMs’ ACCs were
higher by a median of 5% and 49% than the MpWFs’ and SKFs’ ACC, respectively.
While this improvement was not found to be significant, the FPR associated with
the MSLM was higher than the one corresponding to the MpWF by a median of
33%. Although the MSLM’s CTPR was significantly better than the one associated
with the pWF, it wasn’t the case for the SLM and for the MpWF. The Markovian
hypothesis used by the MpWF may have permitted to compensate for a possible
intrinsic lower decoding ability of the post-processed operator which is suggested by
the low CTPR of the pWF.
While the SLM’s CTPR was 7% lower than the MSLM’s one (for a similar TPR),
no significant differences were observed between the MSLM and SLM in terms of
ACC and CTPR.
Table 7.19 shows false activations/deactivations in terms of occurrence num-
ber and duration. The corresponding p-values are gathered in Tables 7.20 and
7.21. The MSLM significantly outperformed the MpWF and SKF in terms of
false activation and/or false deactivation frequencies. While the difference was not
found significant, the SLM corresponded to false deactivations which were 60% more
frequent than the MSLM. This observation seems to support the idea that dynamic
state detection may be advantageously used to limit short spurious activations or
deactivations of a BCI system.
TPR FPR ACC CTPR

MSLM 0.91 0.16 0.87 0.70
SLM 0.87 0.10 0.86 0.65
raw MpWF 0.91 0.34 0.82 0.67
pWF 0.85 0.28 0.80 0.61
SKF 0.64 0.20 0.57 0.20
MSLM 0.94 0.12 0.90 0.70
SLM 0.90 0.07 0.90 0.66
pWF 0.97 0.22 0.84 0.62
SKF 0.66 0.19 0.60 0.22
Table 7.16: Clinical data set, state classification performance, raw and synchronized
signals. Median False Positive Rate (FPR), True Positive Rate (TPR), accuracy
(ACC) and Correct TP Ratio (CTPR) are displayed for the MSLM, SLM, MpWF,
pWF and SKF decoders on the clinical data set (median value over 9 sessions).
Synchronized performance indicators are additionally exposed. True and estimated
switching states y and ŷ were synchronized (see section 5.2.4) before computing
the TPR, FPR, ACC and CTPR (τmax = 100ms, median value over 9 sessions).
TPR FPR ACC CTPR

raw MSLM/MpWF 0.18 0.004* 0.004* 1
MSLM/SKF 0.004* 0.51 0.004* 0.004*
synchronized MSLM/MpWF 0.008* 0.004* 0.004* 1
MSLM/SKF 0.004* 0.51 0.004* 0.004*
Table 7.17: Clinical data set, p-values for state classification performance, MSLM,
MpWF and SKF decoders, raw and synchronized signals (τmax = 100ms). The
TPR FPR ACC CTPR

raw MSLM/SLM 0.18 0.004* 0.51 0.04
MSLM/pWF 0.18 0.04 0.004* 0.004*
synchronized MSLM/SLM 0.18 0.004* 0.51 0.04
MSLM/pWF 0.18 0.04 0.004* 0.04
Table 7.18: Clinical data set, p-values for state classification performance, MSLM,
SLM and pWF decoders, raw and synchronized signals (τmax = 100ms). The

Frequency (1/s) Duration (s) Frequency (1/s) Duration (s)
MSLM 0.07 0.26 0.07 0.20
SLM 0.06 0.17 0.18 0.20
MpWF 0.16 0.24 0.10 0.20
pWF 0.19 0.16 0.33 0.16
SKF 0.08 0.24 0.40 0.34
Table 7.19: Clinical data set, false activation/deactivation median frequency and
duration for MSLM, MpWF and SKF decoders. False activations (respectively, false
deactivations) are blocks of consecutive NC samples misclassified as IC samples
(respectively, a block of IC samples mistaken for NC samples.

MSLM/MpWF 0.004* 1 0.004* 0.04
MSLM/SKF 0.28 1 0.004* 0.004*
Table 7.20: Clinical data set, p-values for the false activation/deactivation frequency
and duration associated with the MSLM, MpWF and SKF decoders. The significance

MSLM/SLM 0.69 0.008* 0.18 0.51
MSLM/pWF 0.004* 0.18 0.18 0.04
Table 7.21: Clinical data set, p-values for the false activation/deactivation frequency
and duration associated with the MSLM, SLM and pWF decoders. The significance
7.2.0.2 IC decoding performance, known states

The relevance of the continuous IC kinematic models embedded in the hybrid
decoders was first assessed in the case where switching state values were known in
the test data sets. Similarly to the study led on the preclinical data set, the MSLM,
MpWF and SKF were applied on complete test data sets {X, z}, i.e. the state
sequence was not hidden during application. Continuous performance indicators
were then computed over the true IC samples exclusively. Table 7.22 shows the IC
continuous models decoding performance, computed for both desynchronized and
synchronized trajectories (τmax = 100ms, ℓ2 -based synchronization for the PCC and
NRMSE, ℓ1 -based synchronization for the NMAE). P-values are gathered in Table
7.23.
The MSLM IC experts significantly outperformed the MpWF continuous models
in terms of NRMSE and NMAE (median improvements of 26% and 22%, respectively).
While the PCC was not found to be significantly different because of the presence of
an outlier in the small data set, it was improved by a median value of 13%. When
synchronized performances indicators were considered, the PCC associated with
the MSLM’s IC experts became significantly better than the one obtained with the
MpWF. Similar results were obtained when the SKF IC experts were compared to
the MSLM’s ones, i.e. the MSLM IC experts surpassed both the MpWF continuous
models and the SKF’s experts. Examples of decoded trajectories are shown in
Figures 7.6 and 7.7.
PCC NRMSE NMAE

MSLM 0.50 0.97 0.86
raw MpWF 0.40 1.24 1.09
SKF 0.31 1.39 1.27
MSLM 0.68 0.80 0.65
SKF 0.43 1.27 1.16
Table 7.22: Clinical data set, known state sequence. Decoding performance during
IC states, with and without synchronization. Synchronization was completed with
τmax = 100ms (ℓ2 -based synchronization for the PCC and NRMSE, ℓ1 -based syn-
chronization for the NMAE). Within each acquisition session, PCC, NRMSE and
NMAE were averaged over the 5 fingers. Median PCC, NRMSE and NMAE are
displayed for the MSLM, MpWF and SKF decoders on the clinical data set (median
value over 9 sessions).
PCC NMRSE NMAE

raw MSLM/MpWF 0.04 0.004* 0.004*
MSLM/SKF 0.04 0.004* 0.004*
synchronized MSLM/MpWF 0.004* 0.004* 0.004*
MSLM/SKF 0.004* 0.004* 0.004*
Table 7.23: Clinical data set, known state sequence. P-values for IC decoding per-
formance for raw and synchronized signals (τmax = 100ms, ℓ2 -based synchronization
for the PCC and NRMSE, ℓ1 -based synchronization for the NMAE). Within each
acquisition session, PCC, NRMSE and NMAE were averaged over the 5 fingers.
The significance of the differences between the decoders’ respective performances
was assessed using the Friedman test with the significance level α = 0.05. Post-hoc
MSLM MpWF SKF

6
y 1 (a.u.)
4
2
0
-2
6
y 2 (a.u.)
4
2
0
-2
6
y 3 (a.u.)
4
2
0
-2
6
y 4 (a.u.)
4
2
0
-2
6
y 5 (a.u.)
4
2
0
-2
35 40 45 50 55 35 40 45 50 55 35 40 45 50 55
observed estimated
Figure 7.6: Clinical data set, known state sequence. Example of observed and
estimated finger trajectories (subject zt). The tracked finger positions are indicated
in solid black lines. Red trajectories represent the estimates yielded by the MSLM,
MpWF and SKF decoders. "a.u." refers to "arbitrary unit".
MSLM MpWF SKF

6
y 1 (a.u.)
4
2
0
-2
-4
2 4 6 8 2 4 6 8 2 4 6 8
y 2 (a.u.)
4
2
0
32 34 36 32 34 36 32 34 36
y 3 (a.u.)
4
2
0
-2
58 60 62 58 60 62 58 60 62
y 4 (a.u.)
4
2
0
28 30 32 28 30 32 28 30 32
y 5 (a.u.)
4
2
0
-2
55 60 55 60 55 60
observed estimated
Figure 7.7: Clinical data set, known state sequence, focus on the IC states. Example
of observed and estimated finger trajectories (subject zt). The tracked finger positions
are indicated in solid black lines. Red trajectories represent the estimates yielded
by the MSLM, MpWF and SKF decoders. "a.u." refers to "arbitrary unit".
7.2.0.3 IC decoding performance, hidden states
IC decoding accuracy was then measured in the case where the values taken by the
switching state were hidden in the test data set. Results are shown in Table 7.24.
The corresponding p-values are displayed in Table 7.25.
The MSLM significantly outperformed the WF in terms of NRMSE and NMAE
(minus 9% and 12%, respectively). The MSLM was additionally associated with a
PCC higher by a median of 7%. The SKF was significantly surpassed by the MSLM
for all performance indicators for both raw (un-synchronized) and synchronized
signals. An example of decoded trajectories is presented on Figure 7.8.
PCC NRMSE NMAE

MSLM 0.29 1.21 1.04
raw MpWF 0.27 1.34 1.17
SKF 0.12 1.62 1.46
MSLM 0.42 1.11 0.90
SKF 0.12 1.60 1.44
Table 7.24: Clinical data set, hidden state sequence. Decoding performance during
IC states, with and without synchronization. Synchronization was completed with
τmax = 100ms (ℓ2 -based synchronization for the PCC and NRMSE, ℓ1 -based syn-
chronization for the NMAE). Within each acquisition session, PCC, NRMSE and
NMAE were averaged over the 5 fingers. Median PCC, NRMSE and NMAE are
displayed for the MSLM, MpWF and SKF decoders on the clinical data set (median
value over 9 sessions).
PCC NMRSE NMAE

raw MSLM/MpWF 0.51 0.004* 0.004*
MSLM/SKF 0.004* 0.004* 0.004*
synchronized MSLM/MpWF 0.04 0.004* 0.004*
MSLM/SKF 0.004* 0.004* 0.004*
Table 7.25: Clinical data set, hidden state sequence. P-values for IC decoding per-
formance for raw and synchronized signals (τmax = 100ms, ℓ2 -based synchronization
for the PCC and NRMSE, ℓ1 -based synchronization for the NMAE). Within each
acquisition session, PCC, NRMSE and NMAE were averaged over the 5 fingers.
The significance of the differences between the decoders’ respective performances
was assessed using the Friedman test with the significance level α = 0.05. Post-hoc
Similarly to the preclinical data set, an additional comparison was completed to

decouple the impact of the state and kinematic decoders on the decoding accuracy
over hidden IC states. An alternative MSLM was built by replacing IC experts
trained on ICi samples by models trained on both NC and IC samples, i.e. by
the linear models applied by the MpWF. The ability of each switching model to
correctly deal with IC periods was compared (see Table 7.26 for the median results,
and Table 7.27 for the corresponding p-values). In contrast with the indicators
computed on true, known IC samples, the decoding performance is here compared
on ICi samples correctly classified as IC by the MSLM gate. As a result, the model
fitted on both NC and IC (i.e., embedded in the MpWF) is not penalized by IC
observations erroneously classified as NC by both the MSLM and MpWF when it is
compared to the MSLM’s IC experts.
Using MSLM experts trained on IC samples exclusively permitted to significantly
improve all performance indicators (median improvements of 15%, 11% and 12%
for the PCC, NRMSE and NMAE, respectively). Significant median improvements
of 32%, 16% and 21% were obtained when synchronized signals were compared
(τmax = 100ms, i.e. 1 sample), suggesting that the MSLM exhibits slight temporal
uncertainties.
PCC NRMSE NMAE

MSLM-IC 0.29 1.21 1.04
raw MSLM-full 0.26 1.35 1.18
MSLM-IC 0.42 1.11 0.90
synchronized MSLM-full 0.29 1.32 1.13
Table 7.26: Clinical data set, hidden state sequence. MSLM IC decoding performance
with experts fitted on IC samples only ("MSLM-IC") or on both NC and IC samples
("MSLM-full"). Within each acquisition session, PCC, NRMSE and NMAE were
averaged over the 5 fingers. Median PCC, NRMSE and NMAE are displayed for
the MSLM-IC and MSLM-full decoders on the clinical data set (median value over
9 sessions).
PCC NMRSE NMAE

raw MSLM-IC/MSLM-full 0.04* 0.004* 0.004*
synchronized MSLM-IC/MSLM-full 0.004* 0.004* 0.004*
Table 7.27: Clinical data set, p-values for IC decoding performance for raw and syn-
chronized signals. Within each acquisition session, PCC, NRMSE and NMAE were
averaged over the 5 fingers. The significance of the differences between the decoders’
respective performances was assessed using the sign test with the significance level
α = 0.05. Significant differences are indicated by a star (*).
MSLM MpWF SKF

8
y 1 (a.u.)
6
4
2
0
-2
8
y 2 (a.u.)
6
4
2
0
-2
8
y 3 (a.u.)
6
4
2
0
-2
8
y 4 (a.u.)
6
4
2
0
-2
8
y 5 (a.u.)
6
4
2
0
-2
35 40 45 50 35 40 45 50 35 40 45 50
observed estimated
Figure 7.8: Clinical data set, hidden state sequence. Example of observed and
estimated finger trajectories (subject zt). The tracked finger positions are indicated
in solid black lines. Red trajectories represent the estimates yielded by the MSLM,
MpWF and SKF decoders. "a.u." refers to "arbitrary unit".
7.2.0.4 Modality influence

Figures 7.10 and 7.9 illustrate the influence of frequency, temporal and spatial
features of the IC experts and the gating network, respectively (subject "zt", 1
session). Figure 7.11 corresponds to the Wiener filter embedded in the MpWF.
Contributions were assessed as the normalized summation of absolute values of
models’ coefficients along each modality.
Chapter 7. Results
w1 w2 w3 w4 w5 w6
Spectral influence
(a.u)
0 100 200 0 100 200 0 100 200 0 100 200 0 100 200 0 100 200
Frequency (Hz) Frequency (Hz) Frequency (Hz) Frequency (Hz) Frequency (Hz) Frequency (Hz)
Temporal influence
(a.u)
-0.8 -0.4 -0.8 -0.4 -0.8 -0.4 -0.8 -0.4 -0.8 -0.4 -0.8 -0.4
Time (s) Time (s) Time (s) Time (s) Time (s) Time (s)
Spatial influence
0.06
(a.u)
0.04
0.02
electrodes electrodes electrodes electrodes electrodes electrodes
(1cm inter-electrode (1cm inter-electrode (1cm inter-electrode (1cm inter-electrode (1cm inter-electrode (1cm inter-electrode
distance) distance) distance) distance) distance) distance)
Figure 7.9: Clinical data set, example of the average contribution of frequency, temporal and spatial modalities to the MSLM’s
gate (subject zt, one session). The vector set {w1 , . . . , w6 } parametrizes the 6-class multinomial logistic regression such that
w x ′ t
P (z t = i|xt ) = qe iw′ xt . Contributions were assessed as the normalized summation of absolute values of each vector wi along each
j e
156
j
modality. "a.u." refers to the utilization of arbitrary units.
Finger 1 Finger 2 Finger 3 Finger 4 Finger 5
(a.u)
Spectral influence
0 100 200 0 100 200 0 100 200 0 100 200 0 100 200
Frequency (Hz) Frequency (Hz) Frequency (Hz) Frequency (Hz) Frequency (Hz)
7.2. Clinical data set
(a.u)
Temporal influence
-1 -0.5 -1 -0.5 -1 -0.5 -1 -0.5 -1 -0.5
Time (s) Time (s) Time (s) Time (s) Time (s)
0.1
(a.u.)
Spatial influence
0
Electrodes Electrodes Electrodes Electrodes Electrodes
(1cm inter-electrode (1cm inter-electrode (1cm inter-electrode (1cm inter-electrode (1cm inter-electrode
distance) distance) distance) distance) distance)
Figure 7.10: Clinical data set, example of the average contribution of frequency, temporal and spatial modalities to the MSLM’s IC
experts (subject zt, one session). Contributions were assessed as the normalized summation of absolute values of the considered
157
expert’s coefficients along each modality. "a.u." refers to the utilization of arbitrary units.
Chapter 7. Results
Finger 1 Finger 2 Finger 3 Finger 4 Finger 5

Spectral influence
(a.u)
0 100 200 0 100 200 0 100 200 0 100 200 0 100 200
Frequency (Hz) Frequency (Hz) Frequency (Hz) Frequency (Hz) Frequency (Hz)
Temporal influence
(a.u)
-1 -0.5 -1 -0.5 -1 -0.5 -1 -0.5 -1 -0.5

Time (s) Time (s) Time (s) Time (s) Time (s)
0.25
Spatial influence
0.2
0.15
(a.u.)
0.1
0.05
Electrodes Electrodes Electrodes Electrodes Electrodes
(1cm inter-electrode (1cm inter-electrode (1cm inter-electrode (1cm inter-electrode (1cm inter-electrode
distance) distance) distance) distance) distance)
Figure 7.11: Clinical data set, example of the average contribution of frequency, temporal and spatial modalities to the MpWF linear
model (subject zt, one session). Contributions were assessed as the normalized summation of absolute values of models’ coefficients
158 along each modality. "a.u." refers to the utilization of arbitrary units.
7.3. Discussion 159
7.3 Discussion
The present results permit to assess the efficiency of the MSLM for accurate asyn-
chronous mono-limb and sequential multi-limb decoding. The relevance of the three
key features of the MSLM, namely switching modelling, discriminative modelling
and dynamic gating, is of particular interest.
The pertinence of switching decoding, which is particularly related to the poten-

tial existence of context-specific modifications in a model of dependence between
neural features and kinematic parameters, has been suggested in several MUA/SUA
BCI studies [Kim et al., 2006c] [Ifft et al., 2013] [Velliste et al., 2014]. It was observed
in [Velliste et al., 2014] that neural patterns generated during NC states did not fit
a linear model trained on IC periods. The models permitting to decode bimanual
movements from MUA/SUA in monkeys were found in [Ifft et al., 2013] to be different
from the models independently trained for the decoding of unimanual left and right
movements. These findings seemed to indicate the existence of an unimanual and a
bimanual control modes associated with different models between neural features
and kinematic parameters. In [Yu et al., 2007], switches between goal-specific models
permitted to improve the decoding of reaching movements executed by monkeys.
Different models were associated with specific phases of reaching movements in [Kim
et al., 2003] and [Kang et al., 2012], and switches between emissions models were
exploited in [Wu et al., 2004] to improve Kalman-based estimation of kinematic
parameters from MUA/SUA signals. The interest of multi-model approaches has
additionally been emphasized in recent reviews or studies on MUA/SUA signals
[Rouse and Schieber, 2015] [Kao et al., 2017]. MUA/SUA and ECoG signals exhibit
different spatial resolutions and spectral contents (see Chapter 1). While the perti-
nence of kinematic decoding is for example well established for MUA/SUA signals, it
is still a matter of discussion for ECoG signals (see Chapter 1). The recent interest
for MUA/SUA switching modelling may thus be irrelevant for ECoG decoding,
for example if the model of dependence between ECoG features and kinematic
parameters is well approximated by a context-independent linear model. While the
use of switching models has been reported in a few ECoG studies for the integration
of NC support into kinematic decoders [Williams et al., 2013] [Bundy et al., 2016]
or for sequential multi-limb decoding [Flamary and Rakotomamonjy, 2012], the
validity and interest of the switching hypothesis has only been partially investigated
in [Williams et al., 2013], where it was observed that neural patterns elicited during
NC and IC states exhibited a different baseline, and where a post-processing strategy
was found suboptimal for the discrimination between NC and IC states. It has not,
however, been clearly established whether fitting kinematic decoders on IC samples
exclusively may permit to improve ECoG decoding accuracy during IC periods. In
[Bundy et al., 2016] for example, the performance of linear kinematic models fitted
on IC samples was not compared to the one of models fitted on both NC and IC
samples, and the performance of these two training strategies was not specifically
evaluated on IC periods in [Flamary and Rakotomamonjy, 2012]. The high decoding
accuracy achieved by means of a post-processed linear model in [Wang et al., 2011]

further illustrates that the relevance of switching models is unclear for ECoG signals.
The results here reported may contribute to the exploration of the relevance of
switching modelling of ECoG signals, in particular for the task of accurate decoding
during IC periods and for NC support integration into mono-limb and/or multi-limb
kinematic decoders.
Several studies have been completed to assess the respective performance of
discriminant (static) and generative (dynamic) models for kinematic decoding of
MUA/SUA and ECoG signals (see Chapter 2). Although generative switching
models have been regularly applied for the decoding of MUA/SUA signals, their
use had not been reported for ECoG signals for which switching modelling is
generally performed within the framework of discriminative models. The relevance
of discriminative or generative frameworks for switching modelling thus remains to
be clarified, here specifically for the task of accurate asynchronous mono-limb and
sequential multi-limb control with linear models.
7.3.1 Mono-limb decoding

The MSLM and its key features were first evaluated for a task of accurate asyn-
chronous mono-limb trajectory reconstruction from ECoG signals.
7.3.1.1 Strategies for NC support
The preclinical data set permitted to complete a first study on the efficiency of the
MSLM for the introduction of NC support into mono-limb kinematic decoders. The
training of the proposed MSLM decoder was considered in the case of complete
training data sets, i.e. parameter identification was supervised with respect to both
trajectory and state sequences. A comparison of the MSLM with a Wiener Filter
with Markovian post-processing and a SKF was performed to assess the relevance
of two of the MSLM key features, namely the use of switching models rather than
of a single post-processed model and the framework of discriminative rather than
generative modelling. An additional comparison with a SLM and pWF, i.e. variants
of the MSLM and MpWF with static state detection, was carried out to evaluate
the interest of the MSLM’s third key feature, namely dynamical gating.
MSLM-based state detection was found to significantly outperform the state

detection achieved by the MpWF as measured by the AUC. The TPR, FPR and
ACC of the MpWF were additionally degraded when compared to the ones associated
with the MSLM. The MSLM’s FPR was in particular lower than the MpWF’one by
a median of 26.6% and 14.4% for the subdural and epidural data sets, respectively.
These median improvements rose to 34.7% and 17.7% when synchronized probabilities
were considered on the subdural and epidural data sets, respectively, which seems
to indicate that the MSLM may exhibit a temporal uncertainty slightly superior
to the WF’s one. MSLM- and the MpWF-based state estimation mainly differ
7.3. Discussion 161
because the MSLM decoder is fed with neural signals while the WF is fed with the
trajectory estimates issued by a single kinematic model, i.e. a continuous model
trained and applied on both NC and IC states. The lower performance of the
MpWF with respect to the MSLM for state detection may indicate that optimal
states estimates cannot be inferred from the kinematic estimates, and thus that the
kinematic estimates yielded by a single NC/IC linear kinematic decoder may be
corrupted because the decoder is trained on both NC and IC samples. This finding
seems to suggest the existence of NC/IC-specific modifications of the considered
linear model of dependence between ECoG signals and kinematic parameters, and
are consistent with the NC/IC-specific changes of ECoG baseline activity which
were reported in [Williams et al., 2013].
In the case of the subdural data set, the MSLM’s IC expert and the MpWF
performed equivalently when they were applied on IC states (Table 7.8): excluding
NC samples when training the MSLM IC expert did not permit to improve the model
between subdural neural signals and kinematic parameters during IC periods. Thus,
although using a single continuous decoder resulted in a degraded state detection,
its ability to decode IC states was not inferior to the one of the IC-specialized expert.
These results suggest that possible modifications of the ECoG encoding may not be
sufficiently important to impact the quality of the linear kinematic model used to
decode neural features during IC periods on the considered subdural ECoG data set.
Different results were obtained on the epidural data set, where the IC experts
yielded IC kinematic estimates on known IC states with slightly but significantly
lower errors than the single linear model embedded in the MpWF model. Similar
results were observed when MSLMs embedding an IC expert trained on either IC
samples exclusively or on both NC and IC samples were compared, as the NMRSE
and NMAE were significantly better on the epidural data set when IC samples
only were used to train the IC expert. These observations are consistent with an
earlier ECoG study [Williams et al., 2013] where it was advocated to utilize NC-
and IC-specific intercepts when applying a linear model to decode both NC and IC
samples, so that as to take into account differences in the baseline activity during
NC and IC states. Interestingly, this study had also been completed with epidural
ECoG signals [Williams et al., 2013]. A median improvement of 6% was additionally
found for the PCC (see Table 7.11). While this improvement was not significant,
the PCC between synchronized trajectories was found to be significant.
The main performance difference between SKF-based and MSLM-based switching

modelling arose for state detection. The AUC and ACC associated with the MSLM
were found to be significantly higher than the ones obtained with the SKF on both
the epidural and subdural data sets. No significant difference was observed between
SKF- and MSLM-based kinematic decoding accuracy during IC state. The MSLM’s
IC expert was fitted using PLS regression. The fact that the SKF was fed with
features obtained after a PLS-based dimensionality reduction may have contributed
to this similarity between the MSLM’s and SKF’s IC decoding performances. If
both the SKF and MSLM are switching decoders which model the switching state
succession as a first-order Markovian process, their respective statistical frameworks

differ. The SKF being a generative model, its state estimation procedure relies on
the consistency of the NC- and IC-specific emission and transition models with the
observed neural signals. Dimensionality reduction is additionally required for online
application of the SKF, and a suboptimal feature selection may result in degraded
kinematic estimates. Because it was designed as a discriminative model, the MSLM
can directly exploit differences between the distributions of NC and IC high dimen-
sional neural features. The generative framework here appeared less relevant than the
discriminative one for the considered task of state detection from high-dimensional
neural signals. The SKF exhibited a higher FPR than the MSLM. This high FPR
may impair the SKF’s applicability for the integration of NC support into kinematic
decoders (see Figure 7.1), for example when high-dimensional neural features are
considered. Stable NC positions were here considered, because it is desirable for the
limbs of CLINATEC’s exoskeleton to be located at an unique position during all
blocks of NC states. As the SKF is theoretically able to handle variable NC po-
sitions, its relevance may be increased for applications with non-unique NC positions.
Finally, the interest of the MSLM’s Markovian hypothesis was illustrated when it
was compared it with a SLM, i.e. its analogue with static gating. While performance
indicators based on the confusion matrix were similar for the MSLM and the SLM,
the MSLM corresponded to significantly fewer false activations and deactivations,
i.e. blocks of consecutive false positives or false negatives, on the epidural data set.
Although it was not found significant, the same trend was observed on the subdural
data set. False activations and deactivations were longer (median increase: 30.1%
and 19.2% respectively, on the subdural data set, and 26.3% and 28.9% respectively,
on the subdural data set) but fewer for the MSLM than for the SLM (median
improvement of the false activations’ and deactivations’ frequency of 32.7% and
16.7%, respectively, on the subdural data set, and of 31.9% and 26.0%, respectively,
on the epidural data set). As a block of adjacent misclassified samples is expected
to be less disturbing to BCI users than a few isolated erroneous state estimates,
these results seem to confirm the interest of the dynamic state detection performed
by the MSLM.
7.3.1.2 Multiple IC experts for accurate decoding
Decoding accuracy during IC is required for users to exert efficient kinematic control
over an effector [Marathe and Taylor, 2011]. Three subdural acquisition sessions
were further analysed because they exhibited complex reaching movements for which
one IC expert could be insufficient for accurate kinematic modelling. MSLM with
k = 2 IC experts were identified using the unsupervised training approach proposed
in the present doctoral work. The majority of decoders which use has been reported
for ECoG kinematic decoding are static or dynamic linear models [Wang et al.,
2013c] [Schalk et al., 2007] [Chao et al., 2010] [Shimoda et al., 2012] [Nakanishi et al.,
2013] [Williams et al., 2013] [Hammer et al., 2013] [Wang et al., 2013c] [Hotson
7.3. Discussion 163
et al., 2014] [Bundy et al., 2016]. They do not exploit potential modifications of
the model of dependence between ECoG signals and kinematic parameters within
movement phases. While the use of mixtures of linear IC models has been considered
for MUA/SUA decoding [Kim et al., 2006c], the relevance of switching modelling
of ECoG signals is not straightforward because of the comparatively lower spatial
resolution and spectral content of ECoG signals which even challenge the efficiency of
generic kinematic decoding. For two of the sessions, the cross-validated PCC, NMAE
and, for one of the session the NRMSE, were significantly improved when computed
over hidden IC samples. The NMAE was significantly improved on the third data set.
The preliminary study led on data sets exhibiting complex, multi-phase movements
suggests that MSLMs with k > 1 IC experts may be profitably used to decode
complex movements, and therefore that phase-specific model modifications may be
observed in ECoG signals.
7.3.2 Multi-limb decoding

The relevance of the MSLM for asynchronous multi-limb decoding was investigated
on a publicly available clinical data set acquired while human subjects were execut-
ing finger movements. Similarly to the mono-limb asynchronous study, the MSLM
was compared to both a SKF and a Wiener filter with Markovian post-processing.
The three models intrinsically prevent large parallel limb activations, as one limb
only is associated with each possible value of the latent hidden state. Parameter
identification was supervised with respect to both trajectory and state sequences.
The MSLM significantly surpassed the MpWF for state detection in terms of
ACC and FPR (median improvements of 4% and 33%, respectively). The observed
state detection improvement suggests that a linear model of dependence between
neural signals and kinematic parameters may depend on the activated limb. Con-
sistently with a possible existence of state-specific modifications of a linear model
of dependence between neural features and kinematic parameters, the IC experts
resulted in kinematic estimates with significantly lower errors for both hidden and
known state sequences, and, in the case of synchronized estimated trajectories,
higher correlation coefficients than the MpWF. Interestingly, significantly higher
PCC were obtained on IC samples when the MSLM embedded experts fitted on IC
rather than on both NC and IC samples (unknown state sequence, see Table 7.27).
This observation is coherent with the results obtained on the epidural preclinical
data set. The MSLM superior state and kinematic decoding abilities were associated
with one drawback, namely the fact that errors were larger when they did occur
(False Positives or activation of a wrong finger) (see Figure 7.8). It is expected that
this drawback is profitably compensated for by the state detection and kinematic
ability of the MSLM, as a slight increase of false activation amplitude is likely to be
of limited importance to subjects.
The SKF was outperformed by the MSLM with respect to both state detection
and trajectory reconstruction during IC periods. Although it had been proved

efficient for IC kinematic decoding when it is correctly initialized at the beginning
of each movement (see section C.3.2 in Appendix C), the SKF here presented
a poor decoding ability when the switching states were known but the SKF was
not reinitialized at the beginning of each movement (see illustration in Figure 7.6).
The SKF performance was further degraded in the case of unknown state sequences.
It exhibited both a poor state detection along with a limited decoding accuracy
during IC periods. These results seem to indicate that the generative framework is
less relevant than the discriminative one for switching modelling with more than 2
experts, in particular that the MSLM’s direct exploitation of the distribution of the
neural features is more efficient for state detection than the SKF’s indirect strategy.
The utilization of a SKF-like model had not yet been proposed for asynchronous
multi-limb decoding, i.e. for K > 2. The results here obtained for the task of
multi-limb decoding suggest that the use of the SKF may not be easily extended for
multi-limb NC support, because it yields poor kinematic estimates which low accu-
racy may hinder its applicability for closed-loop ECoG-driven motor BCI systems.
Further analyses are required to clarify the limits of the SKF. Mixtures of KF were
for example considered in e.g. [Yu et al., 2007], where they were profitably used to
extract goal-directed trajectories from MUA/SUA signals in monkeys. The SKF
efficiency may have been here limited by ECoG spatial resolution or by the loss of
informative neural features during dimensionality reduction.
Finally, similarly to the mono-limb asynchronous study, the performance of

the MSLM for state detection was assessed by comparing it with the SLM, its
analogue with static state detection. The MSLM corresponded to higher median
ACC and CTPR than the SLM. It was additionally associated with less frequent
false deactivations than the SLM.
7.3.3 Absolute decoding performance

