Jurnal 1 C

Maceira-Elvira et al.
Journal of NeuroEngineering and Rehabilitation (2019) 16:142

https://doi.org/10.1186/s12984-019-0612-y
REVIEW Open Access
Wearable technology in stroke

rehabilitation: towards improved diagnosis
and treatment of upper-limb motor
impairment
Pablo Maceira-Elvira1,2, Traian Popa1,2, Anne-Christine Schmid1,2 and Friedhelm C. Hummel1,2,3*
Abstract
Stroke is one of the main causes of long-term disability worldwide, placing a large burden on individuals and
society. Rehabilitation after stroke consists of an iterative process involving assessments and specialized training,
aspects often constrained by limited resources of healthcare centers. Wearable technology has the potential to
objectively assess and monitor patients inside and outside clinical environments, enabling a more detailed
evaluation of the impairment and allowing the individualization of rehabilitation therapies. The present review aims
to provide an overview of wearable sensors used in stroke rehabilitation research, with a particular focus on the
upper extremity. We summarize results obtained by current research using a variety of wearable sensors and use
them to critically discuss challenges and opportunities in the ongoing effort towards reliable and accessible tools
for stroke rehabilitation. Finally, suggestions concerning data acquisition and processing to guide future studies
performed by clinicians and engineers alike are provided.
Keywords: Stroke, Wearable technology, Rehabilitation, Monitor, Motor function, Home-based, Remote, Telemedicine
Introduction differently, the rehabilitative process after stroke de-

Stroke is one of the leading causes of disability world- pends on the effective assessment of motor deficit and
wide [1], with a global prevalence estimated at 42.4 mil- congruent allocation to treatment (diagnostics), accurate
lion in 2015 [2]. Stroke results in permanent motor appraisal of treatment effects (recovery/adaptation evalu-
disabilities in 80% of cases [3]. During the acute and ation), and prolonged treatment for continuous recovery
subacute stages (< 6 months after stroke [4]), patients re- during the chronic stage (extended training).
ceive rehabilitation therapies at specialized healthcare Each of these three aspects present practical chal-
centers, consisting of an iterative process involving im- lenges. Assigned treatments depend on the assessed
pairment assessments, goal definition, intervention, and early-stage disability [3]. A variety of assessment scales
progress evaluation [5]. After being discharged from the exist to evaluate motor impairment after stroke, de-
rehabilitation center (i.e. after entering the chronic stage, signed to capture aspects such as joint range of motion
e.g., 6 months after stroke), 65% of patients are unable to (ROM), synergistic execution of movements, reaching
integrate affected limbs into everyday-life activities [6], and grasping capabilities, object manipulation, etc. [7].
showing a need for further treatment. Phrased These assessments are normally applied by specialized
medical personnel, which entails certain variability be-
* Correspondence: [email protected] tween assessments [8]. Besides consistency in repeated
1
Defitech Chair in Clinical Neuroengineering, Center for Neuroprosthetics measurements, some scales like the Fugl-Meyer assess-
(CNP) and Brain Mind Institute (BMI), Swiss Federal Institute of Technology
(EPFL), 9, Chemin des Mines, 1202 Geneva, Switzerland
ment (FMA) [9], are unable to capture the entire
2
Defitech Chair in Clinical Neuroengineering, Center for Neuroprosthetics spectrum of motor function in patients due to limited
(CNP) and Brain Mind Institute (BMI), Swiss Federal Institute of Technology sensitivity or ceiling effects [10].
(EPFL Valais), Clinique Romande de Réadaptation, 1951 Sion, Switzerland
Full list of author information is available at the end of the article
© The Author(s). 2019 Open Access This article is distributed under the terms of the Creative Commons Attribution 4.0
International License (http://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and
reproduction in any medium, provided you give appropriate credit to the original author(s) and the source, provide a link to
the Creative Commons license, and indicate if changes were made. The Creative Commons Public Domain Dedication waiver
(http://creativecommons.org/publicdomain/zero/1.0/) applies to the data made available in this article, unless otherwise stated.
Maceira-Elvira et al. Journal of NeuroEngineering and Rehabilitation (2019) 16:142 Page 2 of 18
In addition to thorough standardized assessment Wearable devices used in stroke patients

scales, progress in patients is observable during the exe- Recent availability of ever more compact, robust and
cution of activities of daily living (e.g., during occupa- power-efficient wearable devices has presented research
tional therapy sessions). Nevertheless, task completion and development groups in academia and industry with
not always reflects recovery, as patients often adopt dif- the means of studying and monitoring activities per-
ferent synergistic patterns to compensate for lost func- formed by users on a daily basis.
tion [11], and such behavior is not always evident. Over the past years, multiple research groups have
Main provision of rehabilitation therapies occurs at worked towards a reliable, objective and unobtrusive
hospitals and rehabilitation centers. Evidence of en- way of studying human movement. From the array of
hanced recovery related to more extensive training has sensors and devices created, a few have gained popular-
been found [12], but limited resources at these facilities ity in time due to their practicality. The next subsections
often obstruct extended care during the chronic stage. will focus on the wearable devices most frequently used
This calls for new therapeutic options allowing patients in the study of human motion, with special emphasis on
to train intensively and extensively after leaving the monitoring of upper limbs in stroke patients.
treatment center, while ensuring the treatment’s quality,
effectiveness and safety.
Inertial measurement units (IMUs)
Wearable sensors used during regular assessments can
Inertial measurement units (IMUs) are devices combin-
reduce evaluation times and provide objective, quantifi-
ing the acceleration readings from accelerometers and
able data on the patients’ capabilities, complementing
the angular turning rate detection of gyroscopes [13].
the expert yet subjective judgement of healthcare spe-
Recent versions of such devices are equipped with a
cialists. These recordings are more objective and replic-
magnetometer as well, adding an estimation of the
able than regular observations. They have the potential
orientation of the device with respect to the Earth’s mag-
of reducing diagnostic errors affecting the choice for
netic field [14]. A general description of how inertial
therapies and their eventual readjustment. Additional in-
data are used to extract useful information from these
formation (e.g., muscle activity) extracted during the
devices is offered by Yang and Hsu [15]. High-end IMUs
execution of multiple tasks can be used to better
used for human motion tracking, such as the “MTw
characterize motor function in patients, allowing for
Awinda” sensor (Xsens®, Enscheda, Overijssel, The
finer stratification into more specific groups, which can
Netherlands) [16], acquire data at sampling rates as high
then lead to better targeted care (i.e. personalized ther-
as 1 kHz (sensitivities of ±2000 deg/s, ±160 m/s2, ±1.9
apies). These devices also make it possible to acquire
G). More affordable sensors (e.g. “MMR” (mbientlab
data unobtrusively and continuously, which enables the
Inc.®, San Francisco, California, USA) [17]) stream data
study of motor function while patients perform daily-life
at 100 Hz (max sensitivities of ±2000 deg/s, ±16 g, 13 G).
activities. Further, the prospect of remotely acquiring
The necessary sampling rate depends on the application,
data shows promise in the implementation of independ-
and must be defined such that aliasing is avoided (i.e.
ent rehabilitative training outside clinics, allowing pa-
Nyquist rate, 2 times the frequency of the studied
tients to work more extensively towards recovery.
phenomenon). Figure 1 shows an example of motion
The objective of this review is to provide an overview
tracking using these devices.
of wearable sensors used in stroke rehabilitation re-
search, with a particular focus on the upper extremity,
aiming to present a roadmap for translating these tech- Diagnostics
nologies from “bench to bedside”. We selected articles Multiple scales exist for assessing motor function in
based on their reports about tests conducted with actual stroke patients [7]. However, limitations exist in terms
stroke patients, with the exception of conductive elasto- of objectivity and test responsiveness to subtle changes
mer sensors, on which extensive research exists without [18], as well as on the amount of time needed to apply
tests in patients. In the section “Wearable devices used these tests. Therefore, several research groups have fo-
in stroke patients”, we summarize results obtained by cused on the use of IMUs to assess motor function more
current research using a variety of wearable sensors and objectively. Hester et al. [19] were able to predict hand
use them to critically discuss challenges and opportun- and arm stages of the Chedoke-McMaster clinical score,
ities in the ongoing effort towards reliable and accessible while Yu et al. [20] built Brunnstrom stage [21] classi-
tools for stroke rehabilitation. In the “Discussion” sec- fiers, assigning each patient to one of six classes of syn-
tion, we present suggestions concerning data acquisition ergistic movements in affected limbs. The Wolf Motor
and processing, as well as opportunities arising in this test [22–24], the FMA [25, 26] and the Action Research
field, to guide future studies performed by clinicians and Arm Test (ARAT) [27], frequently used to assess motor
engineers alike. function in clinical settings, have also been automated.
interesting possibilities have been explored in every-day-

use devices, such as smartphones [33].
Tracking of the whole body has also been achieved
using sensor networks in an attempt to objectively evalu-
ate movement quality in daily-life situations [34], as well
as tracking of complex upper-limb movements [35].
Extended training
IMUs allow providing immediate feedback to patients
about their performance and posture [36, 37], as well as
the adequate use of equipment (e.g., orthoses) [38],
which presents an opportunity for extended training
(e.g., at home). Wittman and colleagues [39] used an off-
the shelf system to train patients at home, seeing signifi-
cant improvements as assessed by both the FMA and
metrics native to the used IMU system.
Implementation (requirements and challenges)

The complexity of tracking and assessing motion
depends on how constrained the circumstances for the
recordings are. Tracking motion during the execution of
daily-life activities is particularly difficult in stroke pa-
tients, as their movements are often slower, more seg-
mented and more variable than those of healthy
individuals [11]. Prolonged recordings are constrained
by multiple factors, such as battery life of the wearable
devices [40] and orientation drift resulting from the
double integration of angular acceleration [41]. Better-
performing batteries, better communication protocols
(e.g., Bluetooth Low-Energy (BLE) [42]) and algorithms
allowing to sample data at lower rates without losing
much information (e.g., data compression [20]) help
mitigate the former problem, while orientation drift can
be corrected using, for example, the on-board magnet-
ometer [41].
Fig. 1 IMU sensors (orange) used to track arm movements. Sensors Recording over shorter periods, like those during stan-
placed on the back of the hands, forearms and upper arms capture dardized motor function assessment scales, is less vul-
acceleration (linear and angular) and orientation of each segment,
nerable to these limiting factors, but still susceptible to
allowing kinematic reconstruction or movement characterization
other issues. Quantifying movements taking place in a
single plane (e.g., shoulder flexion, with the arm moving
Recovery/adaptation evaluation parallel to the sagittal plane) is straightforward, as re-
IMUs are practical options to assess motor function dur- cordings from either the accelerometer or the gyroscope
ing the execution of activities of daily life. Lee and col- can be sufficient. In contrast, characterizing complex
leagues [28] focused on limb neglect and task execution movements (e.g. flexor synergic movement from the
quality assessment. Limb neglect can be seen by looking FMA) is more challenging and often requires combining
at the symmetry (or lack thereof) in sensor readings data from both the accelerometer and the gyroscope.
from the affected and unaffected sides [29–31]. Zhou Assigning clinically relevant scores (e.g. FMA scores) to
et al. [32] used a single, triple-axis accelerometer to track performed movements requires characterizing the re-
movements of the forearm in a simple manner, but corded signals using a variety of features. These features
tracking of more complex motion requires either more are normally extracted using a sliding-window approach
sensors or alternative data analysis techniques. Harder- along the acquired signals, and the choice of which fea-
to-detect compensatory movements (e.g., of the torso) tures to use depends on the type of movements involved.
can also be identified [19]. Besides using IMU modules Common features used in characterization of IMU data
designed specifically for human movement tracking, are movement intensity, signal amplitude (mean and
standard deviation), signal energy and dominant fre- the circumstances of the recording and its correspond-
quency [43]. After extracting these features, statistical ing analysis. For instance, Ives and Wigglesworth [64]
methods commonly used in machine learning allow clas- showed significant decreases in amplitude (11.4%) and
sifying and assigning grades to the movements that orig- timing (39 ms signal lengthening) when comparing a
inated them; the initial choice of models to test depends sampling rate of 6 kHz to 250 Hz. These differences
on the extracted features [44]. would likely not affect the performance of a classifier if
Problems can arise when studying stroke patients, as the all data were recorded with the same sampling rate, but
acquired inertial signals may not hold enough information might impede classification if sampling rates were too
due to the very low variation of signals during slow move- different because of different amplitudes and timing
ments. An alternative to selecting features would be to shifts. High-end acquisition systems, such as “Ultium”
compare waveforms directly by defining a set of signals as wearable EMG sensors (Noraxon Inc.®, Scottsdale, Ari-
templates for unimpaired movements with signals ac- zona, USA) [65], have sampling rates as high as 4 kHz
quired from patients [45]. Techniques such as Coherent (sensitivity of 0.3 μV in a range of 0–5 V), while more
Point Drift (CPD) [46] or Dynamic Time Warping (DTW) accessible alternatives like the “FreeEMG” (BTS Bio-
[47] may be used. DTW has been used in stroke research engineering®, Garbagnate Milanese, Milan, Italy) [66]
by a number of groups (e.g. [48, 49]), as it allows to com- have a sampling rate of 1 kHz.