The comparative relevance of the three key features of the MSLM has thus been
assessed on both preclinical and clinical ECoG data sets, and the reported results
suggest the interest of using the proposed decoder for ECoG asynchronous kinematic
decoding. The absolute decoding performance of the MSLM, however, is also crucial.
In particular, the ability of users to control reaching movements depends on the
accuracy of the kinematic parameters extracted from their neural signals [Marathe
and Taylor, 2011]. The IC decoding performances obtained on the preclinical data set
(median PCC around 0.7 and 0.3 for the reconstruction of 3D wrist movements during
known IC states for subdural and epidural ECoG signals, respectively) suggest that
kinematic decoding is less challenging for subdural ECoG signals than for epidural
ones. Incidentally, these results are consistent with the PCC reported in [Bundy
et al., 2016], where 3D trajectory reconstructions from subdural ECoG signals
reached average PCCs of 0.49, 0.28 and 0.38 for the x-, y- and z-axis, respectively.
Similarly, state detection quality was higher for subdural acquisition sessions than for
7.4. Conclusion 165
epidural ones (average ACC of 0.89 and 0.96 for the subdurally implanted monkeys
A and K, respectively, and of 0.83 and 0.77 for the epidurally implanted monkeys B
and C). The FPRs, which are particularly important for clinical applications, range
from 0.3 and 0.11 (subdural data set, Monkey K and A, respectively) to 0.14 and
0.17 (epidural data set, Monkey C and B, respectively). While these state detection
performances may appear deterrent for clinical applications, they are consistent with
results reported in [Bundy et al., 2016] (ACC ranging from 0.68 to 0.90 depending on
subject, with an average of 0.80 for 5 subdurally-implanted subjects) and [Wang et al.,
2013b] (average ACC of 0.91 for 3 subdurally-implanted subjects). Studies have
suggested that state detection accuracy may be facilitated during closed-loop control
[Williams et al., 2013] (ACC of 0.98-0.99). The observed difference between subdural
and epidural signals is coherent with the findings reported in [Bundy et al., 2014],
where the quality of signals acquired by subdural micro-ECoG array (inter-electrode
distance of 1mm, against 3.5mm for the preclinical data set analysed in the present
doctoral thesis) is superior to the epidural equivalents. The accuracy of the NC/IC
state detection completed on the clinical data set was stable from one subject to
another, and comparable to the results obtained on subural preclinical data (average
ACC of 0.85 for 9 subjects, standard deviation of 0.02). The discrimination between
fingers was found more challenging (average ratio of correctly classified fingers of
0.65), but results were consistent with results reported in [Saa et al., 2016], where
a CRF permitted to achieve an ACC of 0.65 for 6-class classification. Closed-loop
finger detection was performed with an accuracy of 0.92 for NC/IC detection and
of 0.75 for inter-finger discrimination [Hotson et al., 2016], again suggesting that
performance improvements are observed after training. An average PCC of 0.47
was obtained for MSLM-based multi-finger trajectory reconstruction on known IC
samples. Thus, the IC decoding accuracy was lower for the clinical subdural data set
when compared to the preclinical subdural one. The larger inter-electrode distance
of the clinical ECoG arrays may have contributed to this difference [Bundy et al.,
2014]. The PCC is slightly lower than the PCC reported in the acausal decoding
study completed in [Flamary and Rakotomamonjy, 2012]. Further investigations are
required to assess whether the exploitation of acausal information would permit to
the MSLM to achieve the same decoding performance.
7.4 Conclusion
The decoding performance of the Markov Switching Linear Model has been presented
and compared to the performance of alternative decoders for two decoding tasks,
namely accurate asynchronous mono-limb and sequential multi-limb decoding. The
reported results have permitted to validate the relevance of the three key features
of the MSLM, namely switching modelling, discriminative modelling and dynamic
state detection.
A comparison of the MSLM with a Wiener Filter with Markovian post-processing
(MpWF) suggested the interest of switching modelling of ECoG signals, in particular
to improve the accuracy of kinematic estimates during IC states. This interest did
not clearly arise from earlier studies led on ECoG signals. In particular, it is not well
established whether linear models of dependence between ECoG neural features and
kinematic parameters are state- or context-specific, and whether taking modifications
of this model into account may permit to improve decoding accuracy. Improvements
of the decoding accuracy were here observed for mono-limb kinematic decoding
from epidural preclinical ECoG signals and from clinical signals measured during
multi-limb movements. A reduction of spurious activations during NC states and of
erroneous limb activations during IC periods was also observed. As the unsupervised
training of MSLM with k = 2 IC experts permitted to improve IC decoding accuracy
in three acquisition sessions exhibiting complex, multi-phase movements, switching
modelling may also be of interest for multi-phase ECoG kinematic decoding.
An additional comparison was completed with a Switching Kalman Filter, i.e.
a switching generative model, to assess the relevance of discriminative switching
modelling with respect to generative switching modelling. To the best of our
knowledge, such comparison had not been drawn in earlier MUA/SUA or ECoG
studies. While the SKF permitted to achieve satisfying decoding accuracy over IC
periods on the preclinical data set, it was outperformed by the MSLM in terms of state
detection accuracy. The SKF additionally exhibited poor decoding performances
when applied for multi-limb sequential ECoG decoding. While generative kinematic
decoders are widely popular for kinematic decoding, these results seem to indicate
that the generative framework may be suboptimal for switching ECoG modelling.
The Markovian hypothesis used to perform dynamic state detection was here
found to be advantageous to limit short spurious system activations or deactivations.
The relevance of its three key features and of the associated training strategies
permitted the MSLM to address the specific challenges of this thesis, i.e. namely
accurate asynchronous mono-limb and sequential multi-limb decoding. The MSLM
corresponded to improved state detection and/or kinematic parameter estimation
on the considered ECoG data sets. Finally, the resulting accuracy is expected to be
compatible with closed-loop control.
Chapter 8
Discussion and perspectives
Contents
8.1 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 167
8.1.1 Challenges for motor BCI clinical applications . . . . . . . . . 167
8.1.2 Proposed methodology . . . . . . . . . . . . . . . . . . . . . . 169
8.1.3 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 171
8.1.4 Limits of the present work . . . . . . . . . . . . . . . . . . . . 173
8.2 Perspectives . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 174
8.2.1 Technical optimizations . . . . . . . . . . . . . . . . . . . . . . 174
8.2.2 Closed-loop evaluation during CLINATEC’s clinical trial . . . 176
8.1 Discussion
In spite of proofs of concept in laboratory environments [Collinger et al., 2013]
[Wodlinger et al., 2015], clinical applications of motor BCIs remain rare [Mak and
Wolpaw, 2009]. Technical obstacles to the clinical deployment of motor BCI systems
are addressed in the present doctoral thesis.
8.1.1 Challenges for motor BCI clinical applications

Three technical challenges have been particularly considered, namely asynchronous
mono-limb decoding, asynchronous sequential multi-limb decoding and decoding
accuracy during active states.
A major issue for BCI clinical applications is the ability to provide users with
accurate asynchronous control over the effector [Graimann et al., 2009]. Unlike
synchronous BCIs which are periodically controllable by users, asynchronous BCI
decoders are continuously available. The limitation of spurious effector activations
during No-Control (NC) periods is desirable for asynchronous BCI systems. In
contrast with communication BCIs which generally permit users to control virtual
effectors (e.g., a cursor displayed on a computer screen), motor BCI systems embed
physical effectors such as robotic prostheses or orthoses. False activations are likely
to be particularly disturbing and stressful to users when they physically interact
with the BCI effector. NC support, i.e. the ability of the BCI decoder to output
168 Chapter 8. Discussion and perspectives
neutral values during NC states, is thus especially important for orthosis-based

motor BCI systems.
Although it has been scarcely explored by the BCI community, multi-limb
decoding is expected to improve the independence of severely motor impaired
patients. Upper- and lower-limb control would for example be particularly beneficial
to patients with tetraplegia. The extension of mono-limb asynchronous decoding to
multi-limb asynchronous control is one of the challenges arising for multi-limb BCI
systems. A sequential upper-/lower-limb activation strategy was here considered. It
was chosen for CLINATEC’s BCI system with the aim of improving the system’s
robustness and of facilitating neural control. The corresponding decoding objective
lies in avoiding both spurious system activations and parallel, residual movements
of the momentarily non-controlled limbs.
Finally, the ability of users to execute brain-controlled movements is impaired
when the estimates of the user’s intentions are not sufficiently correlated with his
true intentions [Marathe and Taylor, 2011]. Decoding accuracy during Intentional
Control (IC) periods is thus crucial for motor BCIs.
A decoder, namely a Markov Switching Linear Model (MSLM), has been devel-
oped in the present doctoral work for the task of accurate, asynchronous sequential
multi-limb kinematic ECoG decoding.
The interest of switching models for kinematic reconstruction from ECoG signals
has been investigated. Their relevance depends on the presence of state-related
modifications in linear models of dependence between neural features and kinematic
parameters. While the presence of such changes has regularly been illustrated in
MUA/SUA studies, it has not yet been clearly established for ECoG signals.
The respective relevance of generative and discriminative approaches for switching
modelling of ECoG signals has additionally been explored. Although generative
switching models have been regularly applied for the decoding of MUA/SUA signals,
their use had not been reported for ECoG signals for which switching modelling is
generally performed within the framework of discriminative models. Their respective
relevance of both switching frameworks remained to be investigated for ECoG
signals.
Both supervised and unsupervised training procedures have been presented
for the proposed discriminative Markov Switching Linear Model. An evaluation
procedure, which includes block-wise criteria and permits to separate spatial from
temporal errors, has been proposed to assess the comparative decoding performances
of the MSLM and of alternative decoders which use has been reported within the
BCI community.
Finally, the ability of the proposed model to address 3 challenges associated with
clinical motor BCIs, namely efficient asynchronous mono-limb control, sequential
multi-limb control and accurate decoding of kinematic parameters during active
states, has been evaluated.
8.1. Discussion 169
8.1.2 Proposed methodology

The Markov Switching Linear Model has been designed as a switching discriminative
kinematic decoder with dynamic state detection.
8.1.2.1 Markov Switching Linear Model

The MSLM strategy to support NC periods in the case of mono-limb and multi-limb
sequential asynchronous control consists of switching between NC and IC models
to take into account state-specific modifications of a linear model of dependence
between neural features and kinematic parameters (see Chapter 3). Several IC
experts are used when complex movements are considered. Switches between IC
models are therefore used to address the issue of multi-limb effector control (one IC
model per limb) as well as to improve decoding accuracy during IC periods (several
IC experts to model complex movements executed with one limb). The relevance
of each model, i.e. of each state, is estimated from the neural data and is used to
combine the available continuous models.
Three key features have been used to design the MSLM (see Figure 5.3). First,
the MSLM has been developed as a switching model (or mixture of linear models),
in contrast with post-processed decoders which rely on a single model to describe
the dependence between neural signals and kinematic parameters. Switching models
are liable to improve both the quality of NC support for asynchronous decoding and
the accuracy of kinematic estimates during IC states if the model between neural
signals and kinematic parameters is context-dependent.
Second, it has been conceived as a discriminative switching model, as opposed
to generative switching models like the Switching Kalman Filter which utilization
has been proposed by the BCI community. More precisely, the MSLM has been
designed as an extension of Mixtures of Experts (ME) models. A static linear model
between neural features and kinematic parameters is conditioned on the current
hidden switching neural state. A probabilistic rule is used to weight the available
regression models. This soft weighting approach results in smoother transitions than
the widely used Winner-takes-all hard combination strategy, and is here expected to
facilitate asynchronous effector control.
Finally, in contrast with the previously reported BCI switching regression models,
the MSLM performs dynamic state detection to limit spurious effector activations in
asynchronous settings. More specifically, the sequence of NC and (possibly multiple)
IC states is assumed to be generated by a first-order Markov chain. A Hidden
Markov Model-based (HMM) discrete decoder is used for state estimation. The
corresponding MSLM thus extends Mixtures of Experts models by embedding se-
quential rather than static state detection.
Both supervised and unsupervised training procedures have been presented in

the present dissertation. Supervised training is possible when the values taken
by the switching state in the training data set are known. As NC/IC labels can
be inferred from the tracked limb trajectory, supervised training was utilized to
train 2-state and 6-state MSLMs to introduce NC support into mono-limb and
sequential multi-limb kinematic decoders. Unsupervised training permits to exploit
internal states without precise a priori knowledge on their value in the training
data set. It was here used to fit several IC experts to model complex unimanual
reaching movements. The proposed unsupervised training procedure is based on the
Expectation-Maximization (EM) algorithm, and extends both HMM- and ME- well
known EM training algorithms.
8.1.2.2 Implementation and evaluation approach
The proposed MSLM is flexible inasmuch as it can embed different HMM-based

gate and experts. The relevance of experts and gates may in particular be impacted
by the measurement noise associated with the experts, by the dimensionality of the
neural features and by their underlying distribution. The MSLM gate and expert
structures and/or the corresponding training approaches were therefore optimized
for the considered data sets in a preliminary study (see Chapter 6 and Appendix
B).
Benchmark decoders were chosen so as to reflect the strategies which use had
been reported in earlier studies for asynchronous (possibly multi-limb) kinematic
decoding (see Chapter 5). A Markovian post-processed Wiener Filter (MpWF)
was chosen to represent a first alternative to the MSLM, namely a single model
with post-processing-based (possibly multiple) NC support. A Switching Kalman
Filter (SKF) was additionally implemented to assess the comparative performance
of discriminative and generative switching models.
Performance indicators were selected to assess the ability of the MSLM to support
asynchronous mono-limb and multi-limb and to accurately reconstruct kinematic
parameters during IC periods (see Chapter 5). A first indicator subset focused
on the capacity of each decoder to handle asynchronous mono-limb or multi-limb
decoding, i.e. it gathered criteria measuring the quality of each decoder’s NC support
and/or of the discrimination between multiple active states. The performance of
brain switches is generally evaluated using indicators extracted from the confusion
matrix, for example the number of observations erroneously decoded as IC states. It
was here proposed to additionally consider block-wise criteria, namely to measure the
frequency and duration of false activation blocks. Because they take into account the
dynamic of the errors made by the state decoder, block-wise criteria are expected to
reflect the user’s perception of NC support quality during asynchronous closed-loop
control sessions. A second indicator subset was dedicated to the evaluation of the
accuracy of the kinematic parameters reconstructed by each decoder during IC
states. It was composed of the criteria typically associated with continuously-valued
variables. These generic criteria are sensitive to slight desynchronizations between
observed and estimated trajectories, and they therefore reflect both spatial and
temporal decoding accuracies. For this reason, it was proposed to complement them
by computing their synchronized analogues, i.e. to compute generic performance
8.1. Discussion 171
indicators after having synchronized the estimated trajectory with the observed
(tracked) trajectory. These synchronized criteria thus isolate spatial accuracy from
temporal accuracy.
8.1.3 Results
The ability of the MSLM decoder to address the issue of accurate asynchronous (pos-
sibly multi-limb) decoding was assessed for two decoding tasks, namely asynchronous
uni-manual trajectory reconstruction and multi-limb (multi-finger) trajectory re-
construction (see Chapters 4 and 6). Results suggest that the proposed MSLM
decoder may be profitably used to introduce NC support into ECoG asynchronous
kinematic decoders, possibly with sequential multi-limb control. It can additionally
be used to improve the accuracy of complex reaching movement estimates by com-
bining several IC experts. In particular, the results permit the validate the relevance
of the MSLM three key features (see Chapter 7).
First, the switching framework was found to be more efficient than the post-
processing framework for the two considered state detection tasks, namely NC/IC
and NC/multiple IC state detection. Post-processing-based NC support has the
advantage of a reduced training duration. Its efficiency is nevertheless degraded if
linear models between neural signals and movement kinematics are not consistent
across states, e.g. across NC and IC states. While the presence of such modifications
has been regularly reported in MUA/SUA studies, it has not been clearly established
for ECoG signals. Post-processing-based mono- and multi-limb asynchronous decod-
ing was here less accurate than MSLM-based decoding. In the case of mono-limb
control, a similar superiority of the switching hypothesis over the post-processing
one had been reported in [Williams et al., 2013], where the estimated velocity of a
cursor was found to provide a poor approximation of NC and IC states. Although
a post-processed decoder designed for parallel multi-finger movements has been
proposed in [Wang et al., 2011], to the best of our knowledge, post-processing
approaches had not yet been considered for sequential multi-limb decoding. Our
results suggest they may be less efficient that switching models for asynchronous
multi-limb control. It should finally be noted that although the switching hypothesis
appeared relevant for NC support, it did not systematically result in an improvement
of kinematic decoding accuracy during IC periods. Improvements of the IC decoding
accuracy were, however, observed for the epidural preclinical and clinical data sets.
The utilization of several IC experts did permit, however, to improve the decoding
of the complex movements exhibited in some sessions of the subdural preclinical
data set.
Discriminative switching models were additionally found to be more efficient
than generative switching models for asynchronous mono-limb and sequential multi-
limb accurate decoding. Generative dynamic models, e.g. Kalman Filters, are
frequently used for kinematic decoding. They have the benefit of embedding a
model of the movement dynamic, which can be advantageously used to constrain
the estimated trajectory smoothness [Koyama et al., 2010b]. The application of

generative dynamic models is nevertheless more complex than the application of
generic discriminative models. In particular, Kalman filtering involves recurrent
matrix inversions which are computationally expensive when neural features are
high-dimensional. A supplementary dimensionality reduction step may thus be
required for real-time application of the KF. Numerical issues due to roundoff errors
may additionally arise when a generic implementation of the KF is used [Tusell,
2011]. The ability of generative models to handle (possibly) multi-limb asynchronous
decoding was investigated in the present doctoral thesis. Switching generative models
had been profitably used for NC support in a simulated study [Srinivasan et al., 2007].
Additionally, their application has been proposed to improve MUA/SUA decoding
accuracy during reaching movements by combining trajectory models [Yu et al.,
2007], and to enhance kinematic reconstruction from SUA/MUA signals during IC
states by using several emission models [Wu et al., 2003b] [Wu et al., 2004]. In spite
of these studies, the relevance of generative switching models on ECoG data had not
yet been established, and their performance had not yet been compared with the
one of their discriminative analogues. Despite the advantage conferred by explicitly
modelling the kinematic dynamic during NC states, generative switching models
here appeared as less efficient that discriminative switching models for asynchronous
ECoG decoding. This suggests that Kalman filtering, which is increasingly popular
for kinematic decoding, may suffer from difficulties to handle asynchronous decoding.
Such difficulties were probably encountered in the MUA/SUA-based clinical trial
completed by Hochberg and colleagues [Hochberg et al., 2012], where it was reported
that the KF used for kinematic decoding was re-initialized at the beginning of each
trial. Kalman-based IC experts additionally yielded IC kinematic estimates which
were either equivalent or less accurate than the ones output by the MSLM linear
experts.
Finally, dynamic gating permitted to limit false detections and/or modify their
dynamic, namely to replace short and numerous false activations by fewer (but
longer) spurious activations which are expected to be less disturbing to users. The
relevance of dynamic detection is further suggested by the fact that recent clinical
motor BCI studies relied on a first-order Markovian hypothesis for efficient (possibly
NC) state detection in closed-loop control sessions [Fifer et al., 2014] [Hotson et al.,
2016] [Kao et al., 2017].
The absolute IC decoding performance which was reached by the MSLM on the
considered data sets is consistent with the decoding accuracy reported by several
teams on private data sets (e.g., [Bundy et al., 2016]). It has been shown in [Marathe
and Taylor, 2011] that position decoding with a PCC of 0.5 can be satisfyingly
used to control the velocity of an upper-limb effector during reaching movements.
For this reason, it is expected that the IC decoding ability of the MSLM makes it
compatible with closed-loop kinematic control of upper-limb effectors.
The reported results thus suggest that the proposed decoder may be profitably
used for ECoG-driven accurate closed-loop asynchronous mono-limb or multi-limb
8.1. Discussion 173
decoding.
8.1.4 Limits of the present work

Because the present study was performed to prepare for CLINATEC’s BCI clinical
deployment, it was not completed with optimal chronic clinical ECoG data and
therefore presents several limitations.
8.1.4.1 ECoG array and subjects

The relevance of the MSLM for asynchronous mono-limb decoding was assessed
using preclinical ECoG data. Application of the MSLM on human ECoG data is
needed to confirm the trends observed on an inherently limited monkey model [Ball
et al., 2009b]. Although the MSLM ability to reconstruct asynchronous multi-limb
movements has been measured on clinical ECoG data, additional tests are also
required to address the shortcomings of this data set. While epileptic subjects are
considered as safe models of non-epileptic patients if some guidelines are respected
[Lachaux et al., 2012], data acquisition was performed using ECoG arrays which
characteristics and implantation area had been chosen with the goal of localizing
the patients’ epileptic focus. In particular, the exposed electrode diameter was
2mm, and the 1cm-long inter-electrode distance resulted in a sparse sampling of
the subjects’ motor cortex. Subdural acquisition by macro-ECoG arrays has been
shown to result in signals equivalent to those measured by epidural macro-ECoG
arrays [Bundy et al., 2014]. Similarly, the results presented in [Wang et al., 2016]
suggest that dense ECoG arrays are desirable for BCI applications.
The fact that the considered data sets were acquired from able-bodied subjects is
another limit of the present study. Severely motor-impaired patients cannot execute
overt (real) movements. Decoders are usually initialized by presenting subjects
with instructions (e.g., movements executed by an experimenter), and by asking
them to attempt to perform the proposed movements (e.g., [Hochberg et al., 2012]
[Wodlinger et al., 2015]). A first potential difficulty lies in a possible difference
in the strength of the patterns elicited by attempted movements and the ones
generated by the execution of real movements. It has been suggested in several
studies that attempted movements elicit ECoG patterns which are exploitable for
decoder calibration [Collinger et al., 2014]. An accuracy drop may nevertheless be
observed for motor-impaired patients. Additionally, the intentions of the subject are
not accessible and are approximated by the instructions given to them. An unknown
lag exists between these instructions and the corresponding movement attempts. If
reaction times generally exhibit a small variability, episodic losses of concentration
of the user may occur and be added to the reaction time.
8.1.4.2 Self-paced multi-limb complex 3D movements

Although a task of multi-finger control permitted to measure the MSLM ability to
manage sequential limb activation, the ability of the MSLM to handle upper- and
lower-limb sequential activation remains to be explored. In particular, lower-limb

movements may be harder to detect because they are mainly controlled in a deeper
cortex area than fingers or arm (see the presentation of the somatotopic organization
of the motor cortex in section 1.2.2.1). On the other hand, the approximate
lateralization of motor control may make it easier to exploit spatial patterns to
discriminate between left and right arm movements than to classify the activation
of fingers of the same hand.
1D limb movements were considered in the case of the clinical data set. As for
the preclinical data set, the wrist trajectories used to assess decoders’ performance
were associated with a limited space exploration. It is thus desirable to explore the
ability of the MSLM to decode complex 3D (possibly multi-limb) trajectories. As a
very limited number of complex reaching movements were present in the considered
preclinical data set, additional studies are particularly required to further investigate
the interest of using multiple IC experts to model complex reaching movements.
The NC states present in the preclinical and clinical acquisition sessions exhibited
similar temporal characteristics, i.e. they were a few seconds long and the standard
deviation of their duration was rather low (up to 3 seconds, see Chapter 4). The
ability of the MSLM to detect NC states with more variable durations needs to be
assessed.
8.2 Perspectives
Further investigations are necessary to strengthen and extend the hereby reported
results, and to address the limits of the present study.
8.2.1 Technical optimizations

The relevance of several potential technical improvements remains to be explored.
8.2.1.1 Expert and gate structure
The study which was here completed to choose the gate and expert structure and
training approach (see Appendix B) is not exhaustive. Several options may be
considered to further optimize MSLM gate and experts.
Gate The MSLM gate relies on a HMM for dynamic state estimation. While
the MSLM gate was able to detect state of moderately variable durations (see
data set description in Chapter 4), the fact that the duration of each state is
implicitly modelled by the HMM transition matrix (see section 3.2.1) may hinder
the MSLM efficiency for some applications, for example in the case of highly variable
state durations. Explicit modelling of the movement duration, for example using
explicit-duration HMMs (also referred to as Hidden semi-Markov Models) [Yu, 2010],
may be explored in the future.
8.2. Perspectives 175
Another structural modification could benefit state detection in the case of

multiple IC states, namely the utilization of a hierarchical state detection strategy.
This approach consists of training a first classifier to detect NC and IC states and a
second one to discriminate between multiple IC states. Its use has been reported
in several EEG [Murguialday et al., 2007] and ECoG studies [Hotson et al., 2016]
[Wang et al., 2016].
State-specific experts (MSLM’s continuous models) Although linear ex-

perts have been used in the present work, the utilization of nonlinear experts may be
considered. The use of nonlinear decoders, for example Generalized Linear Models
[Eliseyev and Aksenova, 2014], Multilayer Perceptrons [Kim et al., 2006b] or Support
Vector Machine Regression [Kim et al., 2006b] [Mehring et al., 2003], has been pro-
posed for the estimation of kinematic parameters from ECoG or MUA/SUA signals.
The performance of the MSLM may additionally be compared to the previously
mentioned nonlinear models for the task of asynchronous ECoG decoding.
8.2.1.2 Training procedures
The interest of several modifications of the MSLM training procedures may be

investigated.
Supervised training Improvements may be brought to the gate training pro-

cedure. PLS-based dimensionality was here cascaded with a logistic regression
model for state detection on the preclinical data. Generic and penalized PLS-based
training approaches have been proposed for logistic regression [Marx, 1996] [Fort
and Lambert-Lacroix, 2005], and may permit to perform optimal dimensionality
reduction during the training of the logistic regression model. While their computa-
tional cost was here deemed deterrent, the interest of these identification algorithms
needs to be thoroughly investigated.
The quality of the MSLM linear models may be enhanced by using the efficient
N-way training approaches developed in CLINATEC. In particular, (penalized)
N-way PLS algorithms have been proposed for sparse [Eliseyev et al., 2012] and/or
smooth [Eliseyev and Aksenova, 2016] expert training, and may be profitably used
to train the MSLM experts.
Unsupervised and semi-supervised training The approach here used for the
selection of the number of IC experts may be improved by recoursing to variational
Bayesian model identification methods [Bishop, 2006]. Variational Bayes training
generalizes Expectation-Maximization training by including hyperparameters in the
set of hidden variables. Though more complex, they here have the interest that the
number of active models can thus be directly estimated during training.
State labelling can be a tedious and error-prone process. Difficulties to assign
NC/IC labels to neural data were for example reported in [Suway et al., 2013]. Such
difficulties may additionally be encountered in the case of open-loop, observation-

based model calibration for motor-impaired patients. Although presented with
instructions, subjects are liable to initiate attempted movements with a certain time
lag, and/or to loose concentration during a few instructed movements. Labelling
these movements as IC may degrade the MSLM decoder. Semi-supervised training
strategies may be used to address this issue, for example by using probabilistic
NC/IC labels and by adjusting the probability during training. Semi-supervised
training has for example been considered for HMMs [Bordes and Vandekerkhove,
2005] [Scheffer et al., 2001] [Ozkan et al., 2012]. Its extension to MSLM training
may be explored.
Recursive supervised or unsupervised training Neural patterns have been

shown to change during closed-loop training [Carmena et al., 2003] [Rouse and Moran,
2009]. Turn-taking training strategies, i.e. decoder identification steps inserted
between user training sessions, have therefore been utilized in several motor BCI
clinical studies [Hochberg et al., 2012] [Wang et al., 2013c]. A promising extension
of turn-taking strategies consists of using adaptive training algorithms for online
adjustments of the decoder parameters, i.e. parallel rather than sequential user and
decoder training [Dangi et al., 2014]. Recursive training may thus be advantageously
considered for online adaptation of the MSLM to the new neural patterns. Several
algorithms have been proposed for HMM recursive training [Khreich et al., 2012].
Efficient recursive training algorithms have additionally been proposed in CLINATEC
for N-way PLS model identification [Eliseyev and Aksenova, 2013]. In the case of
supervised training, they may permit to perform recursive training of the MSLM.
8.2.2 Closed-loop evaluation during CLINATEC’s clinical trial

CLINATEC "BCI and Tetraplegia" clinical research protocol, which Principal In-
vestigator is Prof. A.-L. Benabid [ClinicalTrials.gov, 2016], will permit to further
explore the relevance of the MSLM and of the above-mentioned potential improve-
ments. Its objective is the chronic deployment of an accurate, multi-limb, multi-DoF
(Degrees of Freedom) motor BCI in humans. The wireless 64-channel ECoG im-
plant WIMAGINEr has been specifically designed for stable and long-term signal
acquisition [Mestais et al., 2015]. CLINATEC’s BCI platform [Eliseyev et al., 2014]
also includes the 4-limb exoskeleton EMY [Morinière et al., 2015] and the software
environments required for real time processing of the neural signals. WIMAGINE’s
inter-electrode distance of 4mm [Mestais et al., 2015] will permit a dense sampling
of human users’ ECoG motor activity. CLINATEC’s clinical trial will permit to
investigate the ability of the MSLM to decode complex, self-paced multi-limb move-
ments during closed-loop experiments.
As the forward algorithm used by the MSLM for model gating is computationally
efficient [Rabiner, 1989], the MSLM is expected to be compatible with real-time
BCI decoding.
8.2. Perspectives 177
A soft combination strategy has been chosen for the MSLM experts because it
was expected that hard combination results in disturbingly abrupt transitions and
increases the delay the user perceives between his intention and the system’s reaction.
Closed-loop experiments only, however, will permit to validate or invalidate this
choice. In particular, the possible impact of the combination strategy on subject
training is yet to be investigated.
As mentioned ealier, changes of context between open-loop and closed-loop BCI
settings result in differences between open-loop and closed-loop neural patterns
[Leuthardt et al., 2006a] [Jackson and Fetz, 2011] [Jarosiewicz et al., 2013]. Decoder
adaptation during closed-loop experiments has been regularly used in preclinical and
clinical motor BCIs to handle these pattern modifications (e.g., [Wang et al., 2013c]
[Wodlinger et al., 2015]). CLINATEC’s clinical trial will permit to investigate the
need for regular adaptation of the MSLM, and to select the best algorithmic strategy
to achieve the closed-loop training of the MSLM, namely punctual recalibration or
online adaptation via recursive training algorithms.
CLINATEC’s clinical trial will additionally permit to further assess the interest
of the MSLM for the decoding of complex movements, which was here suggested
by the results obtained on a limited data set. Complex movements are crucial
for the execution of everyday tasks, for example object manipulation. Object
manipulation may in particular involve (coordinated) bimanual movements [Swinnen
and Wenderoth, 2004]. It may thus be desirable for the MSLM to have the ability to
handle parallel multi-limb movements, e.g. coordinated bimanual movements. The
study completed by Ifft and colleagues in [Ifft et al., 2013] suggests that the MSLM
switching hypothesis may be relevant to decode parallel multi-limbs movements, as
they reported that modifications of the tuning properties of individual neurons were
induced when NHPs performed unimanual or bimanual reaching movements (see
Chapter 3). The relevance of the MSLM for object manipulation also depends on
its ability to handle multi-phase movements, in particular phases of stabilization
over objects. Stabilization difficulties have been reported in several MUA/SUA
motor BCIs, and have led several teams to address this issue [Kang et al., 2012]
[Golub et al., 2014] [Gowda et al., 2014] [Kao et al., 2017]. Stabilization was for
example improved by reducing the effector speed when error neural signals and/or
hectic trajectories were detected [Gürel and Mehring, 2012] [Golub et al., 2014]. The
ability of the MSLM to handle hold periods will be explored during CLINATEC’s
BCI clinical application. The detection of a "hold-on-target" switching state [Kang
et al., 2012] may permit to help users to stabilize on targets.
Appendix A
Supplementary materials for

MSLM training
Contents
A.1 Supervised training: decoupling of parameter estimation . 179
A.2 Unsupervised training: E-step . . . . . . . . . . . . . . . . . . 182
Supplementary materials on the supervised and unsupervised training of the

MSLM are gathered in the present appendix.
A.1 Supervised training: decoupling of parameter esti-

mation
Let us consider a MSLM composed of K linear experts gated by a HMM-based
sequential decoder. Each linear expert i is parametrized by Bi , where, depending
on context, Bi refers either to the full set of parameters associated with expert
k (i.e., parameters of the linear model and of the noise distribution) or only to
the parameters of the linear model. Let A = (aij ) be the transition matrix aij =
P (z t+1 = j|z t = i), i, j = 1, ..., K and π be the initial state distribution πi = P (z 1 =
zi ), i = 1, ..., K associated with the first-order Markovian hidden state variable
z t . Finally, let {b1 , . . . , bK } gather the parameters necessary to characterize the
distributions P (xt |z t = j), j = 1, . . . , K, e.g. mean and variance in the case of
Gaussian emissions.
Knowledge about the values taken by the switching variable in the training data
set permits to decouple the identification of the gate and experts parameters.
Let us consider the MSLM data complete log-likelihood
Lc (Θ, X, Y, z) = ln P (X, Y, z|Θ), (A.1)

where
P (X, Y, z|Θ) = P (X, Y|z, Θ) P (z|Θ) = P (x1:T , y1:T |z 1:T , Θ) P (z 1:T |Θ). (A.2)
Let us consider the first term of (A.2), i.e. P (x1:T , y1:T |z 1:T , Θ). Following the
MSLM hypotheses presented in section 3.3,
180 Appendix A. Supplementary materials for MSLM training
T 1 2
P (x1:T , y1:T |z 1:T , Θ) = P xt , yt |z 1:T , Θ .
Ù
t=1
As the emission distribution of xt is only conditioned on z t and as yt only depends

on xt and z t ,
T
P (x1:T , y1:T |z 1:T , Θ) = P (xt , yt |z t , Θ),
Ù
t=1
with P (xt , yt |z t , Θ) = P (yt |xt , z t , Θ)P (xt |z t , Θ).