pare time series that are different in length, which is useful
when comparing slower movements in stroke patients to Diagnostics
conventional movements. CPD is a different technique for Wearable EMG sensors have high potential in the study of
registering one set of points to another, which estimates stroke patients. Investigation of neural activity as mea-
the maximum likelihood between pairs of corresponding sured through motor-evoked potentials (MEPs) triggered
points and finds the best fit between them. by Transcranial Magnetic Stimulation (TMS) [67] is sim-
Sensor noise can cause huge detriment to the outcome pler with wireless EMG. EMG sensors can complement
of movement classification or assessment. The main inertial data from IMUs during standardized motor func-
source of noise for short-duration recordings is tion assessments. For example, Li and colleagues [62] im-
quantization noise (i.e., noise resulting from precision proved the correlation in 0.5% between their condensed
loss during analog-digital conversion), while the afore- measure of motor function and the FM score assigned by
mentioned drift rate plagues longer recordings [50]. a clinician. Albeit the modest increase, assessment of dex-
Wearable sensor misplacement or misalignment can also terous movements, grasping exercises and applied force is
affect classifier performance to a large extent, but some not practical with IMUs, but can be characterized with se-
approaches have reportedly maintained precision and re- lected EMG features (e.g. area under the curve correlating
call at high levels (e.g. orientation transformation, Pr. with applied force), which argues in favor of including this
97% and Rc. 98% [51]) during the classification of certain sensor type during motor assessments. Repnik and col-
movements. leagues [27] complemented IMU data with EMG during
Table 1 provides an overview of studies using wearable the assessment of the ARAT test to capture dexterous
sensors to study stroke patients. This table focuses on movements involved in the manipulation of small objects,
studies that included stroke patients in their cohorts. finding significant differences in muscle activation of
healthy subjects according to the size of grasped objects,
Surface electromyography (sEMG) and similar (maximal) muscle activation in more impaired
Surface Electromyography (sEMG) is a technique in patients (ARAT score 2) when grasping the largest object.
which the electrical potential generated whenever mus-
cles contract is measured using electrode pairs placed on Recovery/adaptation evaluation
the skin over the muscles. The electrodes need to be After stroke, patients tend to adopt compensatory strat-
asymmetrically placed with respect to the neuromuscu- egies to accomplish motor tasks, especially in case of
lar plaques in order to capture the electrical potential moderate to severe impairment [11]. These compensa-
difference as the depolarization wave travels along the tory behavior might go unnoticed during a regular as-
muscle cells’ membranes. Figure 2 shows a typical place- sessment, but can be captured and quantified using
ment configuration for EMG devices, intended to record recordings from EMG sensors [68].
activity from contracting muscles involved in elbow and
wrist flexion. Effectively capturing all significant fre- Extended training
quency components of the EMG signal (according to the Wearable EMG sensors allow providing online feedback
Nyquist rate) requires a sampling rate of 1000 Hz, as its during home-based training in a similar way as with
highest frequency components are reportedly around IMUs. Instead of tracking gross arm movements, applied
400–500 Hz [64]. Still, frequencies needed depend on force calculated from recordings of muscle activity can
Table 1 Studies involving the use of wearable sensors in the study of stroke. Only studies including actual patients shown. Most of the studies listed focused on the assessment
of motor function through standardized clinical tests, which focus mainly on movement quality. This might explain the much more common use of IMU’s so far
Device Sensors Product / Units Author Assessment Type Patients / Methods and Results
Category Healthy controls
IMU 6 accelerometers (hand, “Vitaport 3” (Temec®, Heerlen, Hester et al., Upper limb motor 12 / 0 10% relative error in prediction of clinical scores (leave-
forearm, upper arm, sternum) NL) / m/s2 2006 [19] assessment score one-out cross-validation).
prediction
6 accelerometers (hand, “Vitaport 3” / m/s2 Patel et al., FAS- WMFT 24 / 0 5.76% relative error when predicting FAS.
forearm, upper arm, sternum) 2010 [23]
1 IMU worn at the wrist Non-commercial, / m/s2, deg/s Parnandi FAS- WMFT 1/0 “Prediction” error close to zero. *Model likely overfits the
et al., 2010 data, reasoning behind analysis might be incorrect.
[24]
6 accelerometers (hand, “Vitaport 3” / m/s2 Del Din Motor function, FM Test 24 / 0 4-points off when using a single item of the WMFT to
forearm, upper arm, sternum) et al., 2011 predict total UEFM score (max 66).
[25]
1 IMU attached to the MotionNode® (Seattle, USA) / Zhang et al., Upper limb movement 2/1 Similarity between feature vectors for 5 exercises of the
forearm m/s2, deg/s 2012 [48] trajectory comparison UEFM is evaluated using cosine distance. The authors
claim that higher similarity (close to 0.9) corresponds to
higher FM scores assigned by therapists. They compare
feature vectors from affected and unaffected limbs in
patients, but they never show how similar the vectors are
Maceira-Elvira et al. Journal of NeuroEngineering and Rehabilitation
in the healthy person.

10 IMUs (upper and lower Non-commercial / m/s2, deg/s Strohrmann Changes in motor 2/0 Longitudinal look at changes in motor function over the
extremities and trunk) et al., 2013 function over time course of 4 weeks. Used linear regression. Mean RMSE was
[52] of 0.15 and correlation between regression estimate and
ground truth (expert assessment) was of 0.86.
2 accelerometers, one per Bailey et al. Upper-limb bilateral 48 / 74 Had participants wear accelerometers on each wrist for
(2019) 16:142
“GTX+” (ActiGraph®, Pensacola,

wrist USA) / m/s2 2015 [53] activity to detect limb 26 h. Calculated the magnitude of the acceleration vector
neglect during activities every second for each wrist, and a ratio of said vectors
of daily life between the affected and non-affected hands. Were able
to detect limb neglect in impaired patients (impairment
level measured using ARAT test).
IMU from smartphone worn “Blackberry Z10” (BlackBerry®, Capela et al., Human activity 12 / 15 Found common features for healthy individuals (young
on the right-front hip Waterloo, CAN)/ m/s2, deg/s 2015 [33] recognition (6 activities) and elderly) and stroke patients to discriminate between
different conditions of movement and stillness using a
smartphone. Classification accuracy was over 80% for
most of the levels of comparison (e.g. mobile vs.
immobile, large movements vs. stairs, etc.) when using
decision trees, and similar (if slightly lower) when using
SVM or Naive Bayes.
3 IMUs (Lower arm, upper “ArmeoSenso” (Hocoma®, Wittman Home-based 11 / 0 Significant improvement of motor function as assessed by
arm and trunk) Volketswil, CH), “MotionPod 3” et al. 2016 rehabilitative training the FMA (4.1 points) and by metrics native to the
(Movea Inc.®, Pleasanton, USA) [39] “ArmeoSenso” system
/ m/s2, deg/s, Gauss
2 accelerometers (forearm Not specified / m/s2 Yu et al., Brunnstrom stage 23 / 4 Used ELM to classify people into 5 of the 6 stages of the
and upper arm) 2016a [20] classifier Brunnstrom Stage Evaluation. 80% of samples were used
as training set. No cross-validation was done. All patients
Page 5 of 18
of motor function through standardized clinical tests, which focus mainly on movement quality. This might explain the much more common use of IMU’s so far (Continued)
belonged to stages from II to V. Stage VI is considered to
be unimpaired, so data acquired from healthy participants
were used. Data were acquired during a single exercise
(repeated several times) and used to predict Brunnstrom
stage. Accuracy was above 85% when using ELM.
1 IMU worn at the forearm “MTi-300” (Xsens®, Ensched, NL) Zhang et al., Upper limb motion 14 / 0 Recorded inertial data from 14 patients (6 were relatively
/ m/s2, deg/s 2016a [35] classification unimpaired). Used PCA and used top 7 components to
label recordings according to the motion that generated
them. The Fuzzy Kernel algorithm achieved an error rate
of 0% when dealing with the 6 well-recovered patients,
and of 0.56% for more impaired patients.
1 IMU worn at the forearm “MPU-6050” (InvenSense®, San Zhang et al., Upper limb motion 21 / 8 Proposed a mobility index based on DTW to characterize
Jose, USA) / m/s2 2016b [49] assessment patients’ movements and assign them to Brunnstrom
stages from III-VI. Their index used with a KNN classifica-
tion algorithm (k = 3) achieved an accuracy of 82% in
leave-one-out cross validation.
2 accelerometers, one per “LSM9DS0” (Adafruit®, New de Lucena Bimanual symmetry, jerk 9/0 Used PCA and concluded that the first component relates
wrist York, USA)/ m/s2 et al., 2017 and clinical function to to functional status, whereas they suggest the second
[30] explain variance in upper component might be related to movement quality (as it
limb recovery described a strong correlation (≥0.75) between
acceleration asymmetry and jerk asymmetry). Both
principal components were found to explain 86% of the
variance.
1 IMU worn at the wrist “ReSense” [54] / m/s2, deg/s Leuenberger Affected limb neglect 10 / 0 Proposed a new measure for arm use called Gross Arm
(2019) 16:142
et al., 2017 during activities of daily Movements, which detects changes in arm orientation
[29] life larger than 30 degrees. This measure has large correlation
to clinical tests (r > = 0.9) even when not removing
signals acquired while patients walk.
2 IMUs, one per wrist “Shimmer3” (Shimmer Lee et al., Neglect and exercise 20 / 10 Detection accuracy of goal-directed movements was de-
Research®, Ireland) / m/s2, deg/ 2018 [28] quality at home scribed with a ROC curve, with an AUC of 87%. F-score
s (harmonic mean of precision and recall) of 84.3% when
classifying movements into feedback vs no-feedback
groups, an F-score of 73.7% when detecting feedback
due to accuracy issues and an F-score of 65.3% when de-
tecting feedback due to compensatory movement.
EMG 10 EMG electrodes on Noraxon® (Scottsdale, USA) / Lee et al., Classification of hand 20 / 0 LDA to classify signals into 6 hand gestures, with
forearm and hand mV 2011 [55] postures (6 classes) accuracies ranging from 37.9% (severely impaired
subjects, Chedoke stage 2 and 3) to 71.3 (moderately
impaired, Chedoke stage 4 and 5). Single model built for
each patient, gradually adding more data to it.
89 EMG electrodes “Refa 128” (TMSI®, Twente, NL) Zhang & Classification of hand 12 / 0 Used Fisher linear discriminant analysis (PCA + LDA) for
/ mV Zhou, 2012 postures (20 classes) dimensionality reduction. Best performance (96%
[56] classification accuracy) was obtained using time-domain
features and an SVM classifier. Achieved comparable
Page 6 of 18
of motor function through standardized clinical tests, which focus mainly on movement quality. This might explain the much more common use of IMU’s so far (Continued)
results with only 8 electrodes, but do not specify which
ones
2 EMG electrodes Not specified / mV Donoso Home-based gamified, 10 / 0 Proved feasibility of this approach at home, and the
Brown et al., rehabilitative training system was described as engaging and motivating, but
2015 [57] there were no reports of improved functionality
transferred to activities of daily life
Pots. & 2 potentiometers “SP12S-1 K” (ETI Systems®, Durfee et al., Hand joint angles tracked 24 / 0 System of beams tracking wrist and index finger joint
Encoders Carlsbad, USA) / V mapped to 2009 [58] during proof of concept angles. No classification or other form of accuracy
angular displacement rehabilitative game reported, as position was mapped directly from
potentiometer readout.
1 encoder “E4” (US Digital®, Vancouver, Chen et al., Hand joint angles in 4- 10 / 0 Patients trained at home (only 7 finished training) during
USA) / V mapped to angular 2017 [59] bar hand orthosis 4 weeks, training 5 times a week. By the end of training,
displacement patients showed motor improvement of 4.9 +/− 4.1
points in FM score, with a strong correlation (0.90)
between amount of movements performed during
training and score improvement.