P (yt |xt , z t , Θ) = P (yt |xt , Bz t ) is the likelihood of yt for the expert relevant at
time t. P (xt |z t , Θ) is the HMM emission model corresponding to the state z t , and
parametrized by bz t . Thus, P (xt |z t , Θ) = P (xt |bz t ) and
T
1:T 1:T 1:T
P (yt |xt , Bz t )P (xt |bz t ).
Ù
P (x ,y |z , Θ) = (A.3)
t=1
The second term of A.2 can be rewritten as
P (z 1:T |Θ) = P (z T |z 1:T −1 , Θ) P (z 1:T −1 |Θ)
Taking into account the first-order Markovian hypothesis associated with the latent
state variable z t and the chosen parametrization,
P (z 1:T |Θ) = P (z T |z T −1 , A) P (z 1:T −1 |A, π).
By recurrence,
−1
TÙ −1
TÙ
P (z 1:T |Θ) = P (z 1 |π) P (z t+1 |z t , A) = πz 1 az t , z t+1 . (A.4)
t=1 t=1
Finally, from A.2, A.3 and A.4,
−1
TØ T
ln P (yt |xt , Bz t )
Ø
Lc (Θ, X, Y, z) = ln πz 1 + ln az t ,z t+1 +
t=1 t=1
T
ln P (xt |bz t ). (A.5)
Ø
+
t=1
Let us introduce the following notation: δz t ,k = 1 if z t = k and δz t ,k = 0 otherwise,

and ωi,j
t = 1 if z t = i and z t+1 = j, and ω t = 0 otherwise.
i,j
Then A.5 can be rewritten as
A.1. Supervised training: decoupling of parameter estimation 181
K −1 Ø
TØ K
K Ø
t
Ø
i=1 t=1 i=1 j=1
K
T Ø K
T Ø
δz t ,i ln(P (yt |xt , Bi )) + δz t ,i ln P (xt |bi ). (A.6)
Ø Ø
+
t=1 i=1 t=1 i=1
Switching the sum operators,
K −1
K TØ
K Ø
t
Ø Ø
i=1 i=1 j=1 t=1
K Ø
T K Ø
T
t t
δz t ,i ln P (xt |bi ). (A.7)
Ø Ø
+ δz t ,i ln P (y |x , Bi ) +
i=1 t=1 i=1 t=1
As the contribution of the parameters of interest to the data log-likelihood is

decoupled in A.7, the maximization of each term of A.7 can be performed separately.
Maximizing the data log-likelihood amounts to solving the following maximization
problems, where the constraints which pertain to the transition and initial state
probabilities have been indicated:
T
δz t ,i ln P (yt |xt , Bi ) (A.8)
Ø
Bi
t=1
K
Ø
maximize δz 1 ,j ln πj
π
j=1
(A.9)
K
Ø
subject to πj = 1
j=1
K Ø −1
K TØ
t
Ø
maximize ωij ln aij
A
i=1 j=1 t=1
(A.10)
K
Ø
subject to aij = 1 ∀i ∈ {1, . . . , K} .
j=1
T
δz t ,i ln P (xt |bi ) (A.11)
Ø
maximize i = 1, . . . , K.
bi
t=1
The maximization problems (A.8), (A.9), (A.10), and (A.11) correspond to ML esti-
mation of the expert, initial state, transition and emission parameters, respectively.
The impact of the expert and emission parameters on the likelihood is additionally
separable with respect to each possible state (problems (A.8) and (A.11)).
A.2 Unsupervised training: E-step

The MSLM unsupervised training is completed via the Expectation-Maximization
algorithm. During the E-step, the expectation of the data log-likelihood (A.7) is
computed with respect to the latent vector z, given the observations X and Y and
using the current parameter estimate Θ̂:
è é K K TØ
K Ø −1
E Lc (Θ+ , X, Y, z)|Θ = Θ̂, X, Y = γi1 ln(πi+ ) + t
ln(a+
Ø Ø
ξi,j ij )
i=1 i=1 j=1 t=1
T
K Ø T
K Ø
γit ln P (yt |xt , B+ γit ln P (xt |b+
Ø Ø
+ i )+ i ), (A.12)
i=1 t=1 i=1 t=1
where
è é
γi1 = E δz 1 ,i |Θ = Θ̂, X, Y = P (z 1 = i|Θ = Θ̂, x1:T , y1:T ), (A.13)
è é
γit = E δz t ,i |Θ = Θ̂, X, Y = P (z t = i|Θ = Θ̂, x1:T , y1:T ), (A.14)
è é
t
ξi,j t
= E ωi,j |Θ = Θ̂, X, Y = P (z t = i, z t+1 = j|Θ = Θ̂, x1:T , y1:T ). (A.15)
During the M-step, the parameter estimate Θ̂ is updated by maximizing the expecta-
tion of the data log-likelihood. It thus amounts to solving the following maximization
problems:
T
γit ln P (yt |xt , Bi ) (A.16)
Ø
Bi
t=1
K
γj1 ln πj
Ø
maximize
π
j=1
(A.17)
K
Ø
subject to πj = 1
j=1
K Ø −1
K TØ
t
Ø
maximize ξi,j ln aij
A
i=1 j=1 t=1
(A.18)
K
Ø
subject to aij = 1 ∀i ∈ {1, . . . , K} .
j=1
T
γit ln P (xt |bi ) (A.19)
Ø
maximize i = 1, . . . , K.
bi
t=1
(A.16), (A.17), (A.18), and (A.19) correspond to ML estimation of the expert,

initial state, transition and emission parameters, respectively. In contrast with the
A.2. Unsupervised training: E-step 183
maximization problems associated with MSLM supervised training, the impact of

each state-specific parameter set is weighted by the state probabilities γ1t , γit or ξi,j
t
computed during the E-step. A decomposition of these probabilities is presented in

the first two paragraphs below. The derivation of the MSLM-specific forward and
backward algorithms necessary to compute intermediary probabilities is detailed in
the next paragraphs.
State probability estimation The state probabilities γit = P (z t = i|x1:T , y1:T , Θ)

is decomposed as follows:
γit = P (z t = i|x1:T , y1:T , Θ)

P (z t = i, x1:t , xt+1:T , y1:t , yt+1:T |Θ)
=
P (x1:T , y1:T |Θ)
P (yt+1:T , xt+1:T |z t = i, y1:t , x1:t , Θ)P (z t = i, x1:t , y1:t |Θ)
=
P (x1:T , y1:T |Θ)
P (y t+1:T ,x t+1:T |z = i, Θ)P (z t = i, x1:t , y1:t |Θ)
t
γit = .
P (x1:T , y1:T |Θ)
Transition probability estimation The transition probabilities ξi,j

t = P (z t =
i, z t+1 1:T 1:T

= j|x , y , Θ) are expressed as
t
ξi,j = P (z t = i, z t+1 = j|x1:T , y1:T , Θ)
P (z t = i, z t+1 = j, x1:T , y1:T |Θ)
=
P (x1:T , y1:T |Θ)
P (z t = i, z t+1 = j, x1:t , xt+1 , xt+2:T , y1:t , yt+1 yt+2:T |Θ)
=
P (x1:T , y1:T |Θ)
P (xt+2:T , yt+2:T |z t = i, z t+1 = j, x1:t , xt+1 , y1:t , yt+1 , Θ)
=
P (x1:T , y1:T |Θ)
× P (z t = i, z t+1 = j, x1:t , xt+1 , y1:t , yt+1 |Θ)
P (xt+2:T , yt+2:T |z t+1 = j, Θ)
=
P (x1:T , y1:T |Θ)
× P (xt+1 , yt+1 |z t = i, z t+1 = j, x1:t , y1:t , Θ)
× P (z t = i, z t+1 = j, x1:t , y1:t |Θ)
=
P (x1:T , y1:T |Θ)
× P (z t+1 = j|z t = i, x1:t , y1:t , Θ)P (z t = i, x1:t , y1:t |Θ)
=
P (x1:T , y1:T |Θ)
× P (z t+1 = j|z t = iΘ)P (z t = i, x1:t , y1:t |Θ).
The computation of the intermediary probabilities P (z t = i, x1:t , y1:t |Θ), P (xt+1:T , yt+1:T |z t =
j, Θ) and P (x1:T , y1:T |Θ) is performed using extensions of the forward and back-
ward algorithms originally developed for HMMs [Rabiner, 1989]. They extend the
HMM-specific ones by taking into account the likelihood of each continuous expert
when estimating the probability associated with each possible hidden state.
MSLM forward algorithm The MSLM forward algorithm permits to compute

the probabilities P (z t = i, x1:t , y1:t |Θ), t = 1, . . . , T . The recurrence is initialized
by P (z 1 = j, x1 , y1 |Θ) = P (y1 |x1 , z 1 = j, Θ)P (x1 |z 1 = j)P (z 1 = j). The following
recurrence formula is then utilized:
A.2. Unsupervised training: E-step 185
P (z t+1 = j, x1:t+1 , y1:t+1 |Θ) = P (z t+1 = j, x1:t , y1:t , xt+1 , yt+1 |Θ)
= P (xt+1 , yt+1 |x1:t , y1:t , z t+1 = j, Θ)P (x1:t , y1:t , z t+1 = j|Θ)
= P (xt+1 , yt+1 |z t+1 = j, Θ) P (x1:t , y1:t , z t+1 = j, z t = i|Θ)
Ø
i
t+1 t+1 t+1
= P (x ,y |z = j, Θ)
t+1
= j|z t = i, x1:t , y1:t , Θ)P (z t = i, x1:t , y1:t |Θ)
Ø
× P (z
i
t+1 1:t+1 1:t+1 t+1
P (z = j, x ,y |Θ) = P (x , yt+1 |z t+1 = j, Θ)
P (z t+1 = j|z t = i, Θ)P (z t = i, x1:t , y1:t |Θ),
Ø
×
i
where P (z t+1 = j|z t = i, Θ) = aij and P (xt , yt |z t = j, Θ) = P (yt |xt , Bj )P (xt |bj ).
MSLM backward algorithm Similarly, the MSLM backward algorithm permits

to compute P (xt+1:T , yt+1:T |z t = i, Θ). Following the generic HMM-specific back-
ward algorithm [Fink, 2014], the recurrence is initialized with P (xT , yT |z T = j, Θ) =
(1, . . . , 1)′ . The following iterations are then completed:
P (xt+1:T , yt+1:T |z t = i, Θ) = P (xt+1:T , yt+1:T , z t+1 = j|z t = i, Θ)

Ø
P (xt+1 , xt+2:T , yt+1 , yt+2:T , z t+1 = j|z t = i, Θ)

Ø
=
j
P (xt+1 , yt+1 |xt+2:T , yt+2:T , z t+1 = j, z t = i, Θ)

Ø
=
j
× P (xt+2:T , yt+2:T , z t+1 = j|z t = i, Θ)

P (xt+1 , yt+1 |z t+1 = j, Θ)
Ø
=
j
× P (xt+2:T , yt+2:T |z t+1 = j, z t = i, Θ)P (z t+1 = j|z t = i, Θ)

P (xt+1 , yt+1 |z t+1 = j, Θ)
Ø
=
j
× P (xt+2:T , yt+2:T |z t+1 = j, Θ)P (z t+1 = j|z t = i, Θ),
where P (z t+1 = j|z t = i, Θ) = aij and P (xt , yt |z t = j, Θ) = P (yt |xt , Bj )P (xt |bj ).
Appendix B
MSLM: expert training

procedure and gate selection
Contents
B.1 Expert training procedure selection . . . . . . . . . . . . . . 187
B.1.2 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 189
B.1.3 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 192
B.2 Gate selection . . . . . . . . . . . . . . . . . . . . . . . . . . . 193
B.2.2 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 196
B.2.3 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 199
The relevance of training procedures for linear experts and of approaches for the
modelling of the HMM-based gate’s emission probabilities depends on the problem
at hand. In particular, it can be impacted by the measurement noise associated with
the experts, by the dimensionality of the neural features and by their underlying
distribution. A comparative study was therefore completed to choose the training
approaches and/or structures which corresponded to optimal MSLM experts and
gate for the two considered data sets, namely the preclinical and clinical data sets
presented in Chapter 4. Supervised training procedures only were explored in this
study, i.e. NC/IC and NC/ICi , i = 1, . . . , 5 labels were available in the preclinical
and clinical training data subsets, respectively.
B.1 Expert training procedure selection

The MSLM exploits linear regression models (experts) to reconstruct kinematic
parameters during IC periods. It was assumed that these linear models were
associated with Gaussian noises. The corresponding generic ML training approach,
i.e. Optimal Least Squares (OLS), is generally unsuitable for high-dimensional
independent variables. A study was therefore completed to select the best training
procedure to approximate the ML estimator of the IC and ICi , i = 1, . . . , 5 regression
models for the preclinical and clinical data set, respectively.
188 Appendix B. MSLM: expert training procedure and gate selection
B.1.1 Methods and implementation

Penalized and/or projection-based training approaches were chosen on the basis of
their utilization in previous BCI studies, and used to identify the MSLM’s (possibly
multiple) IC experts.
Linear models have been used for kinematic reconstruction from ECoG signals
in several BCI studies (see section 2.4.3.1). Several training strategies have been
resorted to when high-dimensional neural features were considered. The relevance of
popular training approaches, namely ℓ1 -penalized regression (i.e., LASSO), Principal
Component Regression (PCR) and Partial Least Squares Regression (PLSR), was
assessed for trajectory reconstruction during IC periods on both the preclinical and
clinical data sets. All models were trained and tested on IC or ICi , i = 1, . . . , 5
samples exclusively.
Principal Component Regression (PCR) The utilization of PCR, i.e. the

regression of kinematic parameters against neural signals’ principal components,
has been considered in several BCI studies, e.g. [Spüler et al., 2016]. The number
of extracted principal components was here chosen so that the resulting reduced
variable explained a specific percentage of xt variance, here 20%, 40%, 60% and
80% of xt , respectively. The corresponding linear models are referred to as "PCR20 ",
"PCR40 ", "PCR60 " and "PCR80 ", respectively.
Partial Least Squares Regression (PLSR) PLS regression has been found
efficient for kinematic trajectory reconstruction from ECoG features in several
studies [Chao et al., 2010] [Shimoda et al., 2012] [Bundy et al., 2016]. This training
method is able to handle high dimensional and/or correlated explanatory variables,
especially when the explanatory variable’s dimension is higher than the number
of training samples. The PLS estimator is a shrinkage estimator [Lingjaerde and
Christophersen, 2000], i.e. it shrinks parameters towards zero. It thus exploits the
prior knowledge that most features are irrelevant. As a result, the PLS estimator is
biased but its variance is lower than the OLS variance.
PLS-based estimation of the parameters B of a linear model between xt and yt
relies on the assumption that the dependence between xt and yt can be satisfyingly
modelled by a linear model between low-dimensional projections of xt and yt . Let the
latent variables tt ∈ RF and ut ∈ RF be the projections ("scores") of xt and yt onto
the low-dimensional subspaces X̃ ⊂ Rm and Ỹ ⊂ Rn , with dim(X̃) = dim(Ỹ) = F . X̃
and Ỹ are found jointly, on the criterion that they maximize the covariance between
tt and ut . The Nonlinear Iterative Partial Least Squares (NIPALS) [Höskuldsson,
1988] or SIMPLS [de Jong, 1993] algorithms are generally used to perform PLS
regression.
The number of PLS factors used to reflect the dependence between the indepen-
dent and dependent variables is a hyperparameter which value has to be fixed before
performing PLS regression. Although the utilization of information criteria (e.g., the
Bayesian Information Criterion) has recently begun to be explored [Krämer, 2011],
B.1. Expert training procedure selection 189
the optimal dimension F , i.e. the optimal number of PLS factors, is typically found
using k-fold Cross-Validation-based criteria [Geladi and Kowalski, 1986] [Wold et al.,
2001] [Li et al., 2002]. For each considered number of factors f , a PLS model is
trained using (k-1) folds of the data. The corresponding PRESS (PRedicted Error
Sum of Squares) statistic is computed on the remaining fold [Li et al., 2002]. The
total PRESS is obtained by repeating these steps for all (k-1) fold combinations,
and by summing the corresponding fold-specific PRESS statistics [Li et al., 2002].
The utilization of several criteria has been proposed to infer the optimal number of
factors f from the total PRESS statistic. Wold’s R criterion was here applied on
+1)
the 6-fold cross-validated PRESS statistic. It is defined as RW old = PRESS(f
PRESS(f ) [Li
et al., 2002]. The chosen F corresponds to RW old > 1, i.e. the procedure is stopped
when the addition of a new factor increases the error. The corresponding model is
denoted by "PLS".
Penalized regression Penalized training has been used for kinematic decoding in
several BCI studies, in particular [Li et al., 2009] [Suminski et al., 2010] [Flamary and
Rakotomamonjy, 2012] [Shanechi et al., 2013] [Williams et al., 2013] [Willett et al.,
2013] [Spüler et al., 2016]. LASSO regression, which application has been reported
for ECoG kinematic decoding in [Williams et al., 2013] and [Spüler et al., 2016],
was here considered. It relies on ℓ1 -penalization to identify sparse linear models.
The amount λ of ℓ1 -penalization was here chosen by 6-fold cross-validation. LASSO
was chosen over Elastic Net (i.e., linear combination of ℓ1 and ℓ2 regularization)
training because the latter requires the estimation of two hyperparameters, namely
the amount of regularization and the balance between ℓ1 and ℓ2 regularization. Grid
search over two parameters being time-consuming, Elastic Net was deemed less
suited to BCI applications than LASSO training. The open-source glmnet toolbox
[Qian et al., 2013], which relies on an efficient coordinate descend optimization
procedure [Friedman et al., 2010], was used to perform LASSO training.
B.1.2 Results
Comparisons between the 6 training procedures (PCR20 , PCR40 , PCR60 , PCR80 ,

PLS and LASSO) were performed on the preclinical and clinical data sets. Perfor-
mances were assessed by computing the PCC (see section 5.2.3) between the true
and estimated limb positions during IC periods. The Friedman test was used to
assess the significance of the observed performance differences between classifiers
(see section 5.3). Post-hoc testing was carried out by applying the sign test to
the best decoder and each one of the other decoders. The significance level α was
corrected to account for the multiple pairwise comparisons which were completed,
here 5 pairwise comparisons. The corrected significance level αpost−hoc = 0.05
5 = 0.01
was thus utilized to measure the significance of the performance differences observed
between the best decoder and the other ones.
Table B.1: Preclinical data set, median PCC between reconstructed and true
trajectories, linear IC experts (median over 8 and 16 sessions for the subdural and
epidural preclinical data sets, respectively).
PCR20 PCR40 PCR60 PCR80 LASSO PLS

y1 0.26 0.28 0.36 0.24 0.42 0.43
subdural y2 0.51 0.64 0.70 0.63 0.73 0.74
y3 0.42 0.55 0.63 0.59 0.69 0.67
y1 0.08 0.07 0.09 0.10 0.11 0.09
epidural y2 0.20 0.36 0.38 0.36 0.46 0.45
y3 0.21 0.28 0.32 0.27 0.41 0.35
B.1.2.1 Preclinical data set

Reconstruction performances obtained on the preclinical and clinical data sets are
shown in Table B.1. Friedman-based decoder ranking is presented in Table B.2.
The Friedman test indicated that the 6 decoders did not behave equivalently. Post-
hoc pair-wise tests were therefore completed (see corresponding p-values in Table
B.2).
B.1.2.2 Clinical data set

Reconstruction performances obtained on the clinical data set are shown in Table
B.3. Friedman-based decoder ranking is presented in Table B.4. The Friedman test
indicated that the 6 considered linear experts did not behave equivalently. Post-hoc
pairwise tests were therefore completed (see corresponding p-values in Table B.4).
B.1. Expert training procedure selection 191
Table B.2: Preclinical data set, ranking of the linear IC experts’ training approaches.
The Friedman test was used to rank the continuous decoders on the basis of the
associated PCC (averaged over axes). Pairwise post-hoc tests with the corrected
significance level αpost−hoc = 0.01 were performed to assess the significance of the
performance differences between the best decoder and the other ones. Decoders
which exhibited a non-significant performance difference with the best decoder are
indicated with a bold font.
Subdural Epidural
Rank Decoder p-value Decoder p-value
1 LASSO LASSO
2 PLS 0.727 PLS 0.455
3 PCR60 0.008 PCR80 0.02
4 PCR40 0.008 PCR60 0.004
5 PCR80 0.008 PCR40 0.004
6 PCR20 0.008 PCR20 < 0.001
Table B.3: Clinical data set, median PCC between reconstructed and true trajectories,
linear IC experts (median over 9 sessions).
PCR20 PCR40 PCR60 PCR80 LASSO PLS

1 0.05 0.15 0.12 0.20 0.35 0.25
2 0.11 0.22 0.07 0.25 0.49 0.34
fingers 3 0.17 0.24 0.20 0.27 0.51 0.34
4 0.09 0.16 0.12 0.23 0.44 0.26
5 0.12 0.19 0.17 0.28 0.45 0.34
Table B.4: Clinical data set, ranking of the linear IC experts’ training approaches.
The Friedman test was used to rank the continuous decoders on the basis of the
associated PCC (averaged over fingers). Pairwise post-hoc tests with the corrected
significance level αpost−hoc = 0.01 were performed to assess the significance of the
performance differences between the best decoder and the other ones. Decoders
which exhibited a non-significant performance difference with the best decoder are
indicated with a bold font.
Rank Decoder p-value

1 LASSO
2 PLS 0.04
3 PCR80 0.004
4 PCR40 0.004
5 PCR60 0.004
6 PCR20 0.004
B.1.3 Discussion
PCR was found suboptimal when compared to LASSO and PLSR on both the
preclinical and clinical data sets (see Tables B.2 and B.4). The respective per-
formances of PCR and PLSR are consistent with the theoritical properties of the
PLSR, which has been shown to provide a closer fit than the Principal Compo-
nent Regression [Phatak and De Hoog, 2002]. PCR and PLSR performances are
additionally coherent with results reported in an earlier kinematic decoding study
completed with ECoG clinical data [Spüler et al., 2016]. In this study, the extraction
of principal components followed by ridge regression was surpassed by an analogue of
PLSR, namely Canonical-Correlation Analysis [Spüler et al., 2016]. In contrast with
the results obtained with LASSO linear models in this same study, LASSO models
here slightly surpassed PLSR on both data sets. The amount λ of penalization was
chosen with cross-validation. This fine tuning of λ may explain why LASSO models
were outperformed by Canonical-Correlation Analysis in [Spüler et al., 2016], where
λ was fixed at 0.1 rather than optimized for each acquisition session.
LASSO training is computationally more expensive than PLSR. The performance
difference between LASSO and PLSR experts was found to be insignificant on both
the preclinical and clinical data sets. This result may, however, reflect a lack of power
of the statistical test. High p-values were associated with the difference between
LASSO and PLSR on the preclinical data set (0.73 and 0.56 for the subdural and
epidural data sets, respectively), and LASSO experts only outperformed PLSR
experts by a median of 3% and 4% for the PCC on the subdural and epidural data
sets, respectively (PCC averaged over axes). Because of these similar decoding
performances and of its lower computational cost, PLSR was chosen over LASSO
training for fast identification of linear experts on the preclinical data set. The
B.2. Gate selection 193
p-value corresponding to the performance difference between LASSO and PLSR

was lower in the case of the clinical data set (p = 0.04), and was associated with a
median improvement of 35% of the PCC (PCC averaged over fingers). The fact that
the clinical ECoG electrodes were separated by an inter-electrode distance of 1cm
and were not exclusively measuring motor activity (see data description in Chapter
4) may explain this difference between LASSO models and PLSR. Because of this
sparse spatial sampling, it is probable that most electrodes do not contribute to the
encoding of each finger movement. While PLS models are supposed to assign low
weights to the features associated with such electrodes, the fact that these weights
are low rather than null may introduce undesirable noise in the kinematic estimates.
Because of the clear advantage of LASSO training over PLSR, it was decided to use
LASSO training on the clinical data set despite its higher computational cost.
B.2 Gate selection

The state-specific emission distributions associated with the HMM-based gate of the
MSLM can be modelled by means of generic distributions or by using an alternative
approach based on discriminative modelling, i.e. both generative and discriminative
state decoders can be embedded into the dynamic HMM-based gate of the MSLM
(see section 3.2.1). Their relevance may in particular depend on the distribution
of neural features within each state. If for example this distribution cannot be
satisfyingly modelled with a generic distribution, e.g. a Gaussian distribution, using
a discriminative approach to model the emission may permit to improve the gating
procedure. A study was therefore completed to measure the relevance of different
modelling approaches. For the sake of simplicity, this study was carried out on
static decoders, i.e. the ability of several generative and discriminative static state
decoders to distinguish between NC and (possible multiple) IC states was compared
on the preclinical and clinical data sets considered in the present doctoral work. The
best static decoders were chosen to be subsequently integrated into the MSLM’s
dynamic gate.
As BCI classifier performances have been found to depend on the dimension
of the input variable [Bhattacharyya et al., 2011], the considered classifiers were
tested after different projection-based dimensionality reduction procedures had been
performed on the input variable.
B.2.1 Methods and implementation

5 classifiers, namely LDA, QDA, LR, linear and non-linear SVM, were selected on
the basis of their frequent utilization in BCI studies and/or on the high decoding
performance they reached in previous studies. LDA and QDA are generative
classifiers based on Gaussian distributions (see Chapter 2). They are thus static
analogues of generic HMMs with Gaussian emission distributions. By contrast,
LR and SVMs are discriminative classifiers (see Chapter 2). As a result, they
correspond to static analogues of the HMMs which exploit a discriminative approach
to model the emission probabilities. They were combined with unsupervised and
supervised dimensionality reduction steps, and tested for the task of binary and
multi-class state detection on the preclinical and clinical data sets, respectively.
B.2.1.1 Binary and multi-class state decoders
LDA and QDA ML training of the parameters {µi , Σi } associated with the
multivariate Gaussian distributions P (xt |z t = i) = N (µi , Σi ), where i = 1, 2 for the
preclinical data set and i = 1, . . . , 6 for the clinical data set, was performed on the
training samples xt observed at times t such that z t = i. The LDA covariance matrix
Σ = Σi = Σj , i = Ó j being shared by all classes [Friedman et al., 2001], it was fitted
using the totality of the training samples. Equiprobable class prior probabilities
P (z t = i) = P (z t = j) were utilized for both the LDA and QDA classifiers.
Linear and nonlinear SVMs Following [Bhattacharyya et al., 2011] and [Bashashati
et al., 2015], the nonlinear SVM was built using a Radial Basis Function (RBF) kernel.
The one-against-one strategy was utilized to perform multi-class SVM classification
because it has been reported as generally more efficient that the one-against-all one
[Hsu and Lin, 2002] and has been used in several BCI studies [Bashashati et al.,
2015]. One SVM was thus trained for each possible state pair i, j, i = Ó j, and an
Error-Correcting Output Code was used to combine the state estimates yielded
by each binary classifier [Dietterich and Bakiri, 1995]. Each class is associated
with a specific combination of the SVM outputs. During application, the output
of each SVM is computed for the considered observation sample. The class which
corresponds to the combination of SVM outputs which is the closest to the one
obtained on the observation features (i.e., the one which minimize the number of
different outputs) is attributed to the input variable [Dietterich and Bakiri, 1995].
Logistic Regression A LR model and its multi-class extension, namely a multi-

nomial LR, were fitted on each session of the preclinical and clinical data sets. The
ML estimator of LR parameters is known to be unstable when it is applied to
a training data set with linearly separable classes [Bishop, 2006]. Penalized LR
training has been found to be efficient in several BCI studies [Gouy-Pailler et al.,
2009] [Ryali et al., 2010] [Bashashati et al., 2015] [Bundy et al., 2016]. Two training
approaches were thus compared for the identification of LR and multinomial LR
models: ML [Czepiel, 2012] and LASSO training [Friedman et al., 2010]. LASSO
training consists in introducing ℓ1 -penalization to stabilize parameter estimation
[Friedman et al., 2001]. It was again chosen over Elastic Net penalization to limit the
number of hyperparameters (see section B.1.1). The amount λ of ℓ1 regularization
was here found by 6-fold cross-validation.
B.2.1.2 Dimensionality reduction
BCI classifier performances depend on the characteristics of the input variable

[Bashashati et al., 2015], in particular on its dimensionality [Bhattacharyya et al.,
2011]. Supervised and unsupervised projection methods were compared for the
reduction of the dimensionality of the input variable xt . Thus, the above-mentioned
classifiers were trained and compared after 6 dimensionality reduction procedures
had been applied on the input variable.
PCA-based dimensionality reduction PCA-based dimensionality reduction is

frequently performed in BCI studies [Wang et al., 2009b][Ke and Li, 2009] [Suk and
Lee, 2010] [Argunşah and Çetin, 2010] [Bhattacharyya et al., 2011]. PCA-reduced
input variables of different sizes were here considered. Following [Suk and Lee, 2010],
the number of extracted principal components was chosen so that the resulting
reduced variable explained a specific percentage of xt variance, here 20%, 40%, 60%
and 80% of xt , respectively. The corresponding reduced variables are henceforth
referred to as "PCA20 ", "PCA40 ", "PCA60 " and "PCA80 ", respectively.
PLS-based dimensionality reduction The PLS algorithm has been designed

for the identification of regression models. As such, it has been profitably applied
for kinematic decoding in several BCI studies [Shimoda et al., 2012] [Eliseyev and
Aksenova, 2014] [Eliseyev and Aksenova, 2016] [Bundy et al., 2016] [van Gerven
et al., 2012] (see Chapter 2). PLS-based dimensionality reduction combined with
classification has nevertheless been shown to be efficient for the decoding of discrete
variables [Kemsley, 1996] [Nguyen and Rocke, 2002] [Barker and Rayens, 2003]
[Turkmen and Billor, 2012]. Satisfying classification accuracies have been reported
when such combinations have been utilized in BCI studies [Eliseyev et al., 2011]
[Eliseyev et al., 2012].
PLS-based projectors were therefore fitted to perform supervised dimensionality
reduction before applying the considered generic classifiers. The first F PLS-scores
of xt , where the state dummy variable zt was regressed against the neural features
xt [Rosipal and Krämer, 2006], were fed to the classifiers. The optimal number of
PLS factors was found using Wold’s R criterion with 6-fold cross-validation on the
training data set. The resulting reduced input variables are denoted by PLS1 in the
following sections.
The PLS1 reduced input variables and the corresponding classifiers were fitted
on the same data sets. Another training strategy consists in using independent data
sets to train the PLS projectors and the classifiers, thereby reducing the risk of
overfitting. A reduced variable denoted by "PLS2 " was therefore built by training a
PLS model on the first 40% of the training data set. The second part of the training
data set (60%) was used to fit the classifiers.
Finally, the full input variable xt is referred to as "full" in the remaining of the
present appendix.
Table B.5: Preclinical data set, average dimension of the feature subsets. The
average was computed over 5, 3, 6 and 10 acquisition sessions for Monkeys A, K, B
and C, respectively.
PCA20 PCA40 PCA60 PCA80 PLS1 PLS2 full

NHP A 7 ± 3 49 ± 12 231 ± 53 751 ± 105 6±1 5 ± 1 12160 ± 0
NHP K 8 ± 3 56 ± 22 268 ± 133 988 ± 395 8±1 8 ± 1 24747 ± 0
NHP B 2 ± 1 12 ± 4 63 ± 22 332 ± 84 10 ± 1 9 ± 6 24320 ± 0
NHP C 3 ± 2 16 ± 9 61 ± 35 273 ± 196 8±1 7 ± 3 24320 ± 0
B.2.2 Results
P +T N
The classification accuracy ACC = T P +TTN +F P +F N was used to assess the perfor-
mance of each combination reduced feature / classifier for NC/IC discrimination.
In the case of the clinical data set, it was computed after having pooled the active
states ICi , i = 1, . . . , 5 into a global active state IC. The ratio of correctly classified
samples was additionally computed on TP observations to measure the ability of
each decoder to distinguish between fingers. The Friedman test and post-hoc tests
were used to assess the significance of the observed performance differences between
classifiers (see Chapter 5 for details on the validation methodology).
B.2.2.1 Preclinical data set
The number of extracted PCA components and PLS factors varied from one data set
to another. Table B.5 summarizes the average dimensionality of each feature subset.
Some dimensionality reduction/classifier pairs were deemed irrelevant because of
the inadequacy of the classifier for high-dimensional input variables, or for deterrent
training durations. Generic LDA and QDA are for example known to be ill-suited for
high-dimensional feature classification, in particular because of difficulties to estimate
high-dimensional covariance matrices [Friedman, 1989]. Similarly, instabilities of
the ML estimates of LR parameters arise in high-dimensional settings [Zhang et al.,
2007]. LDA, QDA and ML LR were thus trained exclusively when the reduced
variables exhibited limited sizes.
Classification performances obtained on the subdurally and epidurally implanted
monkeys are shown in Table B.6. Discarded dimensionality reduction/classifier
pairs are indicated by grey cells. The Friedman test indicated that all pairs did not
perform similarly. Friedman-based classifier ranking is presented in Table B.7.
B.2.2.2 Clinical data set
The number of extracted PCA components and PLS factors is displayed in Table
B.8. Average performance indicators are shown in Table B.9. Similarly to the
Table B.6: Preclinical data set, binary classification accuracy (ACC). The classifi-
cation accuracy achieved by each couple reduced feature / classifier was averaged
over the subdural and epidural acquisition sessions (8 and 16 sessions, respectively).
Grey cells indicate dimensionality reduction/classifier pairs which were discarded
because the classifier was ill-suited for the corresponding feature dimensionality.
ACC
LDA 0.71 0.87 0.85
QDA 0.79 0.91 0.89
ML-LR 0.86 0.92 0.92
sub. LASSO-LR 0.85 0.90 0.92 0.92 0.92 0.92 0.92
linear-SVM 0.86 0.90 0.91 0.90 0.92 0.92 0.90
rbf-SVM 0.84 0.68 0.68 0.68 0.91 0.89 0.64
LDA 0.56 0.79 0.76
QDA 0.58 0.80 0.76
ML-LR 0.64 0.79 0.77
epi. LASSO-LR 0.64 0.71 0.74 0.78 0.80 0.77 0.81
linear-SVM 0.64 0.71 0.74 0.77 0.79 0.77 0.77
rbf-SVM 0.64 0.60 0.55 0.55 0.76 0.72 0.55
Table B.7: Preclinical data set, ranking of the 6 most efficient couples reduced feature
/ classifier. The Friedman test was used to rank couples reduced feature / classifier.
Pairwise post-hoc tests with the corrected significance level αpost−hoc = 0.05
5 = 0.01
were performed to assess the significance of the performance differences between
the best decoder and the 5 other ones. Decoders which exhibited a non-significant
performance difference with the best decoder are indicated with a bold font.
Subdural Epidural
Rank decoder p-value decoder p-value
1 PLS1 / LASSO LR full / LASSO LR
2 PLS1 / ML LR 1 PLS1 / ML LR 0.02
3 PLS1 / linear SVM 1 PLS1 / QDA 0.02
4 PCA80 / LASSO LR 0.73 PLS1 / LASSO LR 0.02
5 PLS2 / linear SVM 0.29 PLS1 / linear SVM 0.02
6 full / LASSO LR 0.008 PLS1 / LDA 0.02
Table B.8: Clinical data set, average dimension of the feature subsets (9 acquisition
sessions).