Flexible Flex sensors along the dorsal “Flexpoint bend sensor” Prange- Hand gesture tracking 5/0 Presented a glove with two possible modalities (assistive
sensors side of fingers (Flexpoint Sensor Systems®, Lasonder during rehabilitative and rehabilitative). Modest improvement in pinch force
Draper, USA) / V mapped to et al., 2017 training and assistive and execution of other tasks was reported, hinting
joint flexion [60] grasping towards potential benefits of its use as a rehabilitative/
assistive tool.
Combinations 2 accelerometers (forearm “ADXL345” (Analog Devices Yu et al., Motor function, FM Test 24 / 0 Evaluated shortened version of UEFM using
and upper arm) and 7 flex Inc.®, Norwood, USA) / m/s2, V 2016b [61] accelerometers and flex sensors. Model built using SVM
(2019) 16:142
sensors (dorsal side of mapped to joint flexion (after using RRelief algorithm for feature selection) had a
fingers, wrist) 0.92 correlation with clinical scores given by a therapist.
2 IMUs (wrist, upper arm) “MPU-9250” (InvenSense®, San Li et al., 2017 Motor function, FM Test 18 / 16 34-leave-one-out cross validation, achieved determination
and 10 sEMG (forearm, Jose, USA), EMG not stated / [62] coefficients (correlation between their proposed measure
biceps and triceps) m/s2, deg/s, mV and FM score) of 0.85 using different unsupervised (PCA,
NMF) and supervised (LASSO) algorithms.
8 EMG + IMU, worn at the “Myo” armband (Thalmic Labs®, Ryser et al., Control signals for a hand 3 / 0 Classification accuracy between 78 and 98% in stroke
forearm Kitchener, Canada) / m/s2, deg/ 2017 [63] orthosis patients when discriminating between 3 hand gestures
s, arbitrary units (Myo EMG) using SVM.
9 IMUs (hands, wrists, upper “Myo” armband, IMU not stated Repnik et al., ARAT test 28 / 12 Correlation of 0.60 between movement time and
arms, forearms, sternum) and / m/s2, deg/s, arbitrary units 2018 [27] movement smoothness with respect to ARAT score. EMG
16 EMG electrodes (worn at (Myo EMG) data revealed significant differences in muscle activity of
the forearms) healthy subjects when grasping objects of different sizes.
Normalized muscle activation revealed that, in more
affected patients (ARAT score 2), maximal muscle
activation was present when grasping the largest object,
while in healthy participants activation was close to one
third of maximal output.
Page 7 of 18
during motor assessments, is often approached by extract-

ing meaningful features out of a sliding window. Some
groups tried correlating their own measures to scale scores
without a formal validation of their measure, which makes
interpretation difficult and supports an approach of direct
label/score prediction in the context of standardized tests.
As described for IMUs, a sliding-window approach al-
lows extracting significant features for later classification.
Classification is generally performed using signal fea-
tures (i.e. root mean-square, amplitude, etc.) [71] chosen
based on the type of movements in question. Alterna-
tively, extracting many features and applying feature se-
lection criteria afterwards [72] is also possible.
Classification accuracy tends to be high when only a
few (five or six) classes (each corresponding to a gesture
to be identified) are involved, but accuracy frequently
decreases as more gestures are added. Further detriment
to classification performance occurs when dealing with
highly impaired stroke patients, as their muscle signals
tend to be less pronounced [55]. Electrode number and
distribution plays a role as well; high density EMG, with
Fig. 2 EMG sensors (green) placed over biceps and flexor digitorum over 80 electrodes placed as a grid on the upper arm,
superficialis muscles, involved in elbow and wrist flexion, forearm and hand, has yielded high classification accur-
respectively. Electrodes placed asymmetrically with respect to the
neuromuscular plaques allow capturing the electrical potential
acies when dealing with many hand postures, but the
difference as the depolarization wave travels along the muscle cells’ use of only a few well-placed electrodes yields compar-
membranes. Resulting signal (top left) is filtered and amplified for able results [56]. Arrays of electrodes placed on the fore-
further processing arm offer a good tradeoff between relatively simple
setups and useful data acquisition leading to acceptable
serve as a parameter to provide feedback during training. classification accuracies. Pizzolato et al. [73] compared
EMG-based biofeedback has been reported to lead to en- an inexpensive device, consisting of eight single differen-
hanced motor improvements [69], and Donoso Brown tial electrodes worn as a bracelet, to more complex and
and colleagues [57] used it to test a gamified form of much more expensive systems. They reported a reason-
home-based training, although they did not find any im- ably high classification accuracy (69.04% +/− 7.77%) with
proved functionality derived from their intervention. a setup of two adjacent bracelets (16 electrodes).
There are several factors affecting the EMG signal. Re-
Implementation (requirements and challenges) peated recordings performed on the same test subjects
After amplification and preprocessing (e.g. signal filtering during several days has been reported to decrease hand-
for de-noising), these signals can be used to identify pat- gesture classification in close to 30%, compared to re-
terns of activation related to specific movements or pos- sults obtained from repeated measurements taking place
tures. The type of processing applied to recorded signals during the same day [74]. This might result from sensors
depends on the application. For example, continuous re- being placed in slightly different locations, as altering
cordings of muscle activity during the execution of activ- the position of an electrode by just one centimeter can
ities of daily living requires epoching the signals, keeping result in amplitude variations of 200% [75]. Hermens
only relevant segments capturing discrete events of inter- and colleagues offer a series of recommendations on
est. It is possible to do this segmentation manually, but sensor placement and orientation to decrease this vari-
automated methods of threshold detection are a much ability [76].
more practical option [70]. After removing signal seg- Other sources of EMG noise affecting the performance
ments deemed irrelevant, an adequate processing pipeline of used classifiers include cable motion artifacts, power-
must be implemented depending on the information line noise, thermal noise from the sensor’s electronic
sought. Extracting information about motor-unit activity components, electrochemical noise from the interface
while performing e.g. activities of daily living is possible between the electrodes and the skin and mechanical dis-
through wavelet analysis or a variety of time-frequency ap- turbances [70]. Currently-available wearable EMG sen-
proaches [70]. In contrast, identification of gross arm sors are mostly affected by mechanical disturbances,
movements and hand gestures, as well as their assessment which can be filtered out by applying a high pass filter
with cutoff frequency at 20 Hz [77]. The choice for ap- encoders to calculate the joint angles of an arm exoskel-
plied filtering also depends on the application. For ex- eton and using this parameter to adjust therapeutic
ample, low frequencies (i.e. 1–5 Hz) contain important training. Lim et al. [83] combined accelerometers with a
information for hand gesture classification [78], which different encoder using a slitted strip instead of a slitted
would be filtered out with the 20 Hz high-pass filter. disc. This sensor detects the linear displacement of the
strip, which means that laying the strips along the links
Potentiometers and encoders of interest (i.e. fingers) allows the measurement of joint
An accurate way of measuring the angular displacement angles without aligning the rotation axes, facilitating its
around joints is by means of potentiometers and en- use during the execution of daily life activities.
coders. Potentiometers are devices containing a conduct-
ive disc with a certain resistance and two contact points Extended training
on top. The distance between these contact points can Chen and colleagues [59] studied the effects of training
vary, which results in more or less resistive material be- with an encoder-equipped hand orthosis at home, finding
tween the contact points. As resistance varies in an ap- significant improvements in FMA score (4.9 ± 4.1 points).
proximately linear way with changes in arc length, it is
possible to map a direct relationship between resistance Implementation (requirements and challenges)
and angular displacement. This means that aligning the The advantage of not needing to apply machine learning
knob to the rotation axis of a joint allows a good estima- algorithms notwithstanding, the need of a parallel struc-
tion of its angular position. Encoders are optical sensors ture (e.g., exoskeleton) or embedding them in a glove re-
containing a slitted disc. A LED (light-emitting diode) stricts the range of applications these sensors may have
shines against the disc, which allows light to pass for stroke patients. Donning and doffing equipment
through the slits but blocks it otherwise. Presence and might be challenging for patients with low dexterity or
absence of light, detected by a photosensitive compo- high spasticity [60].
nent, is encoded into ones and zeroes and is used to de-
termine angular displacement. Potentiometers are Conductive elastomer (CE) and other flexible sensors
analog sensors with “infinite” resolution, whereas en- Conductive Elastomer (CE) sensors are flexible compo-
coders can have resolutions as high as 1 million counts nents with varying piezo-resistivity. Piezo-resistivity
per revolution [79]. Figure 3 shows an encoder mounted changes due to deformations suffered by a textile substrate
on a hand orthosis to track the fingers’ angular position. deposited with conductive particles (e.g. silver nanoparti-
cles). When placed along a moving body part, such as fin-
gers, it is possible to map the sensor readout related to a
Diagnostics
particular deformation of joint angles. Figure 4 shows an
Encoders and potentiometers can be used in clinical en-
vironments to measure ROM in patients. Researchers at
Peter S. Lum’s lab [80, 81] built an orthosis consisting of
four bars coordinating the movement of the metacarpo-
phalangeal finger joints and the thumb metacarpopha-
langeal joint for home-based training in stroke patients,
using encoders to calculate the joint angles.
Recovery/adaptation evaluation
Chen and Lum [82] focused on an “assists as needed”
approach, using a combination of potentiometers and
Fig. 4 Flexible sensors (red) laid along the fingers. Their flexion
Fig. 3 Encoder (blue) mounted on a hand orthosis, aligned with the results in piezo-resistive changes in the conducting material (e.g.
rotation axis of the index finger. This configuration allows tracking silver nanoparticles), which map directly to different finger positions.
angular displacement of fingers supported by the orthosis Prototype IMU sensor glove by Noitom [84]
example of flexible sensors tracking the position of indi- rehabilitative process relies on three main elements: 1.
vidual finger movements. Diagnosis, in which clinicians use standardized scales to
estimate maximum recovery for every patient [94] and
Diagnostics assign them to rehabilitation therapies accordingly [95].
Yu and colleagues used flexible sensors in combination 2. Evaluation of recovery or adaptation, during which cli-
with IMUs to assess motor function [61], and obtained nicians assess the extent up to which patients can per-
results bearing a high correlation (0.92) with clinical form activities of daily living. 3. Extended training,
scores given by a therapist. Flex sensors are frequently necessary for patients with persistent motor impairment
used as “gold standard” when attempting measurements after entering the chronic stage.
with others setups (e.g. [85]). Conventional motor assessment is vulnerable to
biases derived from measurement errors [96] and ceil-
Recovery/adaptation evaluation ing effects [97], whereas compensatory strategies fre-
Movement tracking using deformable sensors embedded quently adopted by patients while performing different
into clothes would allow monitoring patients as they tasks [11] can complicate the appraisal of recovery.
perform activities of daily living. For example, Tognetti Therapy and training provision at healthcare centers is
et al. [86] embedded CE sensors into clothing with the limited to available resources and restricted by its cor-
objective of classifying body postures and hand gestures responding costs, which obstructs prolonged rehabilita-
(with a reported sensitivity of 11,950 Ω/mm), a work fur- tive training for patients who do not recover fully
ther developed by Giorgino et al. [87, 88] and De Rossi within the first months after stroke.
[89]. A more complex system, combining this technol- A promising option to assess stroke patients objectively
ogy with EMG and IMU data was presented by Lorussi resides in the use of wearable technology. As high-end
et al. [90]. The use of piezo-resistive fabric [91] and sensors become more accessible, more reliable and less
fabric-based microelectromechanical systems (MEMS) obtrusive, the chance of acquiring relevant data during pa-
[92] offer alternatives to CE sensors. All these studies tients’ training or daily routines gets easier. A variety of
show promise in the use of flexible sensors embedded in wearable sensors (e.g. [29, 49, 59, 60, 62, 98]) have been
clothing to monitor stroke patients, but testing with used to assess several aspects of motor performance in
stroke patients is still lacking. stroke patients, going from motor impairment to more
subtle forms of behavior, such as limb neglect.