24 ± 17 209 ± 83 631 ± 135 1393 ± 182 18 ± 6 7±1 21913 ± 3070
Table B.9: Clinical data set, ACC for NC/IC detection and ratio of correctly
classified TP (Correct TP Ratio, CTPR). Both indicators were computed for each
couple reduced feature / classifier, and averaged over the acquisition sessions (9
sessions). Grey cell are associated with the CTPR of some couples reduced feature
/ rbf SVM, because the SVM failed to output TP on some data sets and that the
CTPR was consequently not defined.

LDA 0.71 0.82 0.75
QDA 0.74 0.85 0.80
ML-LR 0.65 0.82 0.80
ACC LASSO-LR 0.64 0.82 0.83 0.83 0.83 0.83 0.86
linear-SVM 0.65 0.81 0.83 0.83 0.73 0.71 0.83
rbf-SVM 0.36 0.36 0.36 0.36 0.80 0.78 0.36
LDA 0.29 0.56 0.29
QDA 0.26 0.56 0.28
CTPR ML-LR 0.27 0.51 0.47
LASSO-LR 0.29 0.49 0.51 0.52 0.55 0.50 0.60
linear-SVM 0.20 0.47 0.48 0.50 0.46 0.30 0.51
rbf-SVM 0.52 0.36
preclinical data set, irrelevant feature subsets / classifier pairs are indicated by grey
cells. Ranking of the feature subset / classifier pairs is presented in Table B.10.
Table B.10: Clinical data set, ranking of the 6 most efficient couples reduced feature
/ classifier. The Friedman test was used to rank couples reduced feature / classifier.
Pairwise post-hoc tests with the corrected significance level αpost−hoc = 0.05
5 = 0.01
were performed to assess the significance of the performance differences between
the best decoder and the 5 other ones. Decoders which exhibited a non-significant
performance difference with the best decoder are indicated with a bold font.
ACC CTPR
Rank decoder p-value decoder p-value
1 full / LASSO LR full / LASSO LR
2 PLS1 / QDA 0.51 PLS1 / LDA 0.04
3 PCA60 / LASSO LR 0.04 PLS1 / QDA 0.51
4 full / linear SVM 0.004 PLS1 / LASSO LR 0.004
5 PCA60 / LASSO LR 0.04 PLS1 / rbf SVM 0.04
6 PLS2 / LASSO LR 0.04 PLS1 / ML LR 0.004
B.2.3 Discussion
The application of generic classifiers on PLS-reduced variables permitted to achieve
high decoding performances for both the subdural and epidural acquisition sessions
of the preclinical data set. In particular, it was observed that classifiers fed with
PCA components either fell behind classifiers fed with PLS factors, or equated these
classifiers. In the latter case, a higher number of PCA components was required to
achieve the same performance as PLS-fed classifiers. These results are consistent
with both theoretical [Barker and Rayens, 2003] and practical [Kemsley, 1996]
[Nguyen and Rocke, 2002] studies previously completed outside the BCI community.
Fitting both PLS regression and classifiers on the same data set was found to yield
better results than separate training. The number of training samples for both the
PLS models and classifiers is reduced when training is performed independently.
This diminution may explain the poor decoding performance observed in the case
of independent training. Satisfying results were obtained with high-dimensional
variables when they were utilized in combination with a penalization approach (i.e.,
LR training with a LASSO penalization).
While linear SVMs were here found to generally outperform nonlinear ones on
the preclinical data set, LDA was surpassed by QDA. These results are consistent
with earlier reports on the variable respective relevance of nonlinear and linear
classifiers for BCI decoding (see Chapter 2).
Generative models (LDA and QDA) were mainly outperformed by discriminant
ones on the preclinica data set, e.g. LR or linear SVM. This observation suggests that
the distribution of neural features within NC and IC states cannot be satisfyingly
modelled for the considered preclinical data set with generic Gaussian distributions.
The modelling of the MSLM’s HMM-based gate emission probabilities is thus likely
to benefit from the utilization of alternative approaches based on discriminative mod-

elling (see section 3.2.1). ML LR fed with PLS-based features was here chosen for
discriminative modelling of the MSLM emission on the preclinical data set because
it exhibited a high decoding accuracy on both the subdural and epidural acquisition
sessions. Its training has the additional advantage of being computationally less
expensive that LASSO training of a LR on the full feature set.
The interest of PLS-based dimensionality reduction was further observed on

the clinical data set. As, however, LASSO-trained LR corresponded to CTPRs in
average 9% higher than the ones associated with LDA or QDA fed with PLS-based
features, it appears that PLS-based dimensionality reduction may be suboptimal
for the multi-class problem at hand, or that discriminative modelling of the MSLM
gate emission probabilities may outperform traditional generative modelling with
generic Gaussian distributions for the considered clinical data set. LASSO LR was
thus chosen to model the MSLM gate emission probabilities on the clinical data set
despite its higher computational cost.
Appendix C
Switching Kalman Filtering for

asynchronous ECoG decoding
Contents
C.1 Introduction: Kalman filter . . . . . . . . . . . . . . . . . . . 201
C.2 Switching Kalman Filter . . . . . . . . . . . . . . . . . . . . . 202
C.2.1 Switching State-Space Model . . . . . . . . . . . . . . . . . . 202
C.2.2 Training . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 203
C.2.3 Application . . . . . . . . . . . . . . . . . . . . . . . . . . . . 203
C.3 Optimization of Kalman Filtering for high-dimensional ECoG
signals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 206
C.3.1 Methods and implementation . . . . . . . . . . . . . . . . . . . 207
C.3.2 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 208
C.3.3 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 208
C.1 Introduction: Kalman filter

The Kalman Filter (KF) is an algorithm which permits to infer a hidden trajectory yt
from noisy measurements xt ∈ Rm [Bishop, 2006] (see Figure C.1). The continuous
response (state) variable yt ∈ Rn is composed by the trajectory coordinates and
derivatives (velocity, acceleration, etc.). The KF applies to linear Gaussian state-
space models [Bishop, 2006]. The following state-space model is generally considered
in BCI studies [Wu et al., 2002] [Pistohl et al., 2008] [Wang et al., 2013b]:
yt+1 = Ayt + wt , (C.1)

xt = Cyt + vt (C.2)
Here, the emission (C.2) and transition (C.1) models are linear with constant
emission and transition matrices (C ∈ Rm×n and A ∈ Rn×n , respectively). The
noise processes wt and vt are independent and identically distributed sequences
of random variables [Krishnamurthy, 2016] with P (wt ) ∼ N (0, Γ), Γ ∈ Rn×n and
P (vt ) ∼ N (0, Σ), Σ ∈ Rm×m . The initial state is characterized by y1 = ŷ0 + u0 ,
with P (u0 ) ∼ N (0, P0 ).
After training, typically performed using Ordinary Least Squares [Wu et al.,
2002], the KF issues the estimate ŷt = E(yt |x1:t ) (see algorithm 1).
Appendix C. Switching Kalman Filtering for asynchronous ECoG
202 decoding
Figure C.1: KF structure. When the KF is used for ECoG filtering, the hidden
trajectory yt corresponds to the effector movement intended by the user, while the
measurement xt refers to the ECoG features.
Algorithm 1 Kalman Filter [Welch and Bishop, 2006]

Notations:
ŷt = E(yt |x1:t ) (C.3)

t t t
P = Var(ŷ − y ) (C.4)
t
ŷprior = E(yt |x1:t−1 ) (C.5)
Ptprior = t
Var(ŷprior − y ).t
(C.6)
Initialization: ŷ0 , P0 .
Time update:
t
ŷprior = Aŷt−1 (C.7)
Ptprior = AP t−1 T
A + Γ. (C.8)
Measurement update:
Kt = Ptprior CT (CPtprior CT + Σ)−1 (C.9)

t t t t T t
ŷ = ŷprior + K (x − C ŷprior ) (C.10)
t t
P = (I − K C)Ptprior . (C.11)
C.2 Switching Kalman Filter

C.2.1 Switching State-Space Model
The Switching Kalman Filter is a hybrid decoder which probabilistically combines
K Kalman Filters [Murphy, 1998]. It extends the KF for hybrid discrete/continuous
decoding.
Emission and transition distributions are conditioned on a switching variable
z ∈ Z, with Z = {z1 , z2 . . . , zK } ⊂ N. z t is not observed, and is assumed to be
t
generated by a first-order Markov chain:

C.2. Switching Kalman Filter 203
z t+1 = Aswitch z t (C.12)

yt+1 = Az t yt + wzt t (C.13)
t t
x = Cz t y + vzt t , (C.14)
where Aswitch is the switching transition matrix. Similarly to the matrices Ak , Ck ,

Γk and Σk , the probability of the initial state y1 is conditioned on z 1 : y1 = ŷz01 +uz 1 ,
with P (uz 1 ) ∼ N (0, P0z 1 ). The initial probabilities P (z 1 = zk ), k = 1 . . . K of the
switching variable are gathered in π ∈ RK . The dependencies between variables are
illustrated in Figure C.2.
Figure C.2: SKF structure. When the SKF is used for ECoG filtering, the hidden
trajectory yt corresponds to the effector movement intended by the user, the hidden
switching variable z t to the neural state (for example, z t = 0 and z t = 1 for NC and
IC states, respectively) and the measurement xt to the ECoG features.
C.2.2 Training
Expectation-Maximization formula for unsupervised Maximum-Likelihood training
of Switching Kalman Filters are presented in [Murphy, 1998]. Supervised Maximum-
Likelihood training is possible when a complete training data set {X, Y, z} is
available. Maximization formula [Murphy, 1998] for EM-based SKF training are
applied, taking into account that the hidden switching state sequence is known. ML
estimates of Ak , Ck , Γk and Σk are computed on the training data sets {Xk , Yk }.
The switching transition matrix Aswitch is fitted on the basis of transition frequencies
in the sequence (z t )Tt=1 .
C.2.3 Application
The algorithms 2 and 3 gather inference formula which permit to perform Switch-
ing Kalman Filtering. In contrast with Kalman filtering, exact estimation of
ŷt = E(yt |x1:t ) is intractable because of the exponential number of possible state
sequences. The SKF therefore relies on approximate solutions [Murphy, 1998]. The
204 decoding
collapsing routine exposed in [Murphy, 1998] and implemented in [Quinn, 2005] is

here presented.
Figure C.3: SKF application.

C.2. Switching Kalman Filter 205
Algorithm 2 Switching Kalman Filter [Murphy, 1998]

Notations:
t|τ
Pj = Cov(yt |x1:τ , z t = j)
t,t−1|τ
Pj = Cov(yt , yt−1 |x1:τ , z t = j)
t,t−1|τ
Pi(j) = Cov(yt , yt−1 |x1:τ , z t−1 = i, z t = j)
M t−1,t|τ (i, j) = P (z t−1 = i, z t = j|x1:τ )
M t|τ (j) = P (z t = j|x1:τ )
Ltj = P (xt |x1:t−1 , z t = j)
Initialization: π, ŷi0 , P0i , i = 1, . . . , K.,
Iterations:
t|t t|t t,t−1|t t−1|t−1 t−1|t−1

(yi(j) , Pi(j) , Pi(j) , Lti(j) ) = filter(yi , Pi , xt ; Aj , Cj , Γj , Σj ) (C.15)
M t−1,t|t (i, j) = P (z t−1 = i, z t = j|x1:t ) (C.16)
Lt (i, j)Z(i, j)M t−1|t−1 (i))
=q q t t−1|t−1 (i)
(C.17)
i j L (i, j)Z(i, j)M
M t|t (j) = M t−1,t|t (i, j)

Ø
(C.18)
i
M t−1,t|t(i,j)
Wi|j = P (z t−1 = i|z t = j, x1:t ) = (C.19)
M t|t(j)
t|t t|t t|t t t|t
(xj , Pj ) = Collapse(yi(j) , Pi|j , Wi|j ) (C.20)
206 decoding
Algorithm 3 Filter operator [Murphy, 1998]:

(yt|t , Pt|t , Pt,t−1|t , Lt ) = filter(yt−1|t−1 , Pt−1|t−1 , xt ; A, C, Γ, Σ)
Prediction:
yt|t−1 = Ayt−1|t−1
Pt|t−1 = APt−1|t−1 A′ + Γ
Innovation:
et = xt − Cyt|t−1
St = CPt|t−1 C′ + Σ
Kt = Pt|t−1 C′ (St )−1
Lt = N (e; 0, St )
Updated estimates:
yt|t = yt|t−1 + Kt et
Pt|t = (I − Kt C)Pt|t−1 = Pt|t−1 − Kt St (Kt )′
Pt,t−1|t = (I − Kt C)APt−1|t−1
Details on the collapsing operator used to approximate mixtures of Gaussians and

mentioned in equation (C.20) can be found in [Murphy, 1998].
C.3 Optimization of Kalman Filtering for high-dimensional

ECoG signals
The SKF exploits Kalman filters to reconstruct kinematic parameters during IC

periods. The accuracy of the Kalman filters can be improved by choosing an optimal
composition for the continuous state variable [Pistohl et al., 2008] and by introducing
a relevant lag between observed neural features and resulting movements in the
emission model [Wu et al., 2002]. The computation of each KF’s update involves the
inversion of a m×m matrix (see section C.1), where m is the dimension of the neural
features. Matrix inversion is liable to be time-consuming when high-dimensional
features are fed to the KF. The reduction of the neural features’ dimension may thus
be additionally required to make KF suitable for online application. A preliminary
study was completed to select relevant state variable compositions, lags between
neural features and kinematic parameters and dimensionality reduction procedures
for efficient SKF-based ECoG decoding on the considered preclinical and clinical
data sets.
C.3. Optimization of Kalman Filtering for high-dimensional ECoG
signals 207
C.3.1 Methods and implementation

C.3.1.1 Optimal lag
Because causal neural patterns are likely to precede the resulting limb movements
with a specific time lag, the accuracy of KF estimates is liable to be improved by
introducing a relevant lag between the observations and the corresponding hidden
trajectory in the emission model. Uniform lags, i.e. lags which are similar for all
neural features [Wu et al., 2002], were here investigated. 2 different lags were tested.
Kalman filters were fed with features either associated with the last 100ms before the
instant t considered for trajectory estimation (10th time bin of the generic feature xt
presented in Chapter 6) or with features observed between t-200 to t-100ms (9th
time bin of the generic feature xt ). The corresponding features are denoted by xKF t .
C.3.1.2 Composition of the state variable
Following earlier ECoG decoding studies [Pistohl et al., 2008] [Wang et al., 2013b]
[Marathe and Taylor, 2013], a first continuous state variable yKFt was composed of
the wrist’s or fingers’ position and velocity. An alternative state variable composed
of position only was additionally considered for comparison purposes.
C.3.1.3 Dimensionality reduction
The recourse to projection approaches such as PCA [Wu et al., 2003b] [Kao et al.,
2013] or Factor Analysis [Sadtler, 2014] and to feature selection strategies [Malik
et al., 2015] has been reported for Bayesian filtering in BCI studies. Projection-based
methods were here considered because of their frequent application [Wu et al., 2003b]
[Kao et al., 2013] [Sadtler, 2014] and of their simple implementation.
PCA Following [Wu et al., 2003b] [Kao et al., 2013], PCA-based dimensionality
was here considered. The number of extracted principal components was chosen so
that the resulting reduced variable explained a specific percentage of xKF
t variance,
here 20%, 40% and 60% of xKF , respectively. Higher percentages were not considered
t
because they often corresponded to a high number of principal components, and

were therefore ill-suited for online application of the KF. The reduced independent
variables are referred to as "PCA20 ", "PCA40 " and "PCA60 ", respectively.
PLS An alternative PLS-based supervised dimensionality reduction was addition-

ally considered. PLS regression between xKF
t and yKFt [Rosipal and Krämer, 2006]
was used to identify an informative low-dimensional subspace, which dimension was
chosen by applying Wold’s R criterion on the 6-fold cross-validated PRESS statistic
[Li et al., 2002].
208 decoding
Table C.1: Preclinical data set, median PCC between reconstructed and true
trajectories. Kalman filters were fed with neural features either associated with the
last 100ms before the considered instant t (10th time bin of the generic feature xt )
or with features observed between t-200 to t-100ms (9th time bin of the generic
feature xt ) (median over 8 and 16 sessions for the subdural and epidural data sets,
respectively). The PCCs were averaged over the 3 considered axes within each
acquisition session.
Position Position and velocity

neural features PCA20 PCA40 PCA60 PLS PCA20 PCA40 PCA60 PLS
sub. t-100:t (ms) 0.22 0.40 0.59 0.64 0.38 0.52 0.64 0.66
t-200:t-100 (ms) 0.23 0.39 0.56 0.64 0.37 0.49 0.59 0.64
epi. t-100:t (ms) 0.16 0.18 0.24 0.36 0.35 0.35 0.39 0.47
t-200:t-100 (ms) 0.16 0.18 0.22 0.35 0.34 0.36 0.39 0.44
OLS estimates of the transition and emission matrices were then computed using
the reduced neural features (observations) and the position/velocity of the limbs of
interest. Similarly, ML estimates of the emission and transition covariance matrices
were computed following [Aggarwal et al., 2013].
C.3.2 Results
Comparisons between the different lags, state variable compositions and dimension-
ality reduction procedures were performed on the preclinical and clinical data sets
presented in Chapter 4. Performance was assessed by computing the PCC (see
section 5.2.3) between the true and estimated limb positions during IC periods.
Importantly, the KFs were re-initialized using the tracked position and/or velocity
at the beginning of each IC (or ICi ) period.
C.3.2.1 Preclinical data set
KF reconstruction performances over IC periods are displayed in Table C.1.
C.3.2.2 Clinical data set
KF reconstruction performances over IC periods are displayed in Table C.2.
C.3.3 Discussion
Continuous state variables composed of limb position and velocity were associated
with higher PCC between true and estimated trajectories than continuous state
variables only constituted by limb positions. These results are consistent with
C.3. Optimization of Kalman Filtering for high-dimensional ECoG
signals 209
Table C.2: Clinical data set, median PCC between reconstructed and true trajectories.
Kalman filters were fed with features associated either with the neural signals elicited
during the last 100ms before the instant t considered for the continuous state variable
t
yKF (10th time bin of the generic feature xt ) or with features observed between
t-200 and t-100ms (9th time bin of the generic feature xt ) . The PCCs were averaged
over the 5 considered fingers within each acquisition session.
Position Position and velocity

neural features xt PCA20 PCA40 PCA60 PLS PCA20 PCA40 PCA60 PLS
t-100:t (ms) 0.28 0.32 0.38 0.41 0.35 0.42 0.45 0.52
t-200:t-100 (ms) 0.24 0.27 0.26 0.32 0.33 0.40 0.36 0.40
findings reported in earlier ECoG studies [Pistohl et al., 2008]. Additionally, the
best decoding accuracy was obtained when the KF was fed with the neural features
corresponding to the neural activity generated during the last 100ms before the
considered instant t (10th time bin of the generic feature xt ). High numbers of
PCA components (i.e., 60% of explained variance) led to improved KF decoding
accuracies when optimal continuous state variable and lag were considered. A similar
observation has been reported in [Kao et al., 2013], where feeding a KF with principal
components accounting for 60% of MUA/SUA features permitted to achieve an online
accuracy similar to the one obtained by using all the neural features. PLS-based
dimensionality reduction nevertheless yielded the best KF IC decoding accuracy.
Thus, somehow counter-intuitively, the application of a discriminant dimensionality
reduction technique proved to be efficient for a subsequent application of a generative
model. Although a better optimization of the number of principal components may
permit to improve the accuracy of PCA-based KF, such procedure is likely to be
time-consuming and to correspond to a less compact feature representation than
the one obtained after PLS-based dimensionality reduction.
Importantly, the high decoding performance exhibited by Kalman filters is
specific to the trial-per-trial analysis which was adopted in the present preliminary
study, i.e. to that fact that the KF were correctly initialized at the beginning of
each movement. In particular, Kalman Filtering has been found to suffer from drift
effects in earlier BCI studies (e.g., [Hochberg et al., 2012]), and a degradation of the
KF decoding accuracy is likely to be observed in real-time settings where positions
and velocity are not correctly re-initialized at the beginning of each movement.
Bibliography
[Acharya et al., 2010] Acharya, S., Fifer, M. S., Benz, H. L., Crone, N. E., and
Thakor, N. V. (2010). Electrocorticographic amplitude predicts finger posi-
tions during slow grasping motions of the hand. Journal of neural engineering,
7(4):046002. (Cited on pages 31 and 46.)
[Aflalo et al., 2015] Aflalo, T., Kellis, S., Klaes, C., Lee, B., Shi, Y., Pejsa, K.,
Shanfield, K., Hayes-Jackson, S., Aisen, M., Heck, C., Liu, C., and Andersen, R.
(2015). Decoding motor imagery from the posterior parietal cortex of a tetraplegic
human. Science, 348(6237):906–910. (Cited on pages 5 and 12.)
[Aggarwal et al., 2008] Aggarwal, V., Acharya, S., Tenore, F., Shin, H. C., Etienne-
Cummings, R., Schieber, M. H., and Thakor, N. V. (2008). Asynchronous decoding
of dexterous finger movements using M1 neurons. IEEE Transactions on Neural
Systems and Rehabilitation Engineering, 16(1):3–14. (Cited on page 47.)
[Aggarwal et al., 2013] Aggarwal, V., Mollazadeh, M., Davidson, A. G., Schieber,
M. H., and Thakor, N. V. (2013). State-based decoding of hand and finger
kinematics using neuronal ensemble and LFP activity during dexterous reach-
to-grasp movements. Journal of neurophysiology, 109(12):3067–81. (Cited on
pages 42, 52, 62, 65, 66, 120, 125 and 208.)
[Aggarwal et al., 2009] Aggarwal, V., Tenore, F., Acharya, Soumyadipta Schieber,
M. H., and Thakor, V. N. (2009). Cortical decoding of individual finger and
wrist kinematics for an upper-limb neuroprosthesis. In Engineering in Medicine
and Biology Society, 2009. EMBC 2009. Annual International Conference of the
IEEE, pages 4535–4538. (Cited on page 63.)
[Ajiboye et al., 2012] Ajiboye, a. B., Simeral, J. D., Donoghue, J. P., Hochberg, L. R.,
and Kirsch, R. F. (2012). Prediction of imagined single-joint movements in a
person with high-level tetraplegia. IEEE Transactions on Biomedical Engineering,
59(10):2755–2765. (Cited on page 46.)
[Alessandri et al., 2008] Alessandri, A., Baglietto, M., and Battistelli, G. (2008).
Maximum-likelihood Kalman Filtering for Switching Discrete-time Linear Systems.
In Proceedings of the 47th IEEE Conference on Decision and Control, pages 3192–
3198. (Cited on page 70.)
[Amiri et al., 2013] Amiri, S., Rabbi, A., Azinfar, L., and Fazel-Rezai, R. (2013).
A Review of P300, SSVEP, and Hybrid P300/SSVEP Brain- Computer Inter-
face Systems. Brain-Computer Interface Systems - Recent Progress and Future
Prospects, 2013:1–8. (Cited on pages 2 and 3.)
212 Bibliography
[Anderson and Sijercic, 1996] Anderson, C. and Sijercic, Z. (1996). Classification of

EEG signals from four subjects during five mental tasks. In Solving engineering
problems with neural networks: proceedings of the conference on engineering
applications in neural networks (EANN’96), pages 407–414. (Cited on pages 13
and 14.)
[Anderson et al., 2012] Anderson, N. R., Blakely, T., Schalk, G., Leuthardt, E. C.,
and Moran, D. W. (2012). Electrocorticographic (ECoG) correlates of human
arm movements. Experimental brain research, 223(1):1–10. (Cited on pages 18,
23, 24, 27, 28, 39, 44 and 115.)
[Argunşah and Çetin, 2010] Argunşah, a. Ö. and Çetin, M. (2010). AR-PCA-HMM

approach for sensorimotor task classification in EEG-based brain-computer in-
terfaces. Proceedings - International Conference on Pattern Recognition, pages
113–116. (Cited on pages 44, 47, 56, 57 and 195.)
[Ashmore et al., 2012] Ashmore, R. C., Endler, B. M., Smalianchuk, I., Degenhart,
A. D., Hatsopoulos, N. G., Tyler-Kabara, E. C., Batista, A. P., and Wang, W.
(2012). Stable online control of an electrocorticographic brain-computer inter-
face using a static decoder. Conference proceedings : ... Annual International
Conference of the IEEE Engineering in Medicine and Biology Society. IEEE
Engineering in Medicine and Biology Society. Annual Conference, 2012:1740–4.
(Cited on pages 17, 18, 26 and 44.)
[Bai et al., 2011] Bai, O., Rathi, V., Lin, P., Huang, D., Battapady, H., Fei, D.-Y.,
Schneider, L., Houdayer, E., Chen, X., and Hallett, M. (2011). Prediction of human
voluntary movement before it occurs. Clinical Neurophysiology, 122(2):364–372.
(Cited on page 53.)
[Ball et al., 2009a] Ball, T., Kern, M., Mutschler, I., Aertsen, A., and Schulze-
Bonhage, A. (2009a). Signal quality of simultaneously recorded invasive and
non-invasive EEG. NeuroImage, 46(3):708–16. (Cited on pages 18 and 39.)
[Ball et al., 2009b] Ball, T., Schulze-Bonhage, A., Aertsen, A., and Mehring, C.
(2009b). Differential representation of arm movement direction in relation to
cortical anatomy and function. Journal of neural engineering, 6(1):016006. (Cited
on pages 23, 24, 27, 42, 43 and 173.)
[Barachant et al., 2010] Barachant, A., Bonnet, S., Congedo, M., and Jutten, C.
(2010). Riemannian geometry applied to BCI classification. Lecture Notes in
Computer Science (including subseries Lecture Notes in Artificial Intelligence and
Lecture Notes in Bioinformatics), 6365 LNCS:629–636. (Cited on page 53.)
[Barachant et al., 2012] Barachant, A., Bonnet, S., Congedo, M., and Jutten, C.
(2012). Multiclass Brain-Computer Interface Classification by Riemannian Ge-
ometry. IEEE Transactions on Biomedical Engineering, 59:920–928. (Cited on
page 53.)
Bibliography 213
[Barker and Rayens, 2003] Barker, M. and Rayens, W. (2003). Partial least squares
for discrimination. Journal of Chemometrics, 17(3):166–173. (Cited on pages 195
and 199.)
[Bashashati et al., 2007a] Bashashati, A., Fatourechi, M., Ward, R. K., and Birch,
G. E. (2007a). A survey of signal processing algorithms in brain-computer
interfaces based on electrical brain signals. Journal of neural engineering, 4(2):R32–
57. (Cited on pages 4, 5, 13, 37, 41 and 75.)
[Bashashati et al., 2007b] Bashashati, A., Ward, R. K., and Birch, G. E.

(2007b). Towards development of a 3-state self-paced brain-computer interface.
Computational intelligence and neuroscience, 2007:84386. (Cited on pages 29
and 63.)
[Bashashati et al., 2015] Bashashati, H., Ward, R. K., Birch, G. E., and Bashashati,
A. (2015). Comparing different classifiers in sensory motor brain computer
interfaces. PLoS ONE, 10(6):1–17. (Cited on pages 43, 52, 53, 54, 55, 56, 119,
194 and 195.)
[Baxter et al., 2013] Baxter, B. S., Decker, A., and He, B. (2013). Noninvasive
control of a robotic arm in multiple dimensions using scalp electroencephalogram.
International IEEE/EMBS Conference on Neural Engineering, NER, pages 45–47.
(Cited on pages 10, 11, 12, 13, 19, 20, 29 and 40.)
[Bell et al., 2008] Bell, C. J., Shenoy, P., Chalodhorn, R., and Rao, R. P. (2008).
Control of a humanoid robot by a noninvasive brain-computer interface in humans.
Journal of neural engineering, 5(2 PG - 214-220):214–220. (Cited on pages 12
and 14.)
[Benavoli et al., 2015] Benavoli, A., Corani, G., and Mangili, F. (2015). Should we
really use post-hoc tests based on mean-ranks? Journal of Machine Learning
Research, pages 1–10. (Cited on pages 112 and 128.)
[Bengio and Frasconi, 1996] Bengio, Y. and Frasconi, P. (1996). Input-output

HMM’s for sequence processing. IEEE Transactions on Neural Networks,
7(5):1231–1249. (Cited on pages 57, 75, 76 and 84.)
[Bhattacharyya et al., 2015] Bhattacharyya, S., Basu, D., Konar, A., and Tibare-
wala, D. N. (2015). Interval type-2 fuzzy logic based multiclass ANFIS algo-
rithm for real-time EEG based movement control of a robot arm. Robotics and
Autonomous Systems, 68:104–115. (Cited on pages 19, 29, 30, 46 and 51.)
[Bhattacharyya et al., 2011] Bhattacharyya, S., Khasnobish, A., Konar, A., Tibare-
wala, D. N., and Nagar, A. K. (2011). Performance analysis of left/right hand
movement classification from EEG signal by intelligent algorithms. IEEE SSCI
2011 - Symposium Series on Computational Intelligence - CCMB 2011: 2011
IEEE Symposium on Computational Intelligence, Cognitive Algorithms, Mind,
214 Bibliography
and Brain, pages 1–8. (Cited on pages 43, 47, 50, 51, 52, 53, 54, 55, 56, 119, 193,
194 and 195.)
[Bilmes, 1998] Bilmes, J. a. (1998). A gentle tutorial of the EM algorithm and its
application to parameter estimation for Gaussian mixture and hidden Markov
models. International Computer Science Institute, 4(510):126. (Cited on page 86.)
[Birbaumer et al., 1999] Birbaumer, N., Ghanayim, N., Hinterberger, T., Iversen,
I., Kotchoubey, B., Kübler, a., Perelmouter, J., Taub, E., and Flor, H. (1999). A
spelling device for the paralysed. Nature, 398(6725):297–298. (Cited on page 13.)
[Bishop, 2006] Bishop, C. M. (2006). Pattern Recognition and Machine Learning,

volume 4. (Cited on pages 47, 51, 52, 53, 54, 61, 62, 72, 73, 119, 123, 175, 194
and 201.)
[Blabe et al., 2015] Blabe, C. H., Gilja, V., Chestek, C. a., Shenoy, K. V., Anderson,
K. D., and Henderson, J. M. (2015). Assessment of brain-machine interfaces
from the perspective of people with paralysis. Journal of neural engineering,
12(4):043002. (Cited on page 8.)
[Blackrock, 2016] Blackrock (2016). Blackrock’s Utah Array.

http://blackrockmicro.com/neuroscience-research-products/low-noise-ephys-elect
rodes/blackrock-utah-array/. (Cited on page 15.)
[Blackrock, 2017] Blackrock (2017). Blackrock Neuromed ECoG grid.

http://blackrockneuromed.com/products/elite/. (Cited on page 18.)
[Blakely et al., 2009] Blakely, T., Miller, K. J., Zanos, S. P., Rao, R. P. N., and
Ojemann, J. G. (2009). Robust, long-term control of an electrocorticographic
brain-computer interface with fixed parameters. Neurosurgical focus, 27(July):E13.
(Cited on pages 17, 18 and 44.)
[Blankertz et al., 2008] Blankertz, B., Tomioka, R., Lemm, S., Kawanabe, M., and
Müller, K. R. (2008). Optimizing spatial filters for robust EEG single-trial analysis.
IEEE Signal Processing Magazine, 25(1):41–56. (Cited on page 41.)
[Boostani et al., 2007] Boostani, R., Graimann, B., Moradi, M. H., and Pfurtscheller,
G. (2007). A comparison approach toward finding the best feature and classifier in
cue-based BCI. Medical and Biological Engineering and Computing, 45(4):403–
412. (Cited on pages 45 and 48.)
[Boostani and Moradi, 2004] Boostani, R. and Moradi, M. H. (2004). A new ap-
proach in the BCI research based on fractal dimension as feature and Adaboost
as classifier. J. Neural Eng, 1:212–217. (Cited on page 45.)
[Bordes and Vandekerkhove, 2005] Bordes, L. and Vandekerkhove, P. (2005). Statis-

tical Inference for Partially Hidden Markov Models. Communications in Statistics
- Theory and Methods, 34(5):1081–1104. (Cited on page 176.)
Bibliography 215
[Bougrain et al., 2012] Bougrain, L., Rochel, O., Boussaton, O., and Havet, L.
(2012). From the decoding of cortical activities to the control of a JACO robotic
arm: a whole processing chain. Control Architecture of Robots (CAR), pages
1–7. (Cited on page 29.)
[Bourlard and Morgan, 1998] Bourlard, H. and Morgan (1998). Hybrid HMM/ANN
systems for speech recognition: Overview and new research directions. Adaptive
Processing of Sequences and Data Structures, 1387(1):389–417. (Cited on pages 77
and 118.)
[Bouton et al., 2016] Bouton, C. E., Shaikhouni, A., Annetta, N. V., Bockbrader,
M. A., Friedenberg, D. A., Nielson, D. M., Sharma, G., Sederberg, P. B., Glenn,
B. C., Mysiw, W. J., Morgan, A. G., Deogaonkar, M., and Rezai, A. R. (2016).
Restoring cortical control of functional movement in a human with quadriplegia.
Nature, 000(7602):1–13. (Cited on pages 12, 15, 29, 31, 35 and 43.)
[Bouveyron and Brunet, 2013] Bouveyron, C. and Brunet, C. (2013). Model-Based

clustering of high-dimensional data : A review. Computational statistics and
data analysis, Elservier. (Cited on page 121.)
[Bradberry et al., 2010] Bradberry, T. J., Gentili, R. J., and Contreras-Vidal, J. L.