Extended training In the present paper, we seek to compare different setups
Prange-Lasonder and colleagues [60] implemented a with the intention of finding the most promising candidates
gamified form of a rehabilitative training using a glove for different applications. There are four main wearable sen-
equipped with flexible sensors, and studied the effects of sors used in the study of stroke: IMUs, EMG, potentiome-
such training at home [93]. Their results proved the ters/encoders and flexible sensors. IMUs allow measuring
feasibility of this approach as a home-based therapy, changes in acceleration, inclination and orientation unobtru-
even though they did not find significant differences in sively. Wireless, energy-efficient [42] transmission of data
comparison to their control intervention. characterizing these sensors enables whole-body recordings
through sensor networks [34], supporting this sensors’ can-
Implementation (requirements and challenges) didacy for movement tracking [28, 35, 49]. Several groups
Flexible sensors embedded into clothing constitute an have used IMUs with diagnostic purposes [19–27] and to
attractive option for unobtrusively tracking movements assess the execution of daily-life activities [19, 28–33]. High
in stroke patients during motor assessments, execution portability and accessible costs further support these sensors
of daily living activities, and rehabilitative training. At as an option for prolonged training during the chronic stage
present, their use in clinical environments and in-home (e.g. at home) [39]. There are general complications
settings is difficult due to practical issues related to don- inherent to the use of these devices, such as estima-
ning, doffing and washing the garments. Furthermore, tion errors derived from accumulated error in the cal-
some sensors require a large amount of wiring [91], culation of orientation from angular acceleration (i.e.
which reduces the degree of unobtrusiveness. Addition- orientation drift [41]) and quantization noise [50]. In
ally, mechanical deformations resulting from, for ex- addition, high movement variability in stroke patients,
ample, wrinkles in the fabric [88] introduce noise to the resulting from adopted compensatory muscle syner-
system, complicating posture and movement tracking. gies and slower, segmented movements [11], compli-
cate data characterization and comparison.
Discussion EMG wearable sensors have also been used for diagno-
Stroke is a frequent disorder that often results in long- sis [27, 62] and first attempts at extended training
lasting loss of motor functions. After stroke, the outside clinical environments [57]. Monitoring the
execution of activities of daily living can benefit from information (e.g. muscle activity) [10] can be acquired
EMG recordings, as these sensors allow capturing using EMG [62]. The best candidate to identify hand ges-
differences in muscle pattern activations resulting from tures (e.g. for orthotic control) amongst the sensors dis-
compensatory movements [68]. These sensors can com- cussed here is likely EMG. EMG allows identifying hand
plement the information obtained with IMUs. Aspects gestures effectively without altering too much the way in
neglected by some assessment scales (e.g. FMA), such as which patients interact with the environment, as would be
applied force [10], can be derived from muscle activation the case with potentiometers and flexible sensors. A pos-
as recorded with EMG. EMG sensors are susceptible to sible alternative would be the use of pressure sensors
different sources of noise, which must be removed be- [100]; Sadarangani and colleagues [98] tried this approach
fore signals can be used [70]. Furthermore, variable with stroke patients and achieved classification accuracies
placement of electrodes can also mislead estimations above 90% (3 classes only). We excluded this type of sen-
and affect the performance of the models used to classify sor from the present review because there is, to the best of
measured activity. our knowledge, no wearable version yet.
Potentiometers and encoders are robust to noise and
require little processing of signals, as the output from Data processing: recommendations
these sensors can be mapped directly to angular dis- As mentioned earlier, the analyses pipeline depends
placement (or linear, in the case of linear encoders). The heavily on the object of study (e.g. movement quality,
range of applications in stroke for these sensors is lim- limb neglect, etc.). There are multiple features to
ited to measuring ROM of limbs, and requires mounting characterize EMG and IMU signals for later classifica-
them on a parallel structure, such as an orthosis, limiting tion (e.g. into classes related to motor function), and the
the degrees of freedom of measured movements. Still, choice depends on the property of interest. For example,
their potential in extensive home-based training is clear muscle force is well- represented using the RMS of the
[59]. The need for an orthosis disappears with the use of EMG signal, whereas movement quality can be better
linear encoders [83] due to integration of the sensors observed by calculating jerk (rate of change in acceler-
into gloves. Nevertheless, the use of both orthoses and ation, capturing movement smoothness) from IMU data.
gloves can be difficult for patients suffering from hand Alternatively, comparing waveforms directly requires ei-
spasticity, which would complicate their use at home. ther normalizing the length of the time series or some-
This problem persists whenever using flexible sensors how matching them to account for different signal
embedded in gloves. Flexible sensors embedded in cloth- durations, such as with DTW.
ing could be a viable option for tracking everyday life ac- For classification problems, it might be better to have
tivities, but practical issues related to washing the many features and then trim them down by means of
garments and to the large amount of wiring required still PCA or other relevance determination algorithms (e.g.
impede their regular use. RRelief). This is a necessary step, as dataset sizes are
As IMU and EMG data cannot be mapped directly often quite small, and keeping too many features might
into the movements and actions that generated them, result in models not generalizing to new data (overfit-
acquired signals must be processed differently. Depend- ting). The choice for the model depends on the applica-
ing on the objective (e.g. assign grades to movements, tion and on its final objective. Several studies discussed
compare patients to healthy controls, etc.) data can ei- in Table 1 used SVM in classification, and some of them
ther be classified using different forms of statistical pro- reported testing more than one model, but this choice is
cessing, such as common methods applied in machine not compulsory. For example, if the objective is to de-
learning [99], or compared using algorithms like DTW ploy an automated tool for assessment of motor function
[48, 49]. Built models often fail to generalize to data from and the ultimate goal is for it to reliably assess function-
highly impaired patients due to lower signal-to-noise ratio ality, many different models can be tested and optimized
(SNR) [55]. Further, results are hard to compare due to a to find the best performer. Alternatively, applications
lack of a unified data acquisition protocol [73]. such as allocating patients to different therapies based of
their specific needs (i.e. individualized care) might bene-
Choosing an adequate setup fit from transparent, easily explained models such as de-
The choice for the best setup depends on the intended ap- cision trees, as the rationale behind a choice for therapy
plication. The best candidate to study movement quality is important.
while remaining unobtrusive and easy to deploy is likely The way in which models are fine-tuned and validated
IMUs. Data from IMUs provide enough information to is an important aspect too. Several studies shown in
characterize movement execution (e.g. [49]), detect limb Table 1 claim performing cross-validation, but its actual
neglect and assess performance of activities of daily life implementation varies a lot between studies. A good ap-
[28]. During motor assessments, overlooked functional proach is to separate a portion of the data as test data
and leave it “untouched” until after fine-tuning the problematic, prices are high, and there are no reports on
model using the remaining data (i.e. training data). Once how well they perform.