(2010). Reconstructing three-dimensional hand movements from noninvasive
electroencephalographic signals. The Journal of neuroscience : the official journal
of the Society for Neuroscience, 30(9):3432–7. (Cited on pages 23, 24, 27 and 59.)
[Bradberry et al., 2009] Bradberry, T. J., Rong, F., and Contreras-Vidal, J. L.

(2009). Decoding center-out hand velocity from MEG signals during visuomotor
adaptation. NeuroImage, 47(4):1691–700. (Cited on page 23.)
[BrainMaps, 2017] BrainMaps (2017). BrainMaps. BrainMaps.org. (Cited on

page 10.)
[Brockwell et al., 2004] Brockwell, a. E., Rojas, a. L., and Kass, R. E. (2004).
Recursive Bayesian Decoding of Motor Cortical Signals by Particle Filtering.
Journal of Neurophysiology, 91(4):1899–1907. (Cited on pages 61 and 62.)
[Brodu et al., 2011] Brodu, N., Lotte, F., and Lécuyer, A. (2011). Comparative
study of band-power extraction techniques for Motor Imagery classification. IEEE
SSCI 2011 - Symposium Series on Computational Intelligence - CCMB 2011: 2011
IEEE Symposium on Computational Intelligence, Cognitive Algorithms, Mind,
and Brain, pages 95–100. (Cited on pages 43 and 44.)
[Brunner et al., 2007] Brunner, C., Naeem, M., Leeb, R., Graimann, B., and
Pfurtscheller, G. (2007). Spatial filtering and selection of optimized compo-
nents in four class motor imagery EEG data using independent components
analysis. Pattern Recognition Letters, 28(8):957–964. (Cited on page 40.)
216 Bibliography
[Brunner et al., 2006] Brunner, C., Scherer, R., Graimann, B., Supp, G., and
Pfurtscheller, G. (2006). Online control of a brain-computer interface using phase
synchronization. IEEE Transactions on Biomedical Engineering, 53(12):2501–
[Bruns, 2004] Bruns, A. (2004). Fourier-, Hilbert- and wavelet-based signal analysis:
are they really different approaches? Journal of neuroscience methods, 137(2):321–
[Bundy et al., 2016] Bundy, D. T., Pahwa, M., Szrama, N., and Leuthardt, E. C.
(2016). Decoding three-dimensional reaching movements using electrocortico-
graphic signals in humans. Journal of Neural Engineering, 13(2):026021. (Cited
on pages 24, 27, 28, 39, 41, 44, 47, 52, 53, 59, 61, 63, 65, 70, 71, 72, 73, 74, 77, 78,
115, 119, 159, 163, 164, 165, 172, 188, 194 and 195.)
[Bundy et al., 2014] Bundy, D. T., Zellmer, E., Gaona, C. M., Sharma, M., Szrama,
N., Hacker, C., Freudenburg, Z. V., Daitch, A., Moran, D. W., and Leuthardt, E. C.
(2014). Characterization of the effects of the human dura on macro- and micro-
electrocorticographic recordings. Journal of neural engineering, 11(1):016006.
[Buzsáki et al., 2012] Buzsáki, G., Anastassiou, C. a., and Koch, C. (2012). The
origin of extracellular fields and currents–EEG, ECoG, LFP and spikes. Nature
reviews. Neuroscience, 13(6):407–20. (Cited on pages 4, 9, 14, 15 and 19.)
[Cacciatore and Nowlan, 1994] Cacciatore, T. W. and Nowlan, S. J. (1994). Mix-

tures of Controllers for Jump Linear and Non-Linear Plants. Advances in Neural
Information Processing Systems 6, (i):719–726. (Cited on page 75.)
[Cappé et al., 2006] Cappé, O., Moulines, E., and Ryden, T. (2006).
Inference in Hidden Markov models. (Cited on pages 81 and 82.)
[Cardillo, 2009] Cardillo, G. (2009). myfriedman: Friedman

test for non parametric two way ANalysis Of VAriance.
http://www.mathworks.com/matlabcentral/fileexchange/25882. (Cited on
page 112.)
[Carmena, 2013] Carmena, J. M. (2013). Advances in neuroprosthetic learning and

control. PLoS biology, 11(5):e1001561. (Cited on pages 6 and 25.)
[Carmena et al., 2003] Carmena, J. M., Lebedev, M. a., Crist, R. E., O’Doherty,
J. E., Santucci, D. M., Dimitrov, D. F., Patil, P. G., Henriquez, C. S., and Nicolelis,
M. a. L. (2003). Learning to control a brain-machine interface for reaching and
grasping by primates. PLoS biology, 1(2):E42. (Cited on pages 15, 16, 29, 46, 59
and 176.)
Bibliography 217
[Carvalhaes and De Barros, 2015] Carvalhaes, C. and De Barros, J. A. (2015). The

surface Laplacian technique in EEG: Theory and methods. International Journal
of Psychophysiology, 97(3):174–188. (Cited on page 41.)
[Cassisi et al., 2012] Cassisi, C., Montalto, P., Aliotta, M., Cannata, A., and Pul-
virenti, A. (2012). Similarity Measures and Dimensionality Reduction Techniques
for Time Series Data Mining. Advances in Data Mining Knowledge Discovery
and Applications, (January 2017):72–95. (Cited on page 109.)
[Chae et al., 2012] Chae, Y., Jeong, J., and Jo, S. (2012). Toward brain-actuated
humanoid robots: Asynchronous direct control using an EEG-Based BCI. IEEE
Transactions on Robotics, 28(5):1131–1144. (Cited on pages 3, 19, 40, 41 and 50.)
[Chao et al., 2010] Chao, Z. C., Nagasaka, Y., and Fujii, N. (2010). Long-term asyn-
chronous decoding of arm motion using electrocorticographic signals in monkeys.
Frontiers in neuroengineering, 3(March):3. (Cited on pages 17, 19, 24, 29, 41, 43,
44, 63, 64, 89, 90, 91, 115, 162 and 188.)
[Chatterjee et al., 2007] Chatterjee, A., Aggarwal, V., Ramos, A., Acharya, S., and
Thakor, N. V. (2007). A brain-computer interface with vibrotactile biofeedback
for haptic information. Journal of neuroengineering and rehabilitation, 4(1):40.
(Cited on page 30.)
[Chen et al., 2014a] Chen, C., Shin, D., Watanabe, H., Nakanishi, Y., Kambara,
H., Yoshimura, N., Nambu, A., Isa, T., Nishimura, Y., and Koike, Y. (2014a).
Decoding grasp force profile from electrocorticography signals in non-human
primate sensorimotor cortex. Neuroscience research, 83:1–7. (Cited on page 46.)
[Chen et al., 2009] Chen, C. W., Lin, C. C. K., and Ju, M. S. (2009). Hand orthosis
controlled using brain-computer interface. Journal of Medical and Biological
Engineering, 29(5):234–241. (Cited on page 29.)
[Chen et al., 2014b] Chen, W., Liu, X., and Litt, B. (2014b). Logistic-weighted
regression improves decoding of finger flexion from electrocorticographic sig-
nals. Conference proceedings : ... Annual International Conference of the IEEE
Engineering in Medicine and Biology Society. IEEE Engineering in Medicine and
Biology Society. Annual Conference, 2014:2629–2632. (Cited on pages 52, 53, 74,
77, 78, 115 and 116.)
[Chiappa and Bengio, 2003] Chiappa, S. and Bengio, S. (2003). Hmm and iohmm
modeling of eeg rhythms for asynchronous bci systems. Technical report, IDIAP.
(Cited on pages 13, 14, 50, 51, 56, 57 and 58.)
[Chin et al., 2007] Chin, C. M., Popovic, M. R., Thrasher, A., Cameron, T., Lozano,
A., and Chen, R. (2007). Identification of arm movements using correlation of
electrocorticographic spectral components and kinematic recordings. Journal of
neural engineering, 4:146–158. (Cited on pages 17, 26, 27, 42, 43, 51, 52 and 54.)
218 Bibliography
[Choi et al., 2009] Choi, K., Hirose, H., Sakurai, Y., Iijima, T., and Koike, Y. (2009).
Prediction of arm trajectory from the neural activities of the primary motor cortex
with modular connectionist architecture. Neural Networks, 22(9):1214–1223.
(Cited on page 46.)
[Cincotti et al., 2007] Cincotti, F., Kauhanen, L., Aloise, F., Palomäki, T., Ca-
porusso, N., Jylänki, P., Mattia, D., Babiloni, F., Vanacker, G., Nuttin, M.,
Marciani, M. G., and Millán, J. D. R. (2007). Vibrotactile feedback for brain-
computer interface operation. Computational Intelligence and Neuroscience, 2007.
(Cited on pages 5 and 30.)
[Cincotti et al., 2008] Cincotti, F., Mattia, D., Aloise, F., Bufalari, S., Schalk, G.,
Oriolo, G., Cherubini, A., Marciani, M. G., and Babiloni, F. (2008). Non-invasive
brain-computer interface system: towards its application as assistive technology.
Brain research bulletin, 75(6):796–803. (Cited on page 5.)
[Cincotti et al., 2003] Cincotti, F., Scipione, A., Timperi, A., Mattia, D., Marciani,
M. G., Millán, J., Salinari, S., Bianchi, L., and Babiloni, F. (2003). Comparison
of different feature classifiers for brain computer interfaces. In International
IEEE/EMBS Conference on Neural Engineering, NER, pages 645–647. (Cited on
pages 53 and 58.)
[ClinicalTrials.gov, 2016] ClinicalTrials.gov (2016). Brain Computer In-

terface: Neuroprosthetic Control of a Motorized Exoskeleton (BCI).
https://clinicaltrials.gov/ct2/show/record/NCT02550522?term=clinatec. (Cited
on pages 31 and 176.)
[Collinger et al., 2014] Collinger, J. L., Vinjamuri, R., Degenhart, A. D., Weber,
D. J., Sudre, G. P., Boninger, M. L., Tyler-Kabara, E. C., and Wang, W. (2014).
Motor-related brain activity during action observation: a neural substrate for
electrocorticographic brain-computer interfaces after spinal cord injury. Frontiers
in integrative neuroscience, 8(February):17. (Cited on page 173.)
[Collinger et al., 2013] Collinger, J. L., Wodlinger, B., Downey, J. E., Wang, W.,
Tyler-Kabara, E. C., Weber, D. J., McMorland, A. J. C., Velliste, M., Boninger,
M. L., and Schwartz, A. B. (2013). High-performance neuroprosthetic control by
an individual with tetraplegia. Lancet, 381(9866):557–64. (Cited on pages 5, 10,
11, 12, 15, 16, 24, 29, 30, 31, 42, 49, 59, 60 and 167.)
[Costecalde, 2012] Costecalde, T. (2012). Brain computer interface with chronic

cortical electrode arrays for motor deficit compensation in motor disabled
patients. Experimental study in rodents. PhD thesis, Université de Grenoble.
(Cited on page 34.)
[Coyle et al., 2005] Coyle, D., Prasad, G., and McGinnity, T. M. (2005). A time-
series prediction approach for feature extraction in a brain-computer interface.
Bibliography 219
IEEE Transactions on Neural Systems and Rehabilitation Engineering, 13(4):461–

[Coyle et al., 2007] Coyle, S. M., Ward, T. E., Markham, C. M., Ward, E., and
Markham, C. M. (2007). Brain-computer interface using a simplified functional
near-infrared spectroscopy system. Journal of neural engineering, 4(3):219–26.
(Cited on page 4.)
[Curran et al., 2004] Curran, E., Sykacek, P., Stokes, M., Roberts, S., Penny, W.,
Johnsrude, I., and Owen, A. (2004). Cognitive Tasks for Driving a Brain Computer
Interfacing System: A Pilot Study. IEEE Transactions on Neural Systems and
Rehabilitation Engineering, 12(1):48–54. (Cited on page 25.)
[Czepiel, 2012] Czepiel, S. A. (2012). Maximum Likelihood Estima-

tion of Logistic Regression Models: Theory and Implementation.
https://pdfs.semanticscholar.org/fd70/35ab3fca912c1f77c34ff0c4435d41ebd0d3.pdf.
(Cited on page 194.)
[Dangi et al., 2014] Dangi, S., Gowda, S., Moorman, H. G., Orsborn, A. L., So,
K., and Shanechi, M. (2014). Continuous Closed-Loop Decoder Adaptation
with a Recursive Maximum Likelihood Algorithm Allows for Rapid Performance
Acquisition in Brain-Machine Interfaces. Neural computation, 1872:1840–1872.
[Darmanjian et al., 2003] Darmanjian, S., Kim, S. P. K. S. P., Nechyba, M., Morri-
son, S., Principe, J., Wessberg, J., and Nicolelis, M. (2003). Bimodal brain-machine
interface for motor control of robotic prosthetic. Proceedings 2003 IEEE/RSJ
International Conference on Intelligent Robots and Systems (IROS 2003) (Cat.
No.03CH37453), 4(October):3–8. (Cited on pages 56 and 57.)
[Davis and Goadrich, 2006] Davis, J. and Goadrich, M. (2006). The Relationship
Between Precision-Recall and ROC Curves. Proceedings of the 23rd international
conference on Machine learning, pages 233–240. (Cited on page 107.)
[de Jong, 1993] de Jong, S. (1993). SIMPLS: An alternative approach to partial least
squares regression. Chemometrics and Intelligent Laboratory Systems, 18(3):251–
[Degenhart et al., 2016] Degenhart, A. D., Eles, J., Dum, R., Mischel, J. L.,
Smalianchuk, I., Endler, B., Ashmore, R. C., Tyler-Kabara, E. C., Hatsopoulos,
N. G., Wang, W., Batista, A. P., and Cui, X. T. (2016). Histological evaluation
of a chronicallyimplanted electrocorticographic electrode grid in a non-human
primate. Journal of Neural Engineering, 13(4):1–14. (Cited on page 19.)
[Delgado Saa and Cetin, 2012] Delgado Saa, J. F. and Cetin, M. (2012). A latent
discriminative model-based approach for classification of imaginary motor tasks
from EEG data. Journal of neural engineering, 9(2). (Cited on pages 57 and 58.)
220 Bibliography
[Demirer et al., 2009] Demirer, R. M., Ozerdem, M. S., and Bayrak, C. (2009).
Classification of imaginary movements in ECoG with a hybrid approach based
on multi-dimensional Hilbert-SVM solution. Journal of Neuroscience Methods,
178(1):214–218. (Cited on page 53.)
[Dempster et al., 1977] Dempster, A., Laird, N., and Rubin, D. B. (1977). Max-
imum likelihood from incomplete data via the EM algorithm. Journal of the
Royal Statistical Society Series B Methodological, 39(1):1–38. (Cited on pages 81
and 82.)
[Demšar, 2006] Demšar, J. (2006). Statistical Comparisons of Classifiers over Mul-

tiple Data Sets. The Journal of Machine Learning Research, 7:1–30. (Cited on
pages 112 and 128.)
[Devulapalli, 1996] Devulapalli, S. (1996). Non-Linear Principal Component

Analysis and Classification of EEG during mental tasks. PhD thesis, Colorado
State University. (Cited on page 47.)
[Dietterich, 2009] Dietterich, T. (2009). Machine learning for sequential data: A

review. Structural, Syntactic, and Statistical Pattern Recognition, 2396:227–246.
(Cited on pages 56, 57, 59, 74 and 81.)
[Dietterich and Bakiri, 1995] Dietterich, T. G. and Bakiri, G. (1995). Solving Multi-
class Learning Problems via Error-Correcting Output Codes. Journal of Artificial
Intelligence Research, 2:263–286. (Cited on page 194.)
[Do et al., 2013] Do, A. H., Wang, P. T., King, C. E., Chun, S. N., and Nenadic, Z.
(2013). Brain-Computer Interface Controlled Robotic Gait Orthosis. Journal of
NeuroEngineering and Rehabilitation, 10(1):111. (Cited on page 29.)
[Doud et al., 2011] Doud, A. J., Lucas, J. P., Pisansky, M. T., and He, B. (2011).
Continuous three-dimensional control of a virtual helicopter using a motor imagery
based brain-computer interface. PloS one, 6(10):e26322. (Cited on pages 13, 19,
20, 29, 30 and 40.)
[Duprès et al., 2014] Duprès, A., Cabestaing, F., and Rouillard, J. (2014). SSVEP-
based BCIs : study of classifier stability over time and effects of human
learning on classification accuracy. AMSE, Journal of the Association for the
Advancement of Modelling and Simulation Techniques in Enterprises (Special
edition HANDICAP). (Cited on page 22.)
[Eberly, 2014] Eberly, D. (2014). Derivative Approximation by Finite Differences

Derivatives of Univariate Functions. Magic Software, Inc, pages 1–7. (Cited on
page 116.)
[Eliseyev and Aksenova, 2013] Eliseyev, A. and Aksenova, T. (2013). Recursive

N-way partial least squares for brain-computer interface. PloS one, 8(7):e69962.
Bibliography 221
[Eliseyev and Aksenova, 2014] Eliseyev, A. and Aksenova, T. (2014). Stable and
artifact-resistant decoding of 3D hand trajectories from ECoG signals using the
generalized additive model. Journal of neural engineering, 11(6):066005. (Cited
on pages 34, 40, 43, 44, 47, 59, 60, 65, 78, 109, 116, 128, 175 and 195.)
[Eliseyev and Aksenova, 2016] Eliseyev, A. and Aksenova, T. (2016). Penalized

Multi-Way Partial Least Squares for Smooth Trajectory Decoding from Electro-
corticographic (ECoG) Recording. Plos One, 11(5):e0154878. (Cited on pages 47,
59, 63, 175 and 195.)
[Eliseyev et al., 2014] Eliseyev, A., Mestais, C., Charvet, G., Sauter, F., Abroug,
N., Arizumi, N., Cokgungor, S., Costecalde, T., Foerster, M., Korczowski, L.,
Morinière, B., Porcherot, J., Pradal, J., Ratel, D., Tarrin, N., Torres, N., Verney,
A., Aksenova, T., and Benabid, A.-l. (2014). CLINATEC ¥ R BCI platform
based on the ECoG-recording implant WIMAGINE ¥ R and the innovative signal-
processing : preclinical results. 36th Annual International Conference of the
IEEE Engineering in Medicince and Biology Society, pages 1222–1225. (Cited on
pages 31, 34 and 176.)
[Eliseyev et al., 2011] Eliseyev, A., Moro, C., Costecalde, T., Torres, N., Gharbi, S.,
Mestais, C., Benabid, A. L., and Aksenova, T. (2011). Iterative N-way partial least
squares for a binary self-paced brain-computer interface in freely moving animals.
Journal of neural engineering, 8(4):046012. (Cited on pages 34, 53 and 195.)
[Eliseyev et al., 2012] Eliseyev, A., Moro, C., Faber, J., Wyss, A., Torres, N., Mes-
tais, C., Benabid, A. L., and Aksenova, T. (2012). L1-penalized N-way PLS for
subset of electrodes selection in BCI experiments. Journal of neural engineering,
9(4):045010. (Cited on pages 34, 43, 44, 53, 56, 175 and 195.)
[Emotiv, 2016] Emotiv (2016). EMOTIV Epoc+. https://www.emotiv.com/epoc/.

(Cited on page 19.)
[Fan et al., 2011] Fan, J., Fan, Y., and Wu, Y. (2011). High-Dimensional Classifica-
tion. High-dimensional Data Analysis, pages 3–37. (Cited on page 51.)
[Fatourechi et al., 2007a] Fatourechi, M., Bashashati, A., Ward, R. K., and Birch,
G. E. (2007a). EMG and EOG artifacts in brain computer interface systems: A
survey. Clinical neurophysiology : official journal of the International Federation
of Clinical Neurophysiology, 118(3):480–94. (Cited on pages 20, 39 and 40.)
[Fatourechi et al., 2007b] Fatourechi, M., Birch, G. E., and Ward, R. K. (2007b).
Application of a hybrid wavelet feature selection method in the design of a self-
paced brain interface system. Journal of neuroengineering and rehabilitation,
4(Ic):11. (Cited on page 48.)
[Fazli et al., 2011] Fazli, S., Danóczy, M., Schelldorfer, J., and Müller, K.-R. (2011).
L1-penalized linear mixed-effects models for high dimensional data with application
to BCI. NeuroImage, 56(4):2100–2108. (Cited on page 48.)
222 Bibliography
[Fazli et al., 2009] Fazli, S., Grozea, C., Danoczy, M., Blankertz, B., Popescu,
F., and Muller, K.-R. (2009). Subject independent EEG-based BCI decoding.
Advances in Neural Information Processing Systems, 22:1–9. (Cited on page 21.)
[Felton et al., 2007] Felton, E. a., Wilson, J. A., Williams, J. C., and Garell, P. C.
(2007). Electrocorticographically controlled brain-computer interfaces using motor
and sensory imagery in patients with temporary subdural electrode implants.
Report of four cases. Journal of neurosurgery, 106(3):495–500. (Cited on page 44.)
[Fernandes et al., 2003] Fernandes, F. C., van Spaendonck, R. L., and Burrus, C. S.
(2003). A new framework for complex wavelet transforms. IEEE Transactions on
signal processing, 51(7):1825–1837. (Cited on page 43.)
[Ferreira et al., 2008] Ferreira, A., Celeste, W. C., Cheein, F. a., Bastos-Filho, T. F.,
Sarcinelli-Filho, M., and Carelli, R. (2008). Human-machine interfaces based
on EMG and EEG applied to robotic systems. Journal of neuroengineering and
rehabilitation, 5:10. (Cited on page 11.)
[Fifer et al., 2014] Fifer, M. S., Hotson, G., Wester, B. A., McMullen, D. P., Wang,
Y., Johannes, M. S., Katyal, K. D., Helder, J. B., Para, M. P., Vogelstein, R. J.,
Anderson, W. S., Thakor, N. V., and Crone, N. E. (2014). Simultaneous neural
control of simple reaching and grasping with the modular prosthetic limb using
intracranial EEG. IEEE Transactions on Neural Systems and Rehabilitation
Engineering, 22(3):695–705. (Cited on pages 17, 18, 31, 40, 41, 44, 50, 52, 55, 56,
57, 58, 74, 76, 118 and 172.)
[Fink, 2014] Fink, G. (2014). Markov models for pattern recognition. Springer
Science & Business Media. (Cited on pages 84 and 185.)
[Fitzsimmons, 2009] Fitzsimmons, N. A. (2009). Extracting kinematic parameters

for monkey bipedal walking from cortical neuronal ensemble activity. Frontiers
in Integrative Neuroscience, 3:3. (Cited on page 24.)
[Flamary and Rakotomamonjy, 2012] Flamary, R. and Rakotomamonjy, A. (2012).

Decoding Finger Movements from ECoG Signals Using Switching Linear Models.
Frontiers in neuroscience, 6(March):29. (Cited on pages 31, 42, 56, 66, 70, 71, 73,
74, 77, 78, 115, 116, 159, 165 and 189.)
[Flesher et al., 2016] Flesher, S. N., Collinger, J. L., Foldes, S. T., Weiss, J. M.,
Downey, J. E., Tyler-Kabara, E. C., Bensmaia, S. J., Schwartz, A. B., Boninger,
M. L., and Gaunt, R. A. (2016). Intracortical microstimulation of human so-
matosensory cortex. Science Translational Medicine, pages 1–11. (Cited on
page 5.)
[Flint et al., 2012] Flint, R. D., Lindberg, E. W., Jordan, L. R., Miller, L. E., and
Slutzky, M. W. (2012). Accurate decoding of reaching movements from field
potentials in the absence of spikes. Journal of Neural Engineering, 9(4):046006.
(Cited on page 60.)
Bibliography 223
[Flint et al., 2013] Flint, R. D., Wright, Z. a., Scheid, M. R., and Slutzky, M. W.
(2013). Long term, stable brain machine interface performance using local field
potentials and multiunit spikes. Journal of Neural Engineering, 10(5):056005.
(Cited on page 42.)
[Flotzinger et al., 1994] Flotzinger, D., Pregenzer, M., and Pfurtscheller, G. (1994).
Feature selection with distinction sensitive learning vector quantisation and
genetic algorithms. Proceedings of 1994 IEEE International Conference on Neural
Networks (ICNN’94), 6:3448–3451. (Cited on pages 41 and 48.)
[Foodeh et al., 2016] Foodeh, R., Khorasani, A., Shalchyan, V., and Daliri, M. R.
(2016). Minimum Noise Estimate filter: a Novel Automated Artifacts Re-
moval method for Field Potentials. IEEE Transactions on Neural Systems and
Rehabilitation Engineering, 4320(c):1–1. (Cited on page 40.)
[Fort and Lambert-Lacroix, 2005] Fort, G. and Lambert-Lacroix, S. (2005). Classifi-

cation using partial least squares with penalized logistic regression. Bioinformatics,
21(7):1104–1111. (Cited on page 175.)
[Friedman et al., 2001] Friedman, J., Hastie, T., and Tibshirani, R. (2001).
The Elements of Statistical Learning. (Cited on pages 49, 50, 53, 59 and 194.)
[Friedman et al., 2010] Friedman, J., Hastie, T., and Tibshirani, R. (2010). Regu-
larization Paths for Generalized Linear Models via Coordinate Descent. Journal
of statistical software, 33(1). (Cited on pages 123, 189 and 194.)
[Friedman, 1989] Friedman, J. H. (1989). Regularized Discriminant Analysis.

Journal of the American Statistical Association, 84(405):165–175. (Cited on
page 196.)
[Fruitet et al., 2010] Fruitet, J., McFarland, D. J., and Wolpaw, J. R. (2010). A
comparison of regression techniques for a two-dimensional sensorimotor rhythm-
based brain-computer interface. Journal of neural engineering, 7:16003. (Cited
on page 19.)
[Galán et al., 2008] Galán, F., Nuttin, M., Lew, E., Ferrez, P. W., Vanacker, G.,
Philips, J., and Millán, J. D. R. (2008). A brain-actuated wheelchair: Asyn-
chronous and non-invasive Brain-computer interfaces for continuous control of
robots. Clinical Neurophysiology, 119(9):2159–2169. (Cited on page 41.)
[Gancet et al., 2012] Gancet, J., Ilzkovitz, M., Motard, E., Nevatia, Y., Letier,
P., Weerdt, D. D., Hoellinger, T., Seetharaman, K., Petieau, M., Ivanenko, Y.,
Molinari, M., Pisotta, I., Tamburella, F., Labini, F. S., Avella, A., Kooij, H. V. D.,
Wang, L., Helm, F. V. D., Wang, S., Zanow, F., Hauffe, R., and Thorsteinsson,
F. (2012). MINDWALKER: Going One Step Further with Assistive Lower
Limbs Exoskeleton for SCI Condition Subjects. IEEE RAS/EMBS International
Conference on Biomedical Robotics and Biomechatronics, 2010:1794–1800. (Cited
on page 29.)
224 Bibliography
[Ganguly and Carmena, 2009] Ganguly, K. and Carmena, J. M. (2009). Emergence

of a stable cortical map for neuroprosthetic control. PLoS biology, 7(7):e1000153.
[Gao et al., 2003] Gao, Y., Black, M. J., Bienenstock, E., and Donoghue, J. P. (2003).
A quantitative comparison of linear and non-linear models of motor cortical activity
for the encoding and decoding of arm motions. First International IEEE EMBS
Conference on Neural Engineering, 2003. Conference Proceedings., pages 189–192.
[Garrett et al., 2003] Garrett, D., Peterson, D. a., Anderson, C. W., and Thaut,
M. H. (2003). Comparison of linear, nonlinear, and feature selection meth-
ods for EEG signal classification. IEEE Transactions on Neural Systems and
Rehabilitation Engineering, 11(2):141–144. (Cited on pages 48, 52, 54 and 55.)
[Geladi and Kowalski, 1986] Geladi, P. and Kowalski, B. R. (1986). Partial least-
squares regression: a tutorial. Analytica Chimica Acta, 185(C):1–17. (Cited on
page 189.)
[Georgopoulos et al., 1982] Georgopoulos, A. P., Kalaska, J. F., Caminiti, R., and
Massey5, J. T. (1982). On the relations between the direction of two-dimensional
arm movements and cell discharge in primate motor cortex. The Journal of
Neuroscience, 2(11):1527–1537. (Cited on pages 23 and 24.)
[Georgopoulos et al., 1986] Georgopoulos, a. P., Schwartz, a. B., Kettner, R. E.,

and Schwartz, a. B. (1986). Neuronal population coding of movement direction.
Science, 233(March):1416–1419. (Cited on page 59.)
[Ghahramani, 2001] Ghahramani, Z. (2001). An Introduction to Hidden Markov

Models and Bayesian Networks. International Journal of Pattern Recognition
and Artificial Intelligence, 15(01):9–42. (Cited on pages 77 and 79.)
[Goldfeld and Quandt, 1973] Goldfeld, S. M. and Quandt, R. E. (1973). A Markov

model for switching regressions. Journal of Econometrics, 1(1):3–15. (Cited on
page 70.)
[Golub et al., 2014] Golub, M. D., Yu, B. M., Schwartz, A. B., and Chase, S. M.
(2014). Motor cortical control of movement speed with implications for brain-
machine interface control. Journal of neurophysiology, 112(2):411–29. (Cited on
page 177.)
[Gouy-Pailler et al., 2009] Gouy-Pailler, C., Mattout, J., Congedo, M., and Jutten,
C. (2009). Uncued Brain-Computer Interfaces: a Variational Hidden Markov
Model of Mental State Dynamics. ESANN 09, Bruges : Belgium (2009). (Cited
on pages 53, 56, 108 and 194.)
[Gowda et al., 2014] Gowda, S., Orsborn, A. L., Overduin, S. a., Moorman, H. G.,
and Carmena, J. M. (2014). Designing dynamical properties of brain-machine
Bibliography 225
interfaces to optimize task-specific performance. IEEE Transactions on Neural

Systems and Rehabilitation Engineering, 22(5):911–920. (Cited on pages 15
and 177.)
[Graimann et al., 2009] Graimann, B., Allison, B., and Pfurtscheller, G. (2009).
Brain-Computer Interfaces: A Gentle Introduction. In Graimann, B., Pfurtscheller,
G., and Allison, B., editors, Brain-Computer Interfaces, The Frontiers Collection,
pages 1–28. Springer Berlin Heidelberg, Berlin, Heidelberg. (Cited on pages 1, 8
and 167.)
[Graimann et al., 2004] Graimann, B., Huggins, J. E., Levine, S. P., and
Pfurtscheller, G. (2004). Toward a direct brain interface based on human sub-
dural recordings and wavelet-packet analysis. IEEE transactions on bio-medical
engineering, 51(6):954–62. (Cited on page 48.)
[Gray, 1918] Gray, H. (1918). Anatomy of the Human Body. Lea \& Febiger. (Cited
on page 10.)
[G.tex, 2016] G.tex (2016). intendiX - User-

Ready Brain-Computer Interface Applications.
http://www.gtec.at/Products/Complete-Solutions/intendiX-Specs-Features.
(Cited on page 30.)
[Gunduz, 2008] Gunduz, A. (2008). Human motor control through

electrocorticographic brain machine interfaces. Doctoral dissertation, Uni-
versity of Florida. (Cited on pages 23, 24 and 27.)
[Gürel and Mehring, 2012] Gürel, T. and Mehring, C. (2012). Unsupervised adap-
tation of brain-machine interface decoders. Frontiers in neuroscience, 6(Novem-
ber):164. (Cited on page 177.)
[Gwin and Ferris, 2011] Gwin, J. T. and Ferris, D. (2011). High-density EEG and
independent component analysis mixture models distinguish knee contractions
from ankle contractions. Proceedings of the Annual International Conference of
the IEEE Engineering in Medicine and Biology Society, EMBS, pages 4195–4198.
(Cited on page 19.)
[Gysels and Celka, 2004] Gysels, E. and Celka, P. (2004). Phase synchronization for
the recognition of mental tasks in a brain-computer interface. IEEE Transactions
on Neural Systems and Rehabilitation Engineering, 12(4):406–415. (Cited on
page 45.)
[Hammer et al., 2013] Hammer, J., Fischer, J., Ruescher, J., Schulze-Bonhage, A.,
Aertsen, A., and Ball, T. (2013). The role of ECoG magnitude and phase in
decoding position, velocity, and acceleration during continuous motor behavior.
Frontiers in neuroscience, 7(November):200. (Cited on pages 27, 28, 41, 44, 45,
46, 63 and 162.)
226 Bibliography
[Hammer et al., 2016] Hammer, J., Pistohl, T., Fischer, J., Kršek, P., Tomášek, M.,
Marusič, P., Schulze-Bonhage, A., Aertsen, A., and Ball, T. (2016). Predominance
of Movement Speed Over Direction in Neuronal Population Signals of Motor
Cortex: Intracranial EEG Data and A Simple Explanatory Model. Cerebral
Cortex, page bhw033. (Cited on pages 27, 28, 42, 46, 61, 78 and 115.)
[Hamner et al., 2011] Hamner, B., Leeb, R., Tavella, M., and Del R. Millán, J.
(2011). Phase-based features for motor imagery brain-computer interfaces.
Proceedings of the Annual International Conference of the IEEE Engineering
in Medicine and Biology Society, EMBS, pages 2578–2581. (Cited on pages 44
and 45.)
[Hasan and Gan, 2009] Hasan, B. a. S. and Gan, J. Q. (2009). Unsupervised adap-
tive GMM for BCI. 2009 4th International IEEE/EMBS Conference on Neural
Engineering, NER ’09, (June):295–298. (Cited on pages 50 and 51.)
[Hasan and Gan, 2010] Hasan, B. A. S. and Gan, J. Q. (2010). Unsupervised move-
ment onset detection from EEG recorded during self-paced real hand movement.
Medical and Biological Engineering and Computing, 48(3):245–253. (Cited on
page 57.)
[Hasan and Gan, 2011] Hasan, B. A. S. and Gan, J. Q. (2011). Temporal modeling
of EEG during self-paced hand movement and its application in onset detection.
Journal of Neural Engineering, 8(5):056015. (Cited on pages 51, 57 and 75.)
[Haselsteiner and Pfurtscheller, 2000] Haselsteiner, E. and Pfurtscheller, G. (2000).

Using time-dependent neural networks for EEG classification. IEEE transactions
on rehabilitation engineering : a publication of the IEEE Engineering in Medicine
and Biology Society, 8(4):457–463. (Cited on pages 52, 54 and 58.)
[Hatsopoulos et al., 2004] Hatsopoulos, N., Joshi, J., and O’Leary, J. G. (2004).
Decoding continuous and discrete motor behaviors using motor and premotor
cortical ensembles. Journal of neurophysiology, 92(2):1165–1174. (Cited on
pages 50, 51, 52, 54 and 60.)
[Héliot et al., 2010] Héliot, R., Ganguly, K., Jimenez, J., and Carmena, J. M.
(2010). Learning in closed-loop brainmachine interfaces: Modeling and experi-
mental validation. IEEE Transactions on Systems, Man, and Cybernetics, Part
B: Cybernetics, 40(5):1387–1397. (Cited on page 21.)
[Herman et al., 2008] Herman, P., Prasad, G., McGinnity, T. M., and Coyle, D.
(2008). Comparative analysis of spectral approaches to feature extraction for
EEG-based motor imagery classification. IEEE Transactions on Neural Systems
and Rehabilitation Engineering, 16(4):317–326. (Cited on pages 39, 41 and 44.)
[Hill et al., 2006] Hill, N. J., Lal, T. N., Schröder, M., Hinterberger, T., Wilhelm, B.,
Nijboer, F., Mochty, U., Widman, G., Elger, C., Schölkopf, B., Kübler, A., and Bir-
baumer, N. (2006). Classifying EEG and ECoG signals without subject training for
Bibliography 227
fast BCI implementation: comparison of nonparalyzed and completely paralyzed

subjects. IEEE transactions on neural systems and rehabilitation engineering,
14(2):183–186. (Cited on pages 27, 44, 51 and 53.)
[Hochberg et al., 2012] Hochberg, L. R., Bacher, D., Jarosiewicz, B., Masse, N. Y.,
Simeral, J. D., Vogel, J., Haddadin, S., Liu, J., Cash, S. S., van der Smagt, P.,
and Donoghue, J. P. (2012). Reach and grasp by people with tetraplegia using a
neurally controlled robotic arm. Nature, 485(7398):372–375. (Cited on pages 5,
12, 15, 16, 23, 24, 25, 29, 30, 31, 35, 49, 61, 62, 64, 172, 173, 176 and 209.)
[Hochberg et al., 2006] Hochberg, L. R., Serruya, M. D., Friehs, G. M., Mukand,
J. a., Saleh, M., Caplan, A. H., Branner, A., Chen, D., Penn, R. D., and Donoghue,
J. P. (2006). Neuronal ensemble control of prosthetic devices by a human with
tetraplegia. Nature, 442(7099):164–71. (Cited on pages 15, 16, 23, 26, 35, 59
and 64.)
[Homer and Nurmikko, 2013] Homer, M. and Nurmikko, A. (2013). Sensors and
decoding for intracortical brain computer interfaces. Annual review of biomedical
engineering, 15:383–405. (Cited on pages 3, 4, 9, 15, 16 and 22.)
[Hong et al., 2015] Hong, K. S., Naseer, N., and Kim, Y. H. (2015). Classification
of prefrontal and motor cortex signals for three-class fNIRS-BCI. Neuroscience
Letters, 587:87–92. (Cited on page 4.)
[Horki et al., 2011] Horki, P., Solis-Escalante, T., Neuper, C., and Müller-Putz,
G. (2011). Combined motor imagery and SSVEP based BCI control of a 2
DoF artificial upper limb. Medical and Biological Engineering and Computing,
49(5):567–577. (Cited on page 14.)
[Hortal et al., 2015] Hortal, E., Planelles, D., Costa, A., Iáñez, E., Úbeda, A.,
Azorín, J. M., and Fernández, E. (2015). SVM-based Brain-Machine Interface for
controlling a robot arm through four mental tasks. Neurocomputing, 151(P1):116–
121. (Cited on pages 20, 30, 42, 46, 50, 51, 52, 55 and 118.)
[Höskuldsson, 1988] Höskuldsson, A. (1988). PLS Regression Methods. Journal of

Chemometrics, 2(July):211–228. (Cited on pages 47, 118 and 188.)
[Hotson et al., 2014] Hotson, G., Fifer, M. S., Acharya, S., Benz, H. L., Anderson,
W. S., Thakor, N. V., and Crone, N. E. (2014). Coarse electrocorticographic
decoding of ipsilateral reach in patients with brain lesions. PLoS ONE, 9(12):1–20.
(Cited on pages 28, 41, 42, 46, 48, 63 and 163.)
[Hotson et al., 2016] Hotson, G., McMullen, D. P., Fifer, M. S., Johannes, M. S.,
Katyal, K. D., Para, M. P., Armiger, R., Anderson, W. S., Thakor, N. V., Wester,
B. A., and Crone, N. E. (2016). Individual finger control of a modular prosthetic
limb using high-density electrocorticography in a human subject. Journal of
Neural Engineering, 13(2):026017. (Cited on pages 12, 17, 18, 31, 46, 49, 50, 52,
55, 56, 57, 58, 74, 75, 76, 118, 165, 172 and 175.)
228 Bibliography
[Hsu and Lin, 2002] Hsu, C. W. and Lin, C. J. (2002). A comparison of methods
for multiclass support vector machines. IEEE Transactions on Neural Networks,
13(2):415–425. (Cited on page 194.)
[Hudson and Burdick, 2007] Hudson, N. and Burdick, J. W. (2007). Learning hybrid
system models for supervisory decoding of discrete state, with applications to
the parietal reach region. Proceedings of the 3rd International IEEE EMBS
Conference on Neural Engineering, pages 587–592. (Cited on page 21.)
[Hyndman and Koehler, 2006] Hyndman, R. J. and Koehler, A. B. (2006). Another

look at measures of forecast accuracy. International Journal of Forecasting,
22(4):679–688. (Cited on page 108.)
[Ifft et al., 2013] Ifft, P. J., Shokur, S., Li, Z., Lebedev, M. A., and Nicolelis, M.
A. L. (2013). A brain-machine interface enables bimanual arm movements in
monkeys. Science translational medicine, 5(210):210ra154. (Cited on pages 12,
15, 16, 21, 24, 30, 31, 35, 61, 62, 67, 70, 159 and 177.)
[Imamizu et al., 2007] Imamizu, H., Sugimoto, N., Osu, R., Tsutsui, K., Sugiyama,
K., Wada, Y., and Kawato, M. (2007). Explicit contextual information selec-
tively contributes to predictive switching of internal models. Experimental Brain
Research, 181(3):395–408. (Cited on page 69.)
[Iriarte et al., 2003] Iriarte, J., Urrestarazu, E., Valencia, M., Alegre, M., Malanda,
A., Viteri, C., and Artieda, J. (2003). Independent component analysis as a
tool to eliminate artifacts in EEG: a quantitative study. Journal of clinical
neurophysiology : official publication of the American Electroencephalographic
Society, 20(4):249–257. (Cited on page 40.)
[Jackson and Fetz, 2011] Jackson, A. and Fetz, E. E. (2011). Interfacing with the
computational brain. IEEE Transactions on Neural Systems and Rehabilitation
Engineering, 19(5):534–541. (Cited on pages 22, 25 and 177.)
[Jacobs et al., 1991] Jacobs, R. a., Jordan, M. I., Nowlan, S. J., and Hinton, G. E.
(1991). Adaptive Mixtures of Local Experts. Neural Computation, 3(1):79–87.
[Jarosiewicz et al., 2013] Jarosiewicz, B., Masse, N. Y., Bacher, D., Cash, S. S.,
Eskandar, E., Friehs, G., Donoghue, J. P., and Hochberg, L. R. (2013). Advantages
of closed-loop calibration in intracortical brain–computer interfaces for people
with tetraplegia. Journal of Neural Engineering, 10(4):046012. (Cited on pages 15,
16, 21, 22 and 177.)
[Jerbi et al., 2011] Jerbi, K., Vidal, J., Mattout, J., Maby, E., Lecaignard, F.,
Ossandon, T., Hamamé, C., Dalal, S., Bouet, R., Lachaux, J.-P., Leahy, R.,
Baillet, S., Garnero, L., Delpuech, C., and Bertrand, O. (2011). Inferring hand
movement kinematics from MEG, EEG and intracranial EEG: From brain-machine
interfaces to motor rehabilitation. IRBM, 32(1):8–18. (Cited on pages 24 and 27.)
Bibliography 229
[Kandel et al., 2000] Kandel, E., Schwartz, J., and Jessell, T. (2000).
Principles Of Neural Science. (Cited on pages 9, 10, 11 and 15.)
[Kang et al., 2012] Kang, X., Schieber, M. H., and Thakor, N. V. (2012). Decoding
of finger, hand and arm kinematics using switching linear dynamical systems with
pre-motor cortical ensembles. Proceedings of the Annual International Conference
of the IEEE Engineering in Medicine and Biology Society, EMBS, pages 1732–
[Kao et al., 2017] Kao, J. C., Nuyujukian, P., Ryu, S. I., and Shenoy, K. V. (2017).
A high-performance neural prosthesis incorporating discrete state selection with
hidden Markov models. IEEE Transactions on Biomedical Engineering, 64(4):935–
945. (Cited on pages 47, 57, 58, 159, 172 and 177.)
[Kao et al., 2013] Kao, J. C., Nuyujukian, P., Stavisky, S., Ryu, S. I., Ganguli, S.,
and Shenoy, K. V. (2013). Investigating the role of firing-rate normalization and
dimensionality reduction in brain-machine interface robustness. In Proceedings of
the Annual International Conference of the IEEE Engineering in Medicine and
Biology Society, EMBS, volume 2010, pages 293–298. (Cited on pages 47, 72, 207
and 209.)
[Kapeller et al., 2015] Kapeller, C., Gergondet, P., Kamada, K., Ogawa, H.,
Takeuchi, F., Ortner, R., Prückl, R., Kheddar, A., Scharinger, J., and Guger, C.
(2015). Online Control of a Humanoid Robot through Hand Movement Imagination
using CSP and ECoG based Features. In Engineering in Medicine and Biology
Society (EMBC), 2015 37th Annual International Conference of the IEEE, pages
1765–1768. (Cited on pages 12, 17, 18, 21, 31, 41, 45, 50, 52, 55 and 118.)
[Karin and Andres, 2014] Karin, C. A. and Andres, V. A. (2014). Virtual hand
prosthesis moved by encephalographic signals. 2014 III International Congress
of Engineering Mechatronics and Automation (CIIMA), pages 1–5. (Cited on
page 40.)
[Kayikcioglu and Aydemir, 2010] Kayikcioglu, T. and Aydemir, O. (2010). A poly-

nomial fitting and k-NN based approach for improving classification of motor
imagery BCI data. Pattern Recognition Letters, 31(11):1207–1215. (Cited on
pages 52, 54 and 55.)
[Ke and Li, 2009] Ke, L. and Li, R. (2009). Classification of EEG signals by multi-
scale filtering and PCA. Proceedings - 2009 IEEE International Conference on
Intelligent Computing and Intelligent Systems, ICIS 2009, 1:362–366. (Cited on
pages 47 and 195.)
[Kellis et al., 2012] Kellis, S., Hanrahan, S., Davis, T., House, P. A., Brown, R., and
Greger, B. (2012). Decoding hand trajectories from micro-electrocorticography
in human patients. In Proceedings of the Annual International Conference of
230 Bibliography
(Cited on pages 10, 11, 12, 13, 17, 18, 30, 42, 62 and 63.)
[Kellis et al., 2015] Kellis, S., Sorensen, L., Darvas, F., Sayres, C., O’Neill, K.,
Brown, R., House, P., Ojemann, J., and Greger, B. (2015). Multi-scale analysis
of neural activity in humans: implications for micro-scale electrocorticography.
Clinical Neurophysiology, pages 1–11. (Cited on page 18.)
[Kelly et al., 2012] Kelly, J. W., Degenhart, A. D., Siewiorek, D. P., Smailagic, A.,
and Wang, W. (2012). Sparse linear regression with elastic net regularization for
brain-computer interfaces. Proceedings of the Annual International Conference of
(Cited on page 48.)
[Kelly et al., 2011] Kelly, J. W., Siewiorek, D. P., Smailagic, A., Collinger, J. L.,
Weber, D. J., and Wang, W. (2011). Fully automated reduction of ocular artifacts
in high-dimensional neural data. IEEE Transactions on Biomedical Engineering,
58(3 PART 1):598–606. (Cited on page 40.)
[Kemere et al., 2008] Kemere, C., Santhanam, G., Yu, B. M., Afshar, A., Ryu, S. I.,
Meng, T. H., and Shenoy, K. V. (2008). Detecting neural-state transitions using
hidden Markov models for motor cortical prostheses. Journal of Neurophysiology,
100(4):2441–2452. (Cited on pages 57 and 75.)
[Kemere et al., 2004a] Kemere, C., Santhanam, G., Yu, B. M., Ryu, S., Meng, T.,
and Shenoy, K. V. (2004a). Model-based decoding of reaching movements for
prosthetic systems. Conference proceedings : ... Annual International Conference
of the IEEE Engineering in Medicine and Biology Society. IEEE Engineering
in Medicine and Biology Society. Conference, 6:4524–4528. (Cited on pages 67
and 71.)
[Kemere et al., 2004b] Kemere, C., Shenoy, K. V., and Meng, T. H. (2004b). Model-
based neural decoding of reaching movements: A maximum likelihood approach.
IEEE Transactions on Biomedical Engineering, 51(6):925–932. (Cited on pages 67
and 71.)
[Kemsley, 1996] Kemsley, E. (1996). Discriminant analysis of high-dimensional data:

a comparison of principal components analysis and partial least squares data
reduction methods. Chemometrics and Intelligent Laboratory Systems, 33(1):47–
[Khreich et al., 2012] Khreich, W., Granger, E., Miri, A., and Sabourin, R. (2012).
A survey of techniques for incremental learning of HMM parameters. Information
Sciences, 197:105–130. (Cited on page 176.)
[Kim, 1994] Kim, C.-J. (1994). Dynamic linear models with Markov-switching.
Journal of Econometrics, 60:1–22. (Cited on page 70.)
Bibliography 231
[Kim et al., 2006a] Kim, H. K., Biggs, S. J., Schloerb, D. W., Carmena, J. M.,
Lebedev, M. a., Nicolelis, M. a. L., and Srinivasan, M. a. (2006a). Continuous
shared control for stabilizing reaching and grasping with brain-machine interfaces.
IEEE transactions on bio-medical engineering, 53(6):1164–73. (Cited on page 8.)
[Kim et al., 2006b] Kim, K. H., Kim, S. S., and Kim, S. J. (2006b). Superiority
of nonlinear mapping in decoding multiple single-unit neuronal spike trains: A
simulation study. Journal of Neuroscience Methods, 150(2):202–211. (Cited on
pages 60 and 175.)
[Kim et al., 2003] Kim, S.-P., Sanchez, J. C., Erdogmus, D., Rao, Y. N., Wessberg,
J., Principe, J. C., and Nicolelis, M. (2003). Divide-and-conquer approach for
brain machine interfaces: nonlinear mixture of competitive linear models. Neural
networks : the official journal of the International Neural Network Society, 16(5-
6):865–71. (Cited on pages 67, 70, 71, 72, 73 and 159.)
[Kim et al., 2006c] Kim, S.-P., Sanchez, J. C., Rao, Y. N., Erdogmus, D., Carmena,
J. M., Lebedev, M. a., Nicolelis, M. a. L., and Principe, J. C. (2006c). A
comparison of optimal MIMO linear and nonlinear models for brain-machine
interfaces. Journal of neural engineering, 3:145–161. (Cited on pages 47, 60, 61,
63, 66, 70, 159 and 163.)
[Kim et al., 2008] Kim, S.-P., Simeral, J. D., Hochberg, L. R., Donoghue, J. P., and
Black, M. J. (2008). Neural control of computer cursor velocity by decoding
motor cortical spiking activity in humans with tetraplegia. Journal of Neural
[Kim et al., 2011] Kim, S.-P., Simeral, J. D., Hochberg, L. R., Donoghue, J. P.,
Friehs, G. M., and Black, M. J. (2011). Point-and-click cursor control with an intra-
cortical neural interface system by humans with tetraplegia. IEEE Transactions
on Neural Systems and Rehabilitation Engineering, 19(2):193–203. (Cited on
pages 15, 16, 30, 51 and 56.)
[Kindermans et al., 2012] Kindermans, P.-J., Verschore, H., Verstraeten, D., and
Schrauwen, B. (2012). A P300 BCI for the Masses: Prior Information Enables In-
stant Unsupervised Spelling. Advances in Neural Information Processing Systems
25, pages 719–727. (Cited on page 2.)
[King et al., 2015] King, C. E., Wang, P. T., McCrimmon, C. M., Chou, C. C., Do,
A. H., and Nenadic, Z. (2015). The feasibility of a brain-computer interface func-
tional electrical stimulation system for the restoration of overground walking after
paraplegia. Journal of NeuroEngineering and Rehabilitation, 12(1):80. (Cited on
pages 19, 29, 46, 63, 74 and 75.)
[King-Stephens et al., 2015] King-Stephens, D., Mirro, E., Weber, P. B., Laxer,
K. D., Van Ness, P. C., Salanova, V., Spencer, D. C., Heck, C. N., Goldman, A.,
Jobst, B., Shields, D. C., Bergey, G. K., Eisenschenk, S., Worrell, G. A., Rossi,
232 Bibliography
M. A., Gross, R. E., Cole, A. J., Sperling, M. R., Nair, D. R., Gwinn, R. P., Park,
Y. D., Rutecki, P. A., Fountain, N. B., Wharen, R. E., Hirsch, L. J., Miller, I. O.,
Barkley, G. L., Edwards, J. C., Geller, E. B., Berg, M. J., Sadler, T. L., Sun,
F. T., and Morrell, M. J. (2015). Lateralization of mesial temporal lobe epilepsy
with chronic ambulatory electrocorticography. Epilepsia, 56(6):959–967. (Cited
on page 19.)
[Koike et al., 2006] Koike, Y., Hirose, H., Sakurai, Y., and Iijima, T. (2006). Pre-
diction of arm trajectory from a small number of neuron activities in the primary
motor cortex. Neuroscience Research, 55(2):146–153. (Cited on page 46.)
[Kousarrizi et al., 2009] Kousarrizi, M., Ghanbari, A., Teshnehlab, M., Aliyari,
M., and Gharaviri, A. (2009). Feature Extraction and Classification of EEG
Signals Using Wavelet Transform, SVM and Artificial Neural Networks for Brain
Computer Interfaces. 2009 International Joint Conference on Bioinformatics,
Systems Biology and Intelligent Computing, pages 16–19. (Cited on page 43.)
[Koyama et al., 2010a] Koyama, S., Castellanos Pérez-Bolde, L., Shalizi, C. R., and
Kass, R. E. (2010a). Approximate Methods for State-Space Models. Journal of
the American Statistical Association, 105(489):170–180. (Cited on page 62.)
[Koyama et al., 2010b] Koyama, S., Chase, S. M., Whitford, A. S., Velliste, M.,
Schwartz, A. B., and Kass, R. E. (2010b). Comparison of brain-computer
interface decoding algorithms in open-loop and closed-loop control. Journal
of computational neuroscience, 29(1-2):73–87. (Cited on pages 59, 61, 62, 63
and 172.)
[Krämer, 2011] Krämer, N. (2011). The Degrees of Freedom of Partial Least Squares
Regression. Journal of the American Statistical Association, 106(494):697–705.
[Krishnamurthy, 2016] Krishnamurthy, V. (2016).

Partially Observed Markov Decision Processes. Cambridge University Press.
[Krusienski et al., 2012] Krusienski, D. J., McFarland, D. J., and Wolpaw, J. R.

(2012). Value of amplitude, phase, and coherence features for a sensorimotor
rhythm-based brain-computer interface. Brain Research Bulletin, 87(1):130–134.
[Krusienski et al., 2006] Krusienski, D. J., Sellers, E. W., Cabestaing, F., Bayoudh,
S., McFarland, D. J., Vaughan, T. M., and Wolpaw, J. R. (2006). A comparison
of classification techniques for the P300 Speller. Journal of neural engineering,
3(4):299–305. (Cited on page 2.)
[Kubánek et al., 2009] Kubánek, J., Miller, K. J., Ojemann, J. G., Wolpaw, J. R.,
and Schalk, G. (2009). Decoding flexion of individual fingers using electrocortico-
Bibliography 233
graphic signals in humans. Journal of Neural Engineering, 6(6):066001. (Cited

on pages 31, 39, 59 and 78.)
[Kuratowski, 2014] Kuratowski, K. (2014). Topology. Elsevier. (Cited on page 110.)
[Kwak et al., 2015] Kwak, N.-S., Müller, K.-R., and Lee, S.-W. (2015). A lower
limb exoskeleton control system based on steady state visual evoked potentials.
Journal of Neural Engineering, 12(5):056009. (Cited on page 29.)
[Lachaux et al., 2012] Lachaux, J.-P., Axmacher, N., Mormann, F., Halgren, E.,
and Crone, N. E. (2012). High-frequency neural activity and human cognition:
Past, present and possible future of intracranial EEG research. Progress in
Neurobiology, 98(3):279–301. (Cited on page 173.)
[Lafferty et al., 2001] Lafferty, J., McCallum, A., and Pereira, F. C. N. (2001).
Conditional random fields: Probabilistic models for segmenting and labeling
sequence data. ICML ’01 Proceedings of the Eighteenth International Conference
on Machine Learning, 8(June):282–289. (Cited on page 57.)
[LaFleur et al., 2013] LaFleur, K., Cassady, K., Doud, A., Shades, K., Rogin, E., and
He, B. (2013). Quadcopter control in three-dimensional space using a noninvasive
motor imagery-based brain-computer interface. Journal of neural engineering,
10(4):046003. (Cited on pages 4, 12, 19, 20, 25, 26, 29, 30, 40, 42, 48 and 50.)
[Lal et al., 2005] Lal, T. T. N., Hinterberger, T., Widman, G., Schröder, M.,
Schroeder, M., Hill, J., Rosenstiel, W., Elger, C., Schölkopf, B., and Birbaumer,
N. (2005). Methods towards invasive human brain computer interfaces. Advances
in Neural Information Processing System, 17:737–744. (Cited on pages 27, 44, 51
and 53.)
[Le Van Quyen et al., 2001] Le Van Quyen, M., Foucher, J., Lachaux, J., Rodriguez,
E., Lutz, a., Martinerie, J., and Varela, F. J. (2001). Comparison of Hilbert
transform and wavelet methods for the analysis of neuronal synchrony. Journal
of neuroscience methods, 111(2):83–98. (Cited on page 44.)
[Lebedev and Nicolelis, 2006] Lebedev, M. a. and Nicolelis, M. a. L. (2006). Brain-

machine interfaces: past, present and future. Trends in neurosciences, 29(9):536–
46. (Cited on pages 2, 4, 5, 6, 11, 20 and 30.)
[Leeb et al., 2007] Leeb, R., Friedman, D., Müller-Putz, G. R., Scherer, R., Slater,
M., and Pfurtscheller, G. (2007). Self-paced (asynchronous) BCI control of a
wheelchair in virtual environments: a case study with a tetraplegic. Computational
intelligence and neuroscience, 2007:79642. (Cited on pages 26, 35 and 46.)
[Lemm et al., 2004] Lemm, S., Schafer, C., and Curio, G. (2004). BCI competi-
tion–Data set III: Probabilistic modelling of sensorimotor m rhythms for classifica-
tion of imaginary hand movements. IEEE Trans. Biomed. Eng, 51(6):1077–1080.
234 Bibliography
[Leuthardt et al., 2011] Leuthardt, E. C., Gaona, C., Sharma, M., Szrama, N.,
Roland, J., Freudenberg, Z., Solis, J., Breshears, J., and Schalk, G. (2011). Using
the electrocorticographic speech network to control a brain-computer interface
in humans. Journal of neural engineering, 8(3):036004. (Cited on pages 11, 31
and 44.)
[Leuthardt et al., 2006a] Leuthardt, E. C., Miller, K. J., Schalk, G., Rao, R. P. N.,
and Ojemann, J. G. (2006a). Electrocorticography-based brain computer interface–
the Seattle experience. IEEE transactions on neural systems and rehabilitation
engineering : a publication of the IEEE Engineering in Medicine and Biology
Society, 14(2):194–8. (Cited on pages 14, 16, 17, 18, 20, 22, 27, 30 and 177.)
[Leuthardt et al., 2006b] Leuthardt, E. C., Schalk, G., Moran, D., and Ojemann,
J. G. (2006b). The emerging world of motor neuroprosthetics: a neurosurgical
perspective. Neurosurgery, 59(1):1–14; discussion 1–14. (Cited on pages 1, 6, 8,
12 and 13.)
[Leuthardt et al., 2004] Leuthardt, E. C., Schalk, G., Wolpaw, J. R., Ojemann, J. G.,
and Moran, D. W. (2004). A brain-computer interface using electrocorticographic
signals in humans. Journal of neural engineering, 1(2):63–71. (Cited on pages 17,
18, 27, 39 and 44.)
[Lew, 2012] Lew, E. (2012). Detection of self-paced reaching movement intention

from EEG signals. Frontiers in Neuroengineering, 5(July):1–17. (Cited on pages 2
and 41.)
[Li et al., 2002] Li, B., Morris, J., and Martin, E. B. (2002). Model selection
for partial least squares regression. Chemometrics and Intelligent Laboratory
Systems, 64:79–89. (Cited on pages 118, 119, 189 and 207.)
[Li, 2014] Li, Z. (2014). Decoding methods for neural prostheses: where have we
reached? Frontiers in Systems Neuroscience, 8(July):1–6. (Cited on pages 46, 58,
59, 61 and 62.)
[Li et al., 2009] Li, Z., O’Doherty, J. E., Hanson, T. L., Lebedev, M. a., Henriquez,
C. S., and Nicolelis, M. a. L. (2009). Unscented Kalman Filter for Brain-Machine
Interfaces. PLoS ONE, 4(7):e6243. (Cited on pages 59, 62, 63 and 189.)
[Liang and Bougrain, 2009] Liang, N. and Bougrain, L. (2009). Decoding Finger
Flexion using Amplitude Modulation form Band-specific ECoG. In European
Symposium on Artificial Neural Networks - ESANN 2009, pages 467–472. (Cited
on pages 78, 115 and 116.)
[Liang and Bougrain, 2012] Liang, N. and Bougrain, L. (2012). Decoding finger
flexion from band-specific ecog signals in humans. Frontiers in Neuroscience,
6(JUN):1–6. (Cited on pages 31, 48, 59 and 63.)
Bibliography 235
[Liehr et al., 1999] Liehr, S., Pawelzik, K., Kohlmorgen, J., and Müller, K.-R. (1999).
Hidden Markov Mixtures of Experts with an Application to EEG Recordings
from Sleep. Theory in Biosciences, 118(3-4):246–260. (Cited on page 70.)
[Lingjaerde and Christophersen, 2000] Lingjaerde, O. C. and Christophersen, N.

(2000). Shrinkage Structure of Partial Least Squares. Scandinavian Journal of
Statistics, 27(3):459–473. (Cited on page 188.)
[Liu et al., 2004] Liu, H., Hild, K. E., Gao, J. B., Erdogmus, D., Príncipe, J. C., and
Chris Sackellares, J. (2004). Evaluation of a BSS algorithm for artifacts rejection
in epileptic seizure detection. Conference proceedings : ... Annual International
Engineering in Medicine and Biology Society. Conference, 1:91–4. (Cited on
page 40.)
[Loboda et al., 2014] Loboda, A., Margineanu, A., and Rotariu, G. (2014). Dis-
crimination of EEG-Based Motor Imagery Tasks by Means of a Simple Phase
Information Method. International Journal of Advanced Research in Artificial
Intelligence (IJARAI), 3(10), 3(10):11–15. (Cited on page 45.)
[López-Larraz et al., 2016] López-Larraz, E., Trincado-Alonso, F., Rajasekaran,

V., Pérez-Nombela, S., Del-Ama, A. J., Aranda, J., Minguez, J., Gil-Agudo,
A., and Montesano, L. (2016). Control of an Ambulatory Exoskeleton with a
Brain-Machine Interface for Spinal Cord Injury Gait Rehabilitation. Frontiers in
neuroscience, 10(August):359. (Cited on pages 29 and 39.)
[Lotte et al., 2007] Lotte, F., Congedo, M., Lécuyer, A., Lamarche, F., and Arnaldi,
B. (2007). A review of classification algorithms for EEG-based brain-computer
interfaces. Journal of Neural Engineering, 4(2):R1–R13. (Cited on pages 53, 56,
58 and 59.)
[Ma et al., 2017] Ma, X., Ma, C., Huang, J., Zhang, P., Xu, J., and He, J. (2017).
Decoding Lower Limb Muscle Activity and Kinematics from Cortical Neural Spike
Trains during Monkey Performing Stand and Squat Movements. Frontiers in
Neuroscience, 11(February):1–16. (Cited on page 24.)
[Mahmoudi and Erfanian, 2002] Mahmoudi, B. and Erfanian, a. (2002). Single-

channel EEG-based prosthetic hand grasp control for amputee subjects.
Proceedings of the Second Joint 24th Annual Conference and the Annual Fall
Meeting of the Biomedical Engineering Society] [Engineering in Medicine and
Biology, 3:2406–2407. (Cited on page 54.)
[Mak and Wolpaw, 2009] Mak, J. N. and Wolpaw, J. R. (2009). Clinical Applications
of Brain – Computer Interfaces : Current State and Future Prospects. IEEE
reviews in biomedical engineering, 2:187–199. (Cited on pages 1, 2, 3, 4, 5, 8, 9,
16, 30, 42 and 167.)
236 Bibliography
[Malik et al., 2015] Malik, W. Q., Hochberg, L. R., Donoghue, J. P., and Brown,
E. N. (2015). Modulation depth estimation and variable selection in state-space
models for neural interfaces. IEEE Transactions on Biomedical Engineering,
62(2):570–581. (Cited on page 207.)
[Marathe and Taylor, 2011] Marathe, a. R. and Taylor, D. M. (2011). Decoding

position, velocity, or goal: does it matter for brain-machine interfaces? Journal
of neural engineering, 8(2):025016. (Cited on pages 6, 7, 27, 46, 162, 164, 168
and 172.)
[Marathe and Taylor, 2013] Marathe, A. R. and Taylor, D. M. (2013). Decoding

continuous limb movements from high-density epidural electrode arrays using
custom spatial filters. Journal of Neural Engineering, 10(3):036015. (Cited on
pages 17, 18, 21, 31, 41, 45, 46, 62, 63, 115 and 207.)
[Marathe and Taylor, 2015] Marathe, a. R. and Taylor, D. M. (2015). The impact
of command signal power distribution, processing delays, and speed scaling on
neurally-controlled devices. Journal of neural engineering, 12(4):046031. (Cited
on pages 6, 8, 64, 78 and 109.)
[Marx, 1996] Marx, B. D. (1996). Iteratively reweighted partial least squares esti-
mation for generalized linear regression. Technometics, 38:374–381. (Cited on
page 175.)
[Mason et al., 2006] Mason, S., Kronegg, J., Huggins, J., Fatourechi, M., and
Schlögl, A. (2006). Evaluating the performance of self-paced brain computer
interface technology. neil squire soc., vancouver, bc. Technical report, Canada,
Tech. Rep. 19. (Cited on pages 35 and 107.)
[Mason and Birch, 2000] Mason, S. G. and Birch, G. E. (2000). A brain-controlled

switch for asynchronous control applications. IEEE transactions on bio-medical
engineering, 47(10):1297–307. (Cited on pages 29, 46, 54, 63 and 75.)
[Maynard et al., 1997] Maynard, E. M., Nordhausen, C. T., and Normann, R. A.

(1997). The Utah Intracortical Electrode Array: A recording structure for potential
brain-computer interfaces. Electroencephalography and Clinical Neurophysiology,
102(3):228–239. (Cited on page 15.)
[McCreadie et al., 2014] McCreadie, K. A., Coyle, D. H., and Prasad, G. (2014).
Is sensorimotor BCI performance influenced differently by mono, stereo, or 3-D
auditory feedback? IEEE Transactions on Neural Systems and Rehabilitation
Engineering, 22(3):431–440. (Cited on page 5.)
[McFarland et al., 2010] McFarland, D. J., Sarnacki, W. A., and Wolpaw, J. R.

(2010). Electroencephalographic (EEG) control of three-dimensional movement.
Journal of neural engineering, 7(3):036007. (Cited on pages 4, 12, 13, 19, 20, 25,
26, 30, 41 and 48.)
Bibliography 237
[McFarland and Wolpaw, 2011] McFarland, D. J. and Wolpaw, J. R. (2011). Brain-

Computer Interfaces for Communication and Control. Communications of the
ACM, 54(5):60–66. (Cited on pages 8, 21 and 22.)
[Mehring et al., 2004] Mehring, C., Nawrot, M. P., de Oliveira, S. C., Vaadia, E.,
Schulze-Bonhage, A., Aertsen, A., and Ball, T. (2004). Comparing information
about arm movement direction in single channels of local and epicortical field
potentials from monkey and human motor cortex. Journal of Physiology-Paris,
98(4-6):498–506. (Cited on page 23.)
[Mehring et al., 2003] Mehring, C., Rickert, J., Vaadia, E., de Oliveira, S. C., Aert-
sen, A., and Rotter, S. (2003). Inference of hand movements from local field
potentials in monkey motor cortex. Nature Neuroscience, 6(12):1253–1254. (Cited
on pages 23, 60 and 175.)
[Meila and Jordan, 1996] Meila, M. and Jordan, M. I. (1996). Learning Fine Motion
by Markov Mixtures of Experts. Advances in Neural Information Processing
Systems 8, (1567):1003–1009. (Cited on pages 75 and 76.)
[Mestais et al., 2015] Mestais, C. S., Charvet, G., Sauter-Starace, F., Foerster, M.,
Ratel, D., and Benabid, A. L. (2015). WIMAGINE: Wireless 64-channel ECoG
recording implant for long term clinical applications. IEEE Transactions on
Neural Systems and Rehabilitation Engineering, 23(1):10–21. (Cited on pages 33
and 176.)
[Milan and Carmena, 2010] Milan, J. D. R. and Carmena, J. (2010). Invasive or

Noninvasive: Understanding Brain-Machine Interface Technology [Conversations
in BME]. IEEE Engineering in Medicine and Biology Magazine, 29(February):16–
[Milekovic et al., 2012] Milekovic, T., Fischer, J., Pistohl, T., Ruescher, J., Schulze-
Bonhage, A., Aertsen, A., Rickert, J., Ball, T., and Mehring, C. (2012). An
online brain-machine interface using decoding of movement direction from the
human electrocorticogram. Journal of neural engineering, 9(4):046003. (Cited on
pages 12, 17, 18, 30, 41 and 42.)
[Millán and Mouriño, 2003] Millán, J. D. R. and Mouriño, J. (2003). Asynchronous

BCI and local neural classifiers: An overview of the adaptive brain interface
project. IEEE Transactions on Neural Systems and Rehabilitation Engineering,
11(2):159–161. (Cited on pages 29, 63 and 75.)
[Miller et al., 2012] Miller, K. J., Hermes, D., Honey, C. J., Hebb, A. O., Ramsey,
N. F., Knight, R. T., Ojemann, J. G., and Fetz, E. E. (2012). Human Motor
Cortical Activity Is Selectively Phase-Entrained on Underlying Rhythms. PLoS
Computational Biology, 8(9). (Cited on page 95.)
238 Bibliography
[Miller and Schalk, 2008] Miller, K. J. and Schalk, G. (2008). Prediction of Finger
Flexion: 4th Brain-Computer Interface Data Competition. BCI Competition IV.
[Moerland, 1997] Moerland, P. (1997). Some methods for training mixtures of

experts. Technical report, IDIAP. (Cited on pages 81 and 123.)
[Moerland, 2000] Moerland, P. (2000). Mixture models for unsupervised and

supervised learning. PhD thesis, Ecole Polytechnique Fédérale de Lausanne.
[Morinière et al., 2015] Morinière, B., Verney, a., Abroug, N., Garrec, P., and Perrot,
Y. (2015). EMY: A dual arm exoskeleton dedicated to the evaluation of Brain
Machine Interface in clinical trials. IEEE International Conference on Intelligent
Robots and Systems, 2015-Decem:5333–5338. (Cited on pages 34 and 176.)
[Morshed and Khan, 2014] Morshed, B. I. and Khan, A. (2014). Biomedical Science
A Brief Review of Brain Signal Monitoring Technologies for BCI Applications
: Challenges and Prospects. Journal of Bioengineering & Biomedical Science,
[Müller et al., 2003] Müller, K.-R., Anderson, C. W., and Birch, G. E. (2003). Linear
and nonlinear methods for brain-computer interfaces. IEEE transactions on neural
systems and rehabilitation engineering, 11(2):165–169. (Cited on pages 52 and 55.)
[Muller, 2007] Muller, M. (2007). Dynamic Time Warping. Information retrieval

for music and motion, pages 69–84. (Cited on page 109.)
[Müller et al., 2011] Müller, S. M. T., Bastos-Filho, T. F., and Sarcinelli-Filho, M.

(2011). Using a SSVEP-BCI to command a robotic wheelchair. In Industrial
Electronics (ISIE), 2011 IEEE International Symposium on, number 1-2, pages
957–962. (Cited on page 2.)
[Müller-Putz et al., 2010] Müller-Putz, G. R., Kaiser, V., Solis-Escalante, T., and
Pfurtscheller, G. (2010). Fast set-up asynchronous brain-switch based on detection
of foot motor imagery in 1-channel EEG. Medical & Biological Engineering &
Computing, 48(3):229–233. (Cited on page 46.)
[Murguialday et al., 2007] Murguialday, A. R., Aggarwal, V., Chatterjee, A., Cho,
Y., Rasmussen, R., O’Rourke, B., Acharya, S., and Thakor, N. V. (2007). Brain-
Computer Interface for a Prosthetic Hand Using Local Machine Control and
Haptic Feedback. 2007 IEEE 10th International Conference on Rehabilitation
Robotics, 00(c):609–613. (Cited on pages 29, 30 and 175.)
[Murphy, 1998] Murphy, K. (1998). Switching Kalman Filters. Technical Report

August. (Cited on pages 104, 202, 203, 204, 205 and 206.)
Bibliography 239
[Murphy, 2002] Murphy, K. (2002). Dynamic Bayesian Networks: Representation,

Inference and Learning. Doctoral dissertation, University of California, Berkeley.
(Cited on page 57.)
[Murphy et al., 2015] Murphy, M. D., Guggenmos, D. J., Bundy, D. T., and Nudo,
R. J. (2015). Current Challenges Facing the Translation of Brain Computer
Interfaces from Preclinical Trials to Use in Human Patients. Frontiers in cellular
neuroscience, 9(January):497. (Cited on page 17.)
[Muthukumaraswamy, 2013] Muthukumaraswamy, S. D. (2013). High-frequency

brain activity and muscle artifacts in MEG/EEG: a review and recommendations.
Frontiers in Human Neuroscience, 7(April):1–11. (Cited on pages 20, 39 and 40.)
[Nakanishi et al., 2014a] Nakanishi, M., Wang, Y.-T. Y., Mitsukura, Y., and Jung,
T.-P. (2014a). A high-speed brain speller using steady-state visual evoked poten-
tials. International journal of neural systems, 24(6):1450019. (Cited on page 2.)
[Nakanishi et al., 2014b] Nakanishi, Y., Yanagisawa, T., Shin, D., Chen, C., Kam-
bara, H., Yoshimura, N., Fukuma, R., Kishima, H., Hirata, M., and Koike,
Y. (2014b). Decoding fingertip trajectory from electrocorticographic signals in
humans. Neuroscience research, 85:20–7. (Cited on pages 31 and 66.)
[Nakanishi et al., 2013] Nakanishi, Y., Yanagisawa, T., Shin, D., Fukuma, R., Chen,
C., Kambara, H., Yoshimura, N., Hirata, M., Yoshimine, T., and Koike, Y. (2013).
Prediction of three-dimensional arm trajectories based on ECoG signals recorded
from human sensorimotor cortex. PloS one, 8(8):e72085. (Cited on pages 63
and 162.)
[Nakayama and Inagaki, 2006] Nakayama, K. and Inagaki, K. (2006). A Brain

Computer Interface Based on Neural Network with Efficient Pre-Processing.
Simulation, (L):6–9. (Cited on pages 52 and 54.)
[Naseer et al., 2014] Naseer, N., Hong, M. J., and Hong, K. S. (2014). Online binary
decision decoding using functional near-infrared spectroscopy for the development
of brain-computer interface. Experimental Brain Research, 232(2):555–564. (Cited
on page 4.)
[Navarro et al., 2005] Navarro, I., Hubais, B., and Sepulveda, F. (2005). A Compar-
ison of Time, Frequency and ICA Based Features and Five Classifiers for Wrist
Movement Classification in EEG Signals. Conference proceedings : ... Annual
International Conference of the IEEE Engineering in Medicine and Biology
Society. IEEE Engineering in Medicine and Biology Society. Conference, 2:2118–
2121. (Cited on pages 52, 53 and 54.)
[Nazarpour et al., 2009] Nazarpour, K., Praamstra, P., Miall, R. C., and Sanei, S.
(2009). Steady-state movement related potentials for brain-computer interfac-
ing. IEEE transactions on bio-medical engineering, 56(8):2104–2113. (Cited on
page 13.)
240 Bibliography
[Ng and Jordan, 2002] Ng, A. Y. and Jordan, M. I. (2002). On discriminative

vs. generative classifiers: A comparison of logistic regression and naive bayes.
Advances in neural information processing systems, pages 841–848. (Cited on
pages 50 and 51.)
[Nguyen and Rocke, 2002] Nguyen, D. V. and Rocke, D. M. (2002). Multi-class

cancer classification via partial least squares with gene expression profiles.
Bioinformatics (Oxford, England), 18(9):1216–1226. (Cited on pages 195 and 199.)
[Nguyen et al., 2013] Nguyen, H. T., Trung, N., Toi, V., and Tran, V. S. (2013).
An autoregressive neural network for recognition of eye commands in an EEG-
controlled wheelchair. International Conference on Advanced Technologies for
Communications, pages 333–338. (Cited on page 2.)
[Nicolas-Alonso and Gomez-Gil, 2012] Nicolas-Alonso, L. F. and Gomez-Gil, J.

(2012). Brain computer interfaces, a review. Sensors (Basel, Switzerland),
[Nicolelis, 2014] Nicolelis, M. a. L. (2014). Brain-to-Brain Interfaces: When Reality

Meets Science Fiction. Cerebrum : the Dana forum on brain science, (3):13.
(Cited on page 29.)
[Nurse et al., 2015] Nurse, E. S., Freestone, D. R., Oxley, T. J., Ackland, D. C.,
Vogrin, S. J., Murphy, M., Brien, T. J. O., Cook, M. J., and Grayden, D. B. (2015).
Movement Related Directional Tuning from Broadband Electrocorticography in
Humans. pages 22–24. (Cited on pages 18, 24, 26, 27, 28, 39 and 115.)
[Obermaier et al., 2001] Obermaier, B., Guger, C., Neuper, C., and Pfurtscheller,
G. (2001). Hidden Markov models for online classification of single trial EEG data.
Pattern Recognition Letters, 22(12):1299–1309. (Cited on pages 26, 56 and 57.)
[O’Doherty et al., 2011] O’Doherty, J. E., Lebedev, M. a., Ifft, P. J., Zhuang, K. Z.,
Shokur, S., Bleuler, H., and Nicolelis, M. a. L. (2011). Active tactile exploration
using a brain–machine–brain interface. Nature, 479(7372):228–231. (Cited on
pages 5 and 30.)
[Ofner and Müller-Putz, 2012] Ofner, P. and Müller-Putz, G. R. (2012). Decoding of

velocities and positions of 3D arm movement from EEG. Conference proceedings
: Annual International Conference of the IEEE Engineering in Medicine and
Biology Society. IEEE Engineering in Medicine and Biology Society. Conference,
2012:6406–9. (Cited on page 24.)
[Onaran et al., 2011] Onaran, I., Ince, N. F., Cetin, a. E., and Abosch, A. (2011).
A hybrid SVM/HMM based system for the state detection of individual finger
movements from multichannel ECoG signals. 2011 5th International IEEE/EMBS
Conference on Neural Engineering, NER 2011, pages 457–460. (Cited on pages 41,
56 and 57.)
Bibliography 241
[Onose et al., 2012] Onose, G., Grozea, C., Anghelescu, A., , and others Daia, C.
(2012). On the feasibility of using motor imagery EEG-based brain–computer
interface in chronic tetraplegics for assistive robotic arm control: a clinical test and
long-term post trial follow-up. Spinal Cord, 50(9):716–716. (Cited on pages 19,
20, 29, 40, 41 and 45.)
[Ordóñez et al., 2013] Ordóñez, F. J., de Toledo, P., and Sanchis, A. (2013). Activity
Recognition Using Hybrid Generative/Discriminative Models on Home Environ-
ments Using Binary Sensors. Sensors, 13(5):5460–5477. (Cited on pages 77
and 118.)
[Orsborn et al., 2011] Orsborn, A. L., Dangi, S., Moorman, H. G., and Carmena,
J. M. (2011). Exploring time-scales of closed-loop decoder adaptation in brain-
machine interfaces. In Proceedings of the Annual International Conference of
[Orsborn et al., 2012] Orsborn, A. L., Dangi, S., Moorman, H. G., and Carmena,
J. M. (2012). Closed-loop decoder adaptation on intermediate time-scales facilitates
rapid BMI performance improvements independent of decoder initialization con-
ditions. IEEE Transactions on Neural Systems and Rehabilitation Engineering,
[Ortner et al., 2011] Ortner, R., Allison, B. Z., Korisek, G., Gaggl, H., and
Pfurtscheller, G. (2011). An SSVEP BCI to control a hand orthosis for per-
sons with tetraplegia. IEEE Transactions on Neural Systems and Rehabilitation
Engineering, 19(1):1–5. (Cited on pages 2, 11, 12, 14 and 29.)
[Ozkan et al., 2012] Ozkan, H., Akman, A., and Kozat, S. S. (2012). A novel
training algorithm for HMMs with partial and noisy access to the states. arXiv
preprint, (March 2012). (Cited on page 176.)
[Palankar et al., 2008] Palankar, M., De Laurentis, K. J., Alqasemi, R., Veras, E.,
Dubey, R., Arbel, Y., and Donchin, E. (2008). Control of a 9-DoF wheelchair-
mounted robotic arm system using a P300 brain computer interface: Initial
experiments. 2008 IEEE International Conference on Robotics and Biomimetics,
ROBIO 2008, pages 348–353. (Cited on pages 2, 12, 14 and 42.)
[Paralikar, 2010] Paralikar, K. J. (2010). New approaches to eliminating common-

noise artifacts in recordings from intracortical microelectrode arrays: inter-
electrode correlation and virtual referencing. Journal of Neuroscience, 181(1):27–
[Parsons et al., 2004] Parsons, L., Parsons, L., Haque, E., Haque, E., Liu, H., and
Liu, H. (2004). Subspace clustering for high dimensional data. ACM SIGKDD
Explorations Newsletter, 6(1):90–105. (Cited on page 121.)
242 Bibliography
[Penny and Roberts, 1999] Penny, W. and Roberts, S. (1999). Experiments with
an eeg-based computer interface. Technical report, Technical report, Department
of Electrical and Electronic Engineering, Imperial College, London. (Cited on
page 25.)
[Penny et al., 2000] Penny, W. D., Roberts, S. J., Curran, E. A., and Stokes, M. J.
(2000). EEG-based communication: a pattern recognition approach. IEEE
transactions on Rehabilitation Engineering, 8(2):214–215. (Cited on pages 13, 14
and 53.)
[Pereira et al., 2015] Pereira, D. G., Afonso, A., and Medeiros, F. M. (2015).
Overview of Friedman’s test and post-hoc analysis. Communications in
Statistics-Simulation and Computation, 44(10):2636–2653. (Cited on page 112.)
[Perruchoud et al., 2016] Perruchoud, D., Pisotta, I., Carda, S., Murray, M. M.,
and Ionta, S. (2016). Biomimetic rehabilitation engineering: the importance of so-
matosensory feedback for brain-machine interfaces. Journal of neural engineering,
13(4):041001. (Cited on page 5.)
[Pfurtscheller et al., 2000] Pfurtscheller, G., Guger, C., Müller, G., Krausz, G., and
Neuper, C. (2000). Brain oscillations control hand orthosis in a tetraplegic.
Neuroscience Letters, 292(3):211–214. (Cited on pages 3, 12, 13, 19, 20, 29, 42,
46, 50 and 52.)
[Pfurtscheller et al., 2010] Pfurtscheller, G., Solis-Escalante, T., Ortner, R., Linort-
ner, P., and Muller-Putz, G. R. (2010). Self-paced operation of an SSVEP-based
orthosis with and without an imagery-based "brain switch": A feasibility study
towards a hybrid BCI. IEEE Transactions on Neural Systems and Rehabilitation
[Phatak and De Hoog, 2002] Phatak, A. and De Hoog, F. (2002). Exploiting the
connection between PLS, Lanczos methods and conjugate gradients: Alternative
proofs of some properties of PLS. Journal of Chemometrics, 16(7):361–367. (Cited
on page 192.)
[Pistohl et al., 2008] Pistohl, T., Ball, T., Schulze-Bonhage, A., Aertsen, A., and
Mehring, C. (2008). Prediction of arm movement trajectories from ECoG-
recordings in humans. Journal of neuroscience methods, 167(1):105–14. (Cited
on pages 24, 25, 26, 27, 28, 42, 43, 62, 63, 115, 116, 120, 201, 206, 207 and 209.)
[Polikov et al., 2005] Polikov, V. S., Tresco, P. a., and Reichert, W. M. (2005).
Response of brain tissue to chronically implanted neural electrodes. Journal of
neuroscience methods, 148(1):1–18. (Cited on page 16.)
[Pourbakhtiar et al., 2013] Pourbakhtiar, A., Shamsi, M., and Farrokhshad, F.

(2013). Neuro-fuzzy classification of brain computer interface data using phase
based feature. 13th Iranian Conference on Fuzzy Systems, IFSC 2013, pages 3–6.
(Cited on page 45.)
Bibliography 243
[Qian et al., 2013] Qian, J., Hastie, T., Friedman, J., Tibshirani, R., and Simon, N.
(2013). Glmnet for Matlab. https://web.stanford.edu/˜hastie/glmnet_matlab/.
[Quandt, 1958] Quandt, R. E. (1958). The Estimation of the Parameters of a Linear

Regression System Obeying Two Separate Regimes. Journal of the American
Statistical Association, 53(284):873–880. (Cited on page 70.)
[Quandt and Ramsey, 1978] Quandt, R. E. and Ramsey, J. B. (1978). Estimat-

ing mixtures of normal distributions and switching regressions. Journal of the
American Statistical Association, 73(364):730. (Cited on page 70.)
[Quinn, 2005] Quinn, J. (2005). Switching Kalman Filter Matlab implementation.

http://air.ug/˜jquinn/downloads/. (Cited on page 204.)
[Rabiner, 1989] Rabiner, L. (1989). A tutorial on hidden Markov models and selected
applications in speech recognition. Proceedings of the IEEE, 77(2):257–286. (Cited
on pages 56, 57, 77, 82, 84, 86, 176 and 184.)
[Rabiner and Juang, 1986] Rabiner, L. and Juang, B. (1986). An introduction to

hidden Markov models. IEEE ASSP Magazine, 3(January):Appendix 3A. (Cited
on page 76.)
[Rak et al., 2012] Rak, R. J., Kołodziej, M., and Majkowski, A. (2012). Brain-
computer interface as measurement and control system - the review paper.
Metrology and Measurement Systems, XIX(3):427–444. (Cited on page 4.)
[Ramsey et al., 2006] Ramsey, N. F., Heuvel, M. P. V. D., Kho, K. H., and Leijten,
F. S. S. (2006). Towards Human BCI Applications Based on Cognitive Brain
Systems: An Investi- gation of Neural Signals Recorded From the Dorsolateral
Prefrontal Cortex. IEEE Transactions on Neural Systems and Rehabilitation,
[Rebsamen et al., 2006] Rebsamen, B., Burdet, E., Guan, C., Zhang, H., Teo, C. L.,
Zeng, Q., Ang, M., and Laugier, C. (2006). A brain-controlled wheelchair
based on P300 and path guidance. Proceedings of the First IEEE/RAS-EMBS
International Conference on Biomedical Robotics and Biomechatronics, 2006,
BioRob 2006, 2006:1101–1106. (Cited on page 2.)
[Renals et al., 1994] Renals, S., Morgan, N., Bourlard, H., Cohen, M., and Franco,
H. (1994). Connectionist Probability Estimation in HMM Speech Recognition.
IEEE Transactions on Speech and Audio Processing, 2(1):161–174. (Cited on
pages 77 and 118.)
[Robotnik, 2016] Robotnik (2016). KINOVA JACO arm.

http://www.robotnik.eu/robotics-arms/kinova-jaco-arm/. (Cited on page 29.)
[Roche, 2011] Roche, A. (2011). EM algorithm and variants: an informal tutorial.

arXiv preprint arXiv:1105.1476, pages 1–17. (Cited on page 121.)
244 Bibliography
[Rosipal and Krämer, 2006] Rosipal, R. and Krämer, N. (2006). Overview and
Recent Advances in Partial Least Squares. Subspace, Latent Structure and
Feature Selection, Saunders, C., et al. (eds.) (Heidelberg: Springer-Verlag, 2006),
3940:34–51. (Cited on pages 195 and 207.)
[Rouse and Moran, 2009] Rouse, A. G. and Moran, D. W. (2009). Neural adapta-
tion of epidural electrocorticographic (EECoG) signals during closed-loop brain
computer interface (BCI) tasks. Conference proceedings : ... Annual International
[Rouse and Schieber, 2015] Rouse, A. G. and Schieber, M. H. (2015). Advancing

brain-machine interfaces: moving beyond linear state space models. Frontiers in
Systems Neuroscience, 9(July):1–13. (Cited on pages 69 and 159.)
[Rouse et al., 2013] Rouse, A. G., Williams, J. J., Wheeler, J. J., and Moran, D. W.
(2013). Cortical Adaptation to a Chronic Micro- Electrocorticographic Brain
Computer Interface. Journal of Neuroscience, 33(4):1326–1330. (Cited on pages 17
and 18.)
[Rouse et al., 2016] Rouse, A. G., Williams, J. J., Wheeler, J. J., and Moran,
D. W. (2016). Spatial co-adaptation of cortical control columns in a micro-ECoG
brain–computer interface. Journal of Neural Engineering, 13(5):056018. (Cited
[Royer et al., 2010] Royer, A. S., Doud, A. J., Rose, M. L., and He, B. (2010). EEG
control of a virtual helicopter in 3-dimensional space using intelligent control
strategies. IEEE transactions on neural systems and rehabilitation engineering : a
publication of the IEEE Engineering in Medicine and Biology Society, 18(6):581–
[Ryali et al., 2010] Ryali, S., Supekar, K., Abrams, D. A., and Menon, V. (2010).
Sparse logistic regression foro whole brain classification of fMRI data. NeuroImage,
51(2):752–764. (Cited on page 194.)
[Saa and Çetin, 2013] Saa, J. F. D. and Çetin, M. (2013). Discriminative Methods
for Classi fi cation of Asynchronous Imaginary Motor Tasks From EEG Data. Ieee
Transactions on Neural Systems and Rehabilitation Engineering, 21(5):716–724.
[Saa et al., 2016] Saa, J. F. D., de Pesters, A., and Cetin, M. (2016). Asynchronous
decoding of finger movements from ECoG signals using long-range dependencies
conditional random fields. Journal of Neural Engineering, 13(3):36017. (Cited on
pages 31, 56, 57, 58, 74, 75 and 165.)
[Sadeghian and Moradi, 2007] Sadeghian, E. B. and Moradi, M. H. (2007). Con-

tinuous detection of motor imagery in a four-class asynchronous BCI. Annual
Bibliography 245
International Conference of the IEEE Engineering in Medicine and Biology -

Proceedings, pages 3241–3244. (Cited on pages 39, 41, 52 and 53.)
[Sadtler, 2014] Sadtler, P. T. (2014). Investigating the Neural Basis of Learning

Using Brain-Computer Interfaces. PhD thesis, University of Pittsburgh. (Cited
on page 207.)
[Sadtler et al., 2014] Sadtler, P. T., Quick, K. M., Golub, M. D., Chase, S. M.,
Ryu, S. I., Tyler-Kabara, E. C., Yu, B. M., and Batista, A. P. (2014). Neural
constraints on learning. Nature, 512(7515):423–426. (Cited on pages 15, 16, 24
and 25.)
[Sanchez et al., 2002] Sanchez, J. C., Erdogmus, D., Principe, J. C., Wessberg, J.,
and Nicolelis, M. (2002). A comparison between nonlinear mappings and linear
state estimation to model the relation from motor cortical neuronal firing to hand
movements. Proceedings of SAB Workshop on Motor Control in Humans and
Robots: On the Interplay of Real Brains and Artificial Devices, (August):59–65.
(Cited on page 60.)
[Sanchez et al., 2008] Sanchez, J. C., Gunduz, A., Carney, P. R., and Principe, J. C.
(2008). Extraction and localization of mesoscopic motor control signals for human
ECoG neuroprosthetics. Journal of Neuroscience Methods, 167(1):63–81. (Cited
on page 17.)
[Schafer, 2011] Schafer, R. (2011). What Is a Savitzky-Golay Filter?: [Lecture

Notes]. IEEE Signal Process. Mag., 28(July):111–117. (Cited on page 116.)
[Schalk et al., 2007] Schalk, G., Kubánek, J., Miller, K. J., Anderson, N. R.,
Leuthardt, E. C., Ojemann, J. G., Limbrick, D., Moran, D., Gerhardt, L. a.,
and Wolpaw, J. R. (2007). Decoding two-dimensional movement trajectories
using electrocorticographic signals in humans. Journal of Neural Engineering,
4(3):264–275. (Cited on pages 17, 26, 27, 28, 41, 42, 44, 48, 63, 78, 95, 115
and 162.)
[Schalk and Leuthardt, 2011] Schalk, G. and Leuthardt, E. C. (2011). Brain-

computer interfaces using electrocorticographic signals. IEEE Reviews in
Biomedical Engineering, 4:140–154. (Cited on pages 12, 15, 17, 18 and 20.)
[Schalk et al., 2008] Schalk, G., Miller, K. J., Anderson, N. R., Wilson, J. a., Smyth,
M. D., Ojemann, J. G., Moran, D. W., Wolpaw, J. R., and Leuthardt, E. C.
(2008). Two-dimensional movement control using electrocorticographic signals in
humans. Journal of neural engineering, 5(1):75–84. (Cited on pages 17, 18, 26,
27, 31, 44 and 60.)
[Scheffer et al., 2001] Scheffer, T., Decomain, C., and Wrobel, S. (2001). Active
hidden markov models for information extraction. Advances in Intelligent Data
Analysis, pages 309–318. (Cited on page 176.)
246 Bibliography
[Scheid et al., 2013] Scheid, M. R., Flint, R. D., Wright, Z. a., and Slutzky, M. W.
(2013). Long-term, stable behavior of local field potentials during brain machine
interface use. Proceedings of the Annual International Conference of the IEEE
Engineering in Medicine and Biology Society, EMBS, pages 307–310. (Cited on
page 60.)
[Scherer et al., 2008] Scherer, R., Lee, F., Schl??gl, A., Leeb, R., Bischof, H., and
Pfurtscheller, G. (2008). Toward self-paced brain-computer communication:
Navigation through virtual worlds. IEEE Transactions on Biomedical Engineering,
55(2):675–682. (Cited on page 48.)
[Schlögl et al., 2005] Schlögl, A., Lee, F., Bischof, H., and Pfurtscheller, G. (2005).
Characterization of Four-Class Motor Imagery EEG Data for the BCI-Competition
2005. Journal of Neural Engineering, 2(4):L14–L22. (Cited on pages 44, 51, 52,
53, 54 and 55.)
[Schwartz et al., 2006] Schwartz, A. B., Cui, X. T., Weber, D. J., and Moran, D. W.
(2006). Brain-controlled interfaces: movement restoration with neural prosthetics.
Neuron, 52(1):205–20. (Cited on page 3.)
[Schwarz et al., 2014] Schwarz, D. a., Lebedev, M. a., Hanson, T. L., Dimitrov,
D. F., Lehew, G., Meloy, J., Rajangam, S., Subramanian, V., Ifft, P. J., Li,
Z., Ramakrishnan, A., Tate, A., Zhuang, K. Z., and Nicolelis, M. a. L. (2014).
Chronic, wireless recordings of large-scale brain activity in freely moving rhesus
monkeys. Nature methods, 11(6):670–6. (Cited on page 15.)
[Scott, 2008] Scott, S. H. (2008). Inconvenient truths about neural processing in

primary motor cortex. The Journal of physiology, 586(5):1217–24. (Cited on
pages 23 and 59.)
[Sellers et al., 2006] Sellers, E. W., Kübler, A., and Donchin, E. (2006). Brain-
computer interface research at the University of South Florida cognitive psy-
chophysiology laboratory: The P300 speller. IEEE Transactions on Neural
Systems and Rehabilitation Engineering, 14(2):221–224. (Cited on page 2.)
[Shanechi et al., 2013] Shanechi, M. M., Williams, Z. M., Wornell, G. W., Hu,
R. C., Powers, M., and Brown, E. N. (2013). A real-time brain-machine interface
combining motor target and trajectory intent using an optimal feedback control
design. PloS one, 8(4):e59049. (Cited on pages 59 and 189.)
[Shenoy and Carmena, 2014] Shenoy, K. V. and Carmena, J. M. (2014). Combin-

ing decoder design and neural adaptation in brain-machine interfaces. Neuron,
84(4):665–80. (Cited on page 23.)
[Shenoy, 2005] Shenoy, P. (2005). Dynamic Bayesian Networks for Brain-Computer

Interfaces. Advances in Neural Information Processing Systems, 17:1265–72.
(Cited on page 57.)
Bibliography 247
[Shenoy et al., 2008] Shenoy, P., Miller, K. J., Ojemann, J. G., and Rao, R. P. N.
(2008). Generalized features for electrocorticographic BCIs. IEEE Transactions
on Biomedical Engineering, 55(1):273–280. (Cited on page 56.)
[Shimoda et al., 2012] Shimoda, K., Nagasaka, Y., Chao, Z. C., and Fujii, N. (2012).
Decoding continuous three-dimensional hand trajectories from epidural electro-
corticographic signals in Japanese macaques. Journal of Neural Engineering,
9(3):036015. (Cited on pages 17, 19, 24, 39, 43, 44, 47, 59, 63, 64, 89, 91, 115,
116, 162, 188 and 195.)
[Shin et al., 2012] Shin, D., Watanabe, H., Kambara, H., Nambu, A., Isa, T.,
Nishimura, Y., and Koike, Y. (2012). Prediction of Muscle Activities from Elec-
trocorticograms in Primary Motor Cortex of Primates. PLoS ONE, 7(10):e47992.
[Simeral et al., 2011] Simeral, J. D., Kim, S.-P., Black, M. J., Donoghue, J. P., and
Hochberg, L. R. (2011). Neural control of cursor trajectory and click by a human
with tetraplegia 1000 days after implant of an intracortical microelectrode array.
Journal of neural engineering, 8(2):025027. (Cited on pages 15, 16, 17, 22 and 30.)
[Sing et al., 2007] Sing, C., Tsui, L., and Gan, J. Q. (2007). Asynchronous BCI
Control of a Robot Simulator with Supervised Online Training. Intelligent Data
Engineering and Automated Learning - IDEAL 2007, pages 125–134. (Cited on
pages 49, 50 and 55.)
[Slutzky et al., 2010] Slutzky, M. W., Jordan, L. R., Krieg, T., Chen, M., Mogul,
D. J., and Miller, L. E. (2010). Optimal Spacing of Surface Electrode Arrays for
Brain Machine Interface Applications. Journal of neural engineering, 7(2):26004.
(Cited on page 18.)
[Souza and Araújo, 2014] Souza, F. a. a. and Araújo, R. (2014). Mixture of partial
least squares experts and application in prediction settings with multiple operating
modes. Chemometrics and Intelligent Laboratory Systems, 130:192–202. (Cited
on page 123.)
[Spüler et al., 2016] Spüler, M., Grimm, F., Gharabaghi, A., Bogdan, M., and Rosen-
stiel, W. (2016). Comparing Methods for Decoding Movement Trajectory from
ECoG in Chronic Stroke Patients. In Advances in Neurotechnology, Electronics
and Informatics, pages 125–139. Springer. (Cited on pages 39, 44, 48, 60, 66, 115,
188, 189 and 192.)
[Spuler et al., 2015] Spuler, M., Sarasola-Sanz, A., Birbaumer, N., Rosenstiel, W.,
and Ramos-Murguialday, A. (2015). Comparing metrics to evaluate perfor-
mance of regression methods for decoding of neural signals. In Proceedings
of the Annual International Conference of the IEEE Engineering in Medicine
and Biology Society, EMBS, volume 2015-Novem, pages 1083–1086. (Cited on
page 108.)
248 Bibliography
[Squire et al., 2013] Squire, L., Berg, D., Bloom, F. E., Du Lac, S., Ghosh, A., and
Spitzer, N. C. (2013). Fundamental Neuroscience. Academic Press. (Cited on
pages 9 and 16.)
[Srinivasan et al., 2007] Srinivasan, L., Eden, U. T., Mitter, S. K., and Brown, E. N.
(2007). General-purpose filter design for neural prosthetic devices. Journal of
Neurophysiology, 98(4):2456–2475. (Cited on pages 29, 62, 65, 70, 71, 74, 76, 104
and 172.)
[Stavisky et al., 2015] Stavisky, S. D., Kao, J. C., Sorokin, J. M., Ryu, S. I., and
Shenoy, K. V. (2015). System identification of brain-machine interface control
using a cursor jump perturbation. International IEEE/EMBS Conference on
Neural Engineering, NER, 2015-July:643–647. (Cited on page 78.)
[Suk and Lee, 2010] Suk, H. I. and Lee, S. W. (2010). Two-layer hidden Markov
models for multi-class motor imagery classification. Proceedings - Workshop on
Brain Decoding: Pattern Recognition Challenges in Neuroimaging, WBD 2010 -
In Conjunction with theInternational Conference on Pattern Recognition, ICPR
2010, pages 5–8. (Cited on pages 47 and 195.)
[Suminski et al., 2010] Suminski, a. J., Tkach, D. C., Fagg, a. H., and Hatsopoulos,
N. G. (2010). Incorporating Feedback from Multiple Sensory Modalities Enhances
Brain-Machine Interface Control. Journal of Neuroscience, 30(50):16777–16787.
[Sutton and Mccallum, 2012] Sutton, C. and Mccallum, A. (2012). An Introduction

to Conditional Random Fields. Foundations and Trends¥ R in Machine Learning
in Machine Learning, 4(4):267–373. (Cited on pages 51 and 57.)
[Suway et al., 2013] Suway, S. B., Tien, R. N., Jeffries, S. M., Zohny, Z., Clanton,
S. T., McMorland, A. J. C., and Velliste, M. (2013). Resting state detection
for gating movement of a neural prosthesis. 2013 6th International IEEE/EMBS
Conference on Neural Engineering (NER), pages 665–668. (Cited on pages 65,
77, 78 and 175.)
[Swinnen and Wenderoth, 2004] Swinnen, S. P. and Wenderoth, N. (2004). Two

hands, one brain: cognitive neuroscience of bimanual skill. Trends in Cognitive
Sciences, 8(1):18–25. (Cited on pages 31 and 177.)
[Tangermann et al., 2012] Tangermann, M., Müller, K. R., Aertsen, A., Birbaumer,
N., Braun, C., Brunner, C., Leeb, R., Mehring, C., Miller, K. J., Müller-Putz,
G. R., Nolte, G., Pfurtscheller, G., Preissl, H., Schalk, G., Schlögl, A., Vidaurre,
C., Waldert, S., and Blankertz, B. (2012). Review of the BCI competition IV.
Frontiers in Neuroscience, 6(JULY):1–31. (Cited on page 95.)
[Taylor et al., 2002] Taylor, D. M., Tillery, S. I. H., and Schwartz, A. B. (2002).
Direct cortical control of 3D neuroprosthetic devices. Science (New York, N.Y.),
296(5574):1829–32. (Cited on pages 15, 16, 30, 59 and 60.)
Bibliography 249
[Teplan, 2002] Teplan, M. (2002). Fundamentals of EEG measurement.

Measurement Science Review, 2:1–11. (Cited on pages 4, 19 and 20.)
[Thomas et al., 2013] Thomas, E., Dyson, M., and Clerc, M. (2013). An analysis
of performance evaluation for motor-imagery based BCI. Journal of Neural
Engineering, 10(3):031001. (Cited on page 107.)
[Tillery et al., 2003] Tillery, S. I. H., Taylor, D. M., and Schwartz, A. B. (2003).
Training in cortical control of neuroprosthetic devices improves signal extraction
from small neuronal ensembles. Reviews in the neurosciences, 14(1-2):107–19.
(Cited on page 22.)
[Tomioka et al., 2007] Tomioka, R., Aihara, K., and Müller, K.-R. (2007). Logistic
regression for single trial EEG classification. Analysis, 19:1377–1384. (Cited on
pages 52 and 53.)
[Townsend et al., 2004] Townsend, G., Graimann, B., and Pfurtscheller, G. (2004).
Continuous EEG classification during motor imagery - Simulation of an asyn-
chronous BCI. IEEE Transactions on Neural Systems and Rehabilitation
Engineering, 12(2):258–265. (Cited on pages 41, 63, 64 and 75.)
[Tranquillo, 2013] Tranquillo, J. V. (2013). Biomedical Signals and Systems.

Synthesis Lectures on Biomedical Engineering, 8(3):1–233. (Cited on page 108.)
[Trejo et al., 2006] Trejo, L. J., Rosipal, R., and Matthews, B. (2006). Brain-
computer interfaces for 1-D and 2-D cursor control: Designs using volitional
control of the EEG spectrum or steady-state visual evoked potentials. IEEE
Transactions on Neural Systems and Rehabilitation Engineering, 14(2):225–229.
[Turkmen and Billor, 2012] Turkmen, A. and Billor, N. (2012). Partial least
squares classification for high dimensional data using the PCOUT algorithm.
Computational Statistics, 28(2):771–788. (Cited on page 195.)
[Tusell, 2011] Tusell, F. (2011). Kalman filtering in R. Journal Of Statistical

Software, 39(2). (Cited on page 172.)
[Unicare, 2017] Unicare (2017). Unicare EEG cap.

http://www.eeg-sensor.com/show.php?id=56. (Cited on page 20.)
[Urigüen and Garcia-Zapirain, 2015] Urigüen, J. A. and Garcia-Zapirain, B. (2015).

EEG artifact removal-state-of-the-art and guidelines. Journal of Neural
Engineering, 12(3):31001. (Cited on page 40.)
[Valstar and Pantic, 2007] Valstar, M. F. and Pantic, M. (2007). Combined Support
Vector Machines and Hidden Markov Models for Modeling Facial Action Temporal
Dynamics. In International Workshop on Human-Computer Interaction, pages
118–127. (Cited on pages 77 and 118.)
250 Bibliography
[van Gerven et al., 2012] van Gerven, M. a. J., Chao, Z. C., and Heskes, T. (2012).
On the decoding of intracranial data using sparse orthonormalized partial least
squares. Journal of Neural Engineering, 9(2):026017. (Cited on pages 47, 59
and 195.)
[Vansteensel et al., 2010] Vansteensel, M. J., Hermes, D., Aarnoutse, E. J., Ble-
ichner, M. G., Schalk, G., Van Rijen, P. C., Leijten, F. S. S., and Ramsey,
N. F. (2010). Brain-computer interfacing based on cognitive control. Annals of
Neurology, 67:809–816. (Cited on pages 13, 14, 17, 18, 30 and 31.)
[Velliste et al., 2014] Velliste, M., Kennedy, S. D., Schwartz, A. B., Whitford, A. S.,
Sohn, J.-W., and McMorland, A. J. C. (2014). Motor cortical correlates of arm
resting in the context of a reaching task and implications for prosthetic control.
The Journal of neuroscience : the official journal of the Society for Neuroscience,
34(17):6011–22. (Cited on pages 29, 52, 62, 64, 65, 66, 70, 102 and 159.)
[Velliste et al., 2008] Velliste, M., Perel, S., Spalding, M. C., Whitford, A. S., and
Schwartz, A. B. (2008). Cortical control of a prosthetic arm for self-feeding.
Nature, 453(7198):1098–101. (Cited on pages 12, 15, 16, 29, 59 and 60.)
[Vidaurre et al., 2016] Vidaurre, C., Klauer, C., Schauer, T., Ramos-Murguialday,
A., and Müller, K.-R. (2016). EEG-based BCI for the linear control of an upper-
limb neuroprosthesis. Medical Engineering & Physics, 0:1–10. (Cited on pages 29,
40, 45, 50, 52 and 55.)
[Vidaurre et al., 2009] Vidaurre, C., Krämer, N., Blankertz, B., and Schlögl, A.
(2009). Time Domain Parameters as a feature for EEG-based Brain-Computer
Interfaces. Neural Networks, 22(9):1313–1319. (Cited on page 45.)
[Vigário, 1997] Vigário, R. N. (1997). Extraction of ocular artefacts from EEG

using independent component analysis. Electroencephalography and Clinical
Neurophysiology, 103(3):395–404. (Cited on page 40.)
[Vouga et al., 2017] Vouga, T., Zhuang, K., Olivier, J., Lebedev, M., Nicolelis, M.,
Bouri, M., and Bleuler, H. (2017). EXiO - A Brain-Controlled Lower Limb
Exoskeleton for Rhesus Macaques. IEEE Transactions on Neural Systems and
Rehabilitation Engineering, 25(2):1–1. (Cited on page 31.)
[Waldert et al., 2009] Waldert, S., Pistohl, T., Braun, C., Ball, T., Aertsen, A., and
Mehring, C. (2009). A review on directional information in neural signals for
brain-machine interfaces. Journal of physiology, Paris, 103(3-5):244–54. (Cited
on pages 2, 3, 13, 14, 15, 18, 19, 20, 23, 25, 26, 27 and 41.)
[Waldert et al., 2008] Waldert, S., Preissl, H., Demandt, E., Braun, C., Birbaumer,
N., Aertsen, a., and Mehring, C. (2008). Hand Movement Direction Decoded from
MEG and EEG. Journal of Neuroscience, 28(4):1000–1008. (Cited on pages 23,
24 and 27.)
Bibliography 251
[Wallstrom et al., 2004] Wallstrom, G. L., Kass, R. E., Miller, A., Cohn, J. F.,
and Fox, N. A. (2004). Automatic correction of ocular artifacts in the EEG:
A comparison of regression-based and component-based methods. International
Journal of Psychophysiology, 53(2):105–119. (Cited on page 40.)
[Wang et al., 2009a] Wang, B., Wong, C. M., Wan, F., Mak, P. U., Mak, P. I., and
Vai, M. I. (2009a). Comparison of different classification methods for EEG-based
brain computer interfaces: A case study. 2009 IEEE International Conference on
Information and Automation, ICIA 2009, pages 1416–1421. (Cited on pages 53,
54, 55 and 56.)
[Wang et al., 2015] Wang, J. J., Xue, F., and Li, H. (2015). Simultaneous channel
and feature selection of fused EEG features based on Sparse Group Lasso. BioMed
Research International, 2015. (Cited on page 48.)
[Wang et al., 2016] Wang, P. T., King, C. E., McCrimmon, C. M., Lin, J. J., Sazgar,
M., Hsu, F. P. K., Shaw, S. J., Millet, D. E., Chui, L. a., Liu, C. Y., Do,
A. H., and Nenadic, Z. (2016). Comparison of decoding resolution of standard
and high-density electrocorticogram electrodes. Journal of Neural Engineering,
13(2):026016. (Cited on pages 18, 50, 51, 173 and 175.)
[Wang et al., 2013a] Wang, P. T., King, C. E., Schombs, A., Lin, J. J., Sazgar, M.,
and Hsu, F. P. K. (2013a). Electrocorticographic Gamma Band Power Encodes
the Velocity of Upper Extremity Movements. International BCI Meeting. (Cited
[Wang et al., 2013b] Wang, P. T., Puttock, E. J., King, C. E., Schombs, A., Lin,
J. J., Sazgar, M., Hsu, F. P. K., Shaw, S. J., Millett, D. E., Liu, C. Y., Chui,
L. a., Do, A. H., and Nenadic, Z. (2013b). State and trajectory decoding of
upper extremity movements from electrocorticogram. International IEEE/EMBS
Conference on Neural Engineering, NER, pages 969–972. (Cited on pages 43, 62,
63, 65, 66, 102, 103, 165, 201 and 207.)
[Wang et al., 2007] Wang, W., Chan, S. S., Heldman, D. a., and Moran, D. W.
(2007). Motor cortical representation of position and velocity during reaching.
Journal of neurophysiology, 97(6):4258–70. (Cited on page 61.)
[Wang et al., 2013c] Wang, W., Collinger, J. L., Degenhart, A. D., Tyler-Kabara,
E. C., Schwartz, A. B., Moran, D. W., Weber, D. J., Wodlinger, B., Vinjamuri,
R. K., Ashmore, R. C., Kelly, J. W., and Boninger, M. L. (2013c). An electrocor-
ticographic brain interface in an individual with tetraplegia. PloS one, 8(2):e55344.
(Cited on pages 3, 10, 11, 12, 13, 17, 18, 21, 23, 26, 27, 29, 30, 31, 44, 60, 61, 63,
162, 176 and 177.)
[Wang et al., 2009b] Wang, W., Degenhart, a. D., Collinger, J. L., Vinjamuri, R., Su-
dre, G. P., Adelson, P. D., Holder, D. L., Leuthardt, E. C., Moran, D. W., Boninger,
M. L., Schwartz, a. B., Crammond, D. J., Tyler-Kabara, E. C., and Weber, D. J.
252 Bibliography
(2009b). Human motor cortical activity recorded with Micro-ECoG electrodes, dur-
ing individual finger movements. Conference proceedings : Annual International
(Cited on pages 18, 31, 47, 53, 54 and 195.)
[Wang et al., 2013d] Wang, X., Mueen, A., Ding, H., Trajcevski, G., Scheuermann,
P., and Keogh, E. (2013d). Experimental comparison of representation methods
and distance measures for time series data. Data Mining and Knowledge Discovery,
26(2):275–309. (Cited on page 109.)
[Wang et al., 2006] Wang, Y., Hong, B., Gao, X., and Gao, S. (2006). Phase
synchrony measurement in motor cortex for classifying single-trial EEG during
motor imagery. Annual International Conference of the IEEE Engineering in
Medicine and Biology - Proceedings, pages 75–78. (Cited on page 45.)
[Wang et al., 2012] Wang, Z., Gunduz, A., Brunner, P., Ritaccio, A. L., Ji, Q., and
Schalk, G. (2012). Decoding onset and direction of movements using Electrocortico-
graphic (ECoG) signals in humans. Frontiers in neuroengineering, 5(August):15.
[Wang et al., 2011] Wang, Z., Ji, Q., Miller, K. J., and Schalk, G. (2011). Prior
knowledge improves decoding of finger flexion from electrocorticographic signals.
Frontiers in neuroscience, 5(November):127. (Cited on pages 63, 64, 66, 67, 102,
103, 160 and 171.)
[Waser and Garn, 2013] Waser, M. and Garn, H. (2013). Removing cardiac interfer-
ence from the electroencephalogram using a modified Pan-Tompkins algorithm and
linear regression. Conference proceedings : ... Annual International Conference
of the IEEE Engineering in Medicine and Biology Society. IEEE Engineering in
Medicine and Biology Society. Annual Conference, 2013:2028–2031. (Cited on
page 40.)
[Waterhouse, 1997] Waterhouse, S. R. (1997). Classification and Regression using

Mixtures of Experts. PhD thesis. (Cited on pages 72, 73, 75 and 123.)
[Webb et al., 2012] Webb, J., Xiao, Z. G., Aschenbrenner, K. P., Herrnstadt, G.,
and Menon, C. (2012). Towards a portable assistive arm exoskeleton for stroke
patient rehabilitation controlled through a brain computer interface. Proceedings
of the IEEE RAS and EMBS International Conference on Biomedical Robotics
and Biomechatronics, pages 1299–1304. (Cited on pages 19, 20, 29 and 40.)
[Wei et al., 2006] Wei, Q., Gao, X., and Gao, S. (2006). Feature extraction and
subset selection for classifying single-trial ECoG during motor imagery. Annual
International Conference of the IEEE Engineering in Medicine and Biology -
Proceedings, pages 1589–1592. (Cited on pages 45 and 48.)
Bibliography 253
[Wei and Tu, 2008] Wei, Q. and Tu, W. (2008). Channel selection by genetic
algorithms for classifying single-trial ECoG during motor imagery. Conference
proceedings : ... Annual International Conference of the IEEE Engineering in
Medicine and Biology Society. IEEE Engineering in Medicine and Biology Society.
Conference, 2008:624–7. (Cited on page 41.)
[Wei et al., 2007] Wei, Q., Wang, Y., Gao, X., and Gao, S. (2007). Amplitude and
phase coupling measures for feature extraction in an EEG-based brain-computer
interface. Journal of neural engineering, 4(2):120–129. (Cited on page 45.)
[Weigend et al., 1995] Weigend, a. S., Mangeas, M., and Srivastava, a. N. (1995).
Nonlinear gated experts for time series: discovering regimes and avoiding overfit-
ting. International journal of neural systems, 6(4):373–99. (Cited on page 81.)
[Welch and Bishop, 2006] Welch, G. and Bishop, G. (2006). An Introduction to the
Kalman Filter. https://www.cs.unc.edu/˜welch/media/pdf/kalman_intro.pdf,
7(1):1–16. (Cited on page 202.)
[Whitham et al., 2007] Whitham, E. M., Pope, K. J., Fitzgibbon, S. P., Lewis, T.,
Clark, C. R., Loveless, S., Broberg, M., Wallace, A., DeLosAngeles, D., Lillie,
P., Hardy, A., Fronsko, R., Pulbrook, A., and Willoughby, J. O. (2007). Scalp
electrical recording during paralysis: Quantitative evidence that EEG frequencies
above 20 Hz are contaminated by EMG. Clinical Neurophysiology, 118(8):1877–
[Willett et al., 2013] Willett, F. R., Suminski, A. J., Fagg, A. H., and Hatsopoulos,
N. G. (2013). Improving brain-machine interface performance by decoding intended
future movements. Journal of neural engineering, 10(2):026011. (Cited on pages 8,
59, 109 and 189.)
[Williams et al., 2013] Williams, J. J., Rouse, A. G., Thongpang, S., Williams, J. C.,
and Moran, D. W. (2013). Differentiating closed-loop cortical intention from
rest: building an asynchronous electrocorticographic BCI. Journal of neural
engineering, 10(4):046001. (Cited on pages 17, 18, 27, 29, 53, 59, 63, 65, 70, 71,
73, 74, 103, 159, 161, 162, 165, 171 and 189.)
[Wilson et al., 2006] Wilson, J. A., Felton, E. a., Garell, P. C., Schalk, G., and
Williams, J. C. (2006). ECoG factors underlying multimodal control of a brain-
computer interface. IEEE transactions on neural systems and rehabilitation
engineering : a publication of the IEEE Engineering in Medicine and Biology
Society, 14(2):246–50. (Cited on pages 17 and 18.)
[Wilson et al., 2012] Wilson, J. A., Walton, L. M., Tyler, M., and Williams, J.
(2012). Lingual electrotactile stimulation as an alternative sensory feedback
pathway for brain-computer interface applications. Journal of neural engineering,
9(4):045007. (Cited on page 5.)
254 Bibliography
[Wissel et al., 2013] Wissel, T., Pfeiffer, T., Frysch, R., Knight, R. T., Chang, E. F.,
Hinrichs, H., Rieger, J. W., and Rose, G. (2013). Hidden Markov model and
support vector machine based decoding of finger movements using electrocorticog-
raphy. Journal of Neural Engineering, 10(5):056020. (Cited on pages 31, 48, 56
and 57.)
[Wodlinger et al., 2015] Wodlinger, B., Downey, J. E., Tyler-Kabara, E. C.,

Schwartz, a. B., Boninger, M. L., and Collinger, J. L. (2015). Ten-dimensional
anthropomorphic arm control in a human brain-machine interface: difficulties,
solutions, and limitations. Journal of neural engineering, 12(1):016011. (Cited on
pages 5, 10, 11, 12, 15, 16, 23, 24, 25, 29, 30, 31, 35, 42, 49, 59, 60, 61, 64, 167,
173 and 177.)
[Wold et al., 2001] Wold, S., Sjöström, M., and Eriksson, L. (2001). PLS-regression:
A basic tool of chemometrics. In Chemometrics and Intelligent Laboratory
Systems, volume 58, pages 109–130. (Cited on page 189.)
[Wolpaw et al., 2002] Wolpaw, J. R., Birbaumer, N., Mcfarland, D. J., Pfurtscheller,
G., and Vaughan, T. M. (2002). Brain – computer interfaces for communication
and control. Clinical neurophysiology, 113:767–791. (Cited on pages 3, 4, 6 and 8.)
[Wolpaw and McFarland, 2004] Wolpaw, J. R. and McFarland, D. J. (2004). Control

of a two-dimensional movement signal by a noninvasive brain-computer interface
in humans. Proceedings of the National Academy of Sciences of the United States
of America, 101(51):17849–54. (Cited on pages 3, 10, 11, 12, 13, 19, 26 and 30.)
[Wong et al., 2013] Wong, Y. T., Putrino, D., Weiss, A., and Pesaran, B. (2013).
Utilizing movement synergies to improve decoding performance for a brain ma-
chine interface. Proceedings of the Annual International Conference of the IEEE
Engineering in Medicine and Biology Society, EMBS, pages 289–292. (Cited on
page 46.)
[Wood et al., 2005] Wood, F., Prabhat, Donoghue, J., and Black, M. (2005). Infer-
ring attentional state and kinematics from motor cortical firing rates. Conference
proceedings : ... Annual International Conference of the IEEE Engineering in
Medicine and Biology Society. IEEE Engineering in Medicine and Biology Society.
Conference, 1:149–152. (Cited on pages 65 and 70.)
[Wu et al., 2003a] Wu, W., Black, M. J., Gao, Y., Bienenstock, E., Serruya, M.,
Shaikhouni, A., and Donoghue, J. P. (2003a). Neural Decoding of Cursor Motion
Using a Kalman Filter. Advances in Neural Information Processing Systems 15:
Proceedings of the 2002 Conference, (1):133–140. (Cited on pages 61, 62, 70
and 104.)
[Wu et al., 2002] Wu, W., Black, M. J., Gao, Y., Bienenstock, E., Serruya, M. D.,
and Donoghue, J. P. (2002). Inferring hand motion from multi-cell recordings
in motor cortex using a Kalman Filter. SAB’02- Workshop on Motor Control
Bibliography 255
in Humans and Robots: On the Interplay of Real Brains and Artificial Devices,
pages 66–73. (Cited on pages 61, 62, 70, 201, 206 and 207.)
[Wu et al., 2003b] Wu, W., Black, M. J., Mumford, D., Gao, Y., Bienenstock,
E., and Donoghue, J. P. (2003b). A Switching Kalman Filter Model for the
Motor Cortical Coding of Hand Motion. In Engineering in Medicine and Biology
Society, 2003. Proceedings of the 25th Annual International Conference of the
IEEE, pages 2083–2086. (Cited on pages 47, 71, 72, 172 and 207.)
[Wu et al., 2004] Wu, W., Black, M. J., Mumford, D., Gao, Y., Bienenstock, E.,
and Donoghue, J. P. (2004). Modeling and decoding motor cortical activity
using a switching Kalman filter. IEEE transactions on bio-medical engineering,
51(6):933–42. (Cited on pages 67, 70, 71, 74, 104, 159 and 172.)
[Wu et al., 2008] Wu, W., Gao, X., Hong, B., and Gao, S. (2008). Classifying single-
trial EEG during motor imagery by iterative spatio-spectral patterns learning
(ISSPL). IEEE Transactions on Biomedical Engineering, 55(6):1733–1743. (Cited
on page 45.)
[Xu et al., 1994] Xu, L., Jordan, M. I., and Hinton, G. E. (1994). An alternative
model for mixtures of experts. Nips, (7):633–640. (Cited on page 123.)
[Yanagisawa et al., 2012a] Yanagisawa, T., Hirata, M., Saitoh, Y., Kishima, H.,
Matsushita, K., Goto, T., Fukuma, R., Yokoi, H., Kamitani, Y., and Yoshimine,
T. (2012a). Electrocorticographic control of a prosthetic arm in paralyzed patients.
Annals of Neurology, 71(3):353–361. (Cited on pages 8, 13, 17, 18, 26, 30, 44, 51
and 53.)
[Yanagisawa et al., 2012b] Yanagisawa, T., Hirata, M., Saitoh, Y., Kishima, H.,
Matsushita, K., Goto, T., Fukuma, R., Yokoi, H., Kamitani, Y., and Yoshimine,
T. (2012b). Electrocorticographic control of a prosthetic arm in paralyzed patients.
Annals of Neurology, 71. (Cited on pages 21, 42, 43, 49, 50, 51 and 52.)
[Yin et al., 2015] Yin, E., Zhou, Z., Jiang, J., Yu, Y., and Hu, D. (2015). A
dynamically optimized SSVEP brain-computer interface (BCI) speller. IEEE
Transactions on Biomedical Engineering, 62(6):1447–1456. (Cited on page 2.)
[Yu et al., 2007] Yu, B. M., Kemere, C., Santhanam, G., Afshar, A., Ryu, S. I.,
Meng, T. H., Sahani, M., and Shenoy, K. V. (2007). Mixture of trajectory models
for neural decoding of goal-directed movements. Journal of neurophysiology,
97(5):3763–3780. (Cited on pages 67, 70, 71, 159, 164 and 172.)
[Yu, 2010] Yu, S.-Z. (2010). Hidden semi-Markov models. Artificial Intelligence,
174(2):215–243. (Cited on page 174.)
[Yuan et al., 2007] Yuan, H., Doud, A. J., Gururajan, A., and He, B. (2007). Local-
ization of Event-related ( de ) Synchronization of Cerebral Cortex during Online
Control of Brain-Computer Interface Using Minimum-norm Estimates in the
256 Bibliography
Frequency Domain. International Journal of Bioelectromagnetism, 9(2):109–110.

[Yuan and He, 2014] Yuan, H. and He, B. (2014). Brain-Computer Interfaces Us-
ing Sensorimotor Rhythms: Current State and Future Perspectives. IEEE
Transactions on Biomedical Engineering, 61(5):1425–1435. (Cited on pages 4,
5, 8, 12, 17, 23 and 27.)
[Yuksel et al., 2012] Yuksel, S. E., Wilson, J. N., and Gader, P. D. (2012). Twenty
years of mixture of experts. IEEE Transactions on Neural Networks and Learning
Systems, 23(8):1177–1193. (Cited on pages 73, 75, 79 and 82.)
[Zhang et al., 2007] Zhang, C., Fu, H., Jiang, Y., and Yu, T. (2007). High-
dimensional pseudo-logistic regression and classification with applications to gene
expression data. Computational Statistics and Data Analysis, 52(1):452–470.
258 Bibliography
ECoG signal processing for Brain Computer Interface

with multiple degrees of freedom for clinical application
Brain-Computer Interfaces (BCI) are systems that allow severely motor-impaired
patients to use their brain activity to control external devices, for example upper-
limb prostheses in the case of motor BCIs. The user’s intentions are estimated
by applying a decoder on neural features extracted from the user’s brain activity.
Signal processing challenges specific to the clinical deployment of motor BCI systems
are addressed in the present doctoral thesis, namely asynchronous mono-limb or
sequential multi-limb decoding and accurate decoding during active control states.
A switching decoder, namely a Markov Switching Linear Model (MSLM), has been
developed to limit spurious system activations, to prevent parallel limb movements
and to accurately decode complex movements. The MSLM associates linear models
with different possible control states, e.g. activation of a specific limb, specific
movement phases. Dynamic state detection is performed by the MSLM, and the
probability of each state is used to weight the linear models. The performance of
the MSLM decoder was assessed for asynchronous wrist and multi-finger trajectory
reconstruction from electrocorticographic signals. It was found to outperform
previously reported decoders for the limitation of spurious activations during no-
control periods and permitted to improve decoding accuracy during active periods.
Keywords: Asynchronous brain-computer-interface, Hidden Markov Models,
Mixtures of Experts
Traitement du signal ECoG pour Interface Cerveau Machine

à grand nombre de degrés de liberté pour application clinique
Les Interfaces Cerveau-Machine (ICM) sont des systèmes qui permettent à des
patients souffrant d’un handicap moteur sévère d’utiliser leur activité cérébrale
pour contrôler des effecteurs, par exemple des prothèses des membres supérieurs
dans le cas d’ICM motrices. Les intentions de mouvement de l’utilisateur sont
estimées en appliquant un décodeur sur des caractéristiques extraites de son activité
cérébrale. Des challenges spécifiques au déploiement clinique d’ICMs motrices
ont été considérés, à savoir le contrôle mono-membre ou séquentiel multi-membre
asynchrone et précis. Un décodeur, le Markov Switching Linear Model (MSLM), a été
développé pour limiter les activations erronées de l’ICM, empêcher des mouvements
parallèles des effecteurs et décoder avec précision des mouvements complexes. Le
MSLM associe des modèles linéaires à différents états possibles, e.g. le contrôle d’un
membre spécifique ou une phase de mouvement particulière. Le MSLM réalise une
détection d’état dynamique, et les probabilités des états sont utilisées pour pondérer
les modèles linéaires. La performance du décodeur MSLM a été évaluée pour la
reconstruction asynchrone de trajectoires de poignet et de doigts à partir de signaux
electrocorticographiques. Il a permis de limiter les activations erronées du système
et d’améliorer la précision du décodage du signal cérébral.
Mots-clefs: Interface Cerveau-Machine Asynchrone, Modèles de Markov Cachés,
Mélanges d’Experts

ECoG of BCI Thesis

Transféré par

Droits d'auteur :

Formats disponibles

ECoG of BCI Thesis

Transféré par

Informations du document

Copyright

Formats disponibles

Partager ce document

Partager ou intégrer le document

Options de partage

Avez-vous trouvé ce document utile ?

Ce contenu est-il inapproprié ?

Droits d'auteur :

Formats disponibles

ECoG of BCI Thesis

Transféré par

Droits d'auteur :

Formats disponibles

ECoG signal processing for Brain Computer Interface

with multiple degrees of freedom for clinical application

To cite this version:

HAL Id: tel-01763451

HAL is a multi-disciplinary open access L’archive ouverte pluridisciplinaire HAL, est

préparée au sein du Laboratoire CLINATEC

Traitement du signal ECoG pour

International conference proceedings

M.C. Schaeﬀer, E. Labyt, V. Rohu, N. Tarrin, I. Vergara, S. Cokgungor, A. Eliseyev,

M.C. Schaeﬀer, S. Cokgungor, N. Tarrin, V. Rohu, E. Labyt, C. Mestais, A.L. Ben-

Décodeurs pour ICM motrices précises et asynchrones Diﬀérents outils al-

signaux cérébraux. Deux stratégies de décodage sont principalement exploitées par

Méthodes Des modiﬁcations de la dépendence entre signaux cérébraux et paramètres

dans la présente thèse. La procédure d’apprentissage supervisé est pertinente quand

Discussion La pertinence du modèle proposé pour réaliser un décodage asyn-

l’évaluation de la performance du MSLM lors de sessions de contrôle en boucle

Abbreviations and notations

M1 Primary Motor Cortex

UKF Unscented Kalman Filter

2 Data-driven transducers for motor BCIs 37

2.6.3 Accurate decoding during IC states . . . . . . . . . . . . . . . 66

3 Switching models for accurate asynchronous trajectory decoding 69

4 Data set presentation 89

5 Evaluation and comparison 101

8 Discussion and perspectives 167

A Supplementary materials for MSLM training 179

B MSLM: expert training procedure and gate selection 187

C Switching Kalman Filtering for asynchronous ECoG decoding 201

1.1 Brain-Computer Interfaces

1.1.2 Brain pattern elicitation

A distinction is drawn between externally- and internally-paced brain patterns

Brain pattern elicitation sometimes requires the execution of movements of

1.1.3 BCI components

ticographic (ECoG) or electroencephalographic (EEG) arrays [Lebedev and Nicolelis,

Transducer The BCI transducer translates brain activity measurements into

by applying model identiﬁcation algorithms on a data set that contains samples of

Feedback Natural volitional motor control is permitted by the perception and

1.2 Motor BCIs

1.2.1 Towards clinical motor BCIs

Chronicity Long-term validity is commonly aimed at during the development

Multi-limb control Full independence of severely motor-impaired patients, e.g.

Degrees of Freedom Upper- and lower-limb prostheses have multiple controllable

Robustness Most BCI studies are currently completed in controlled laboratory

Figure 1.2: Ability to execute reaching movements depending on the correlation R

Asynchronicity Most BCI demonstrations are conducted using synchronous, cue-

Embeddability Bulky hardware components for signal acquisition, processing

1.2.2 Signal acquisition

1.2.2.1 Cortical patterns for prosthetic control

Electrophysiologial signal acquisition systems mostly give access to the cerebral

Cortical electro-physiological activity Cortical electrophysiological activities

• Delta: 0.5Hz < fδ < 4Hz

• Theta: 4Hz < fθ < 7Hz

• Mu: 8Hz < fµ < 13Hz (over the motor cortex)

• Beta: 13Hz < fβ < 30Hz

• Gamma: 30Hz < fγ

Figure 1.3: Cerebral cortex. A. Cerebral cortex (represented in dark purple)

Functional organization Most motor BCIs exploit neural activities generated

Electro-physiological sources of control for motor BCIs Most reported

Cortical neurons The voluntary modulation of the ﬁring rate of neurons in

Event-related potentials A few teams have reported P300-based neural