more, a fraction of these data is set aside, this time as
validation data, while using the rest to fit the model. Re- Alternatives and possibilities
peating this process with the training data and averaging Easily deployed, inexpensive IMU devices are available off-
(or “voting”, i.e. selecting most frequent labels) the re- the-shelf. Mbientlab [17], for example, offers a wide array
sults will yield a less-biased model. Subsampling of data of what seems to be modular and flexible IMU setups
for every iteration can be done with replacement (bag- allowing prolonged recordings with multiple sensors sim-
ging) or without (pasting). After fine-tuning the model’s ultaneously. Beange and colleagues [107] compared one of
parameters, plugging-in the test data gives a more realis- the IMU modules to a motion capture system and found
tic impression of how well the model will generalize to its performance acceptable. High- end systems such as the
new data. In the end, results obtained will depend on Xsens [14] perform excellently, but their prohibitive cost
the quality of used features and on the amount of infor- limits the range of possible applications; such a system
mation contained in them. The optimization of the could only be used for measurements in high-end, special-
models is relatively trivial, in the sense that there are ized clinics, failing to solve the problem of limited re-
many available tools to do so. Time and effort must be sources of common healthcare centers.
invested in feature engineering, as models can only per- As for the acquisition of EMG data, we were not able
form as well as the quality of the information used to to find a low-cost solution providing quality data while
build them. remaining simple to use. Systems built by companies like
In general, the more data is available to train models, Noraxon [65], Delsys® (Natick, Massachusetts, USA)
the better. The most effective algorithms used across do- [108] or Cometa® (Bareggio, Milan, Italy) [109] provide
mains, such as neural networks, are only useful if used high quality data, but at a high cost. Less expensive sys-
on large amounts of data. For this reason, initiatives like tems like “FreeEMG” [66] or “Biometrics’ sEMG sensors”
the “NinaPro” database [101] should be supported and (Biometrics Ltd.®, Newport, UK) [110] are more access-
contributed-to, such that data acquired on different sites ible, but are still suboptimal in the sense of requiring
might be pooled together. Data acquisition and sharing careful placement of gel electrodes, which makes it im-
between different sites brings along its own challenges practical for unsupervised patient use at home.
and escapes the scope of this review, but standardized Presenting a similar design to that of the “Myo” arm-
protocols like the “NinaPro” and guidelines for sensor band, Yang and colleagues [111] built a bracelet equipped
placement (e.g. [76]) will be crucial towards this effort. with textile electrodes, reporting high classification accur-
acy (close to 100%) in hold-out cross-validation. The study
An empty niche involved only three healthy participants, and training and
An EMG + IMU device that had been gaining momen- testing data used in cross-validation came from the same
tum in multiple scientific domains was the “Myo” arm- subject (no inter-subject validation). Still, the design of
band (Thalmic Labs®, Kitchener, Ontario, Canada) [102]. this device seems promising.
This device consists of an array of eight single differen- A different approach trying to enhance EMG systems
tial electrodes and a 9-axis IMU, presented as a bracelet, with near-infrared spectroscopy (NIRS) was taken ini-
transmitting data through BLE. Its affordability and tially by Herrmann and Buchenrieder [112] in an at-
user-friendliness made it an attractive alternative for tempt to reduce electrode crosstalk. This approach was
prolonged, possibly unsupervised recordings. Further- also pursued by a couple other groups [113, 114], but
more, a formal comparison between this armband and challenges related to the time resolution of NIRS limit
several high-end EMG systems showed similar classifica- the applications possible for these devices.
tion accuracies when using two armbands at the same Interesting possibilities exist in the realm of printable
time [73] to classify signals into 40 different movements, (i.e. epidermal electrodes [115]) and temporary tattoo
further supporting the use of this device in research. Ap- electrodes [116], but these are not yet readily avail-
plications for motor assessments [27], orthotic [63] and able for deployment. For the time being, the choice
prosthetic [103] control, gesture recognition [104], etc. of a device to acquire inertial and EMG data simul-
have benefited from this device. CTRL-Labs® (New York taneously in an inexpensive, easy to deploy fashion
City, New York, USA) [105] is developing a new device remains an open question.
combining these sensors, but this important niche is, at
present, unattended. Some institutions in China have From bench to bedside
started selling products significantly inspired by the Wearable sensors in clinical environments
“Myo”, such as OYMotion® (Beijing, China) [106], but The processing steps and the implementation challenges
their acquisition in Europe and America can be described before may appear daunting when thinking
about integrating these sensors into clinical practice. Escorpizo and colleagues [123] proposed two main ac-
The importance of discussing these challenges lies in the tions towards the integration of the ICF into clinical
joint effort towards democratizing these technologies practice, one of which was the use of the ICF’s Core sets
such that their advantages might be widespread, access- for specific conditions, which contains a list of categories
ible to all, their performance and reliability ensured. To describing the most salient aspects of disability related
achieve this goal further research is necessary, and re- to these. In this case, some of the components belonging
search can greatly benefit from knowledge acquired in to body functions (i.e. muscle power), and activities and
the clinic. participation (e.g. walking, eating, dressing) of the Core
There is a variety of readily available systems dedicat- Set defined for stroke [124], could be assessed using
ing wearable sensors to rehabilitation. For example, the wearable sensors.
“ArmeoSenso” system (Hocoma®, Volketswil, Zürich, The ICF seeks to provide comparable/replicable statis-
Switzerland) [117] uses IMU’s alongside a gamified form tics of disability as a whole. The ICF’s performance and
of training (this is the system used for home-based train- capacity qualifiers describe activities of daily living in
ing, mentioned before [39]). For EMG, products like natural environments and execution of specific tasks, re-
Cometa’s “EMG Easy Report” [118] or Noraxon’s “myo- spectively, which correspond to the “Diagnostics” and
Muscle” [119] allow simplified analyses, like pairing re- the “Evaluation of Recovery/Adaptation” dimensions de-
cordings to video, to look at muscle activity related to scribed before for each sensor type. The “Extended
specific movements. The use of these systems in the Training” dimension addresses some social factors like
clinic provides further insights into practical aspects to degree of independence and integration to society by
consider when developing new products, and allows allowing patients to continue recovering after leaving the
fitting these technologies to the patients’ needs. Their rehabilitation facility.
functionality may be limited to certain aspects and Baets and colleagues [125] reviewed the literature on
system errors might display these techniques as less effi- shoulder assessment by means of IMUs, in the context
cient than conventional approaches, but the develop- of the ICF. They found that even though some measured
ment of flexible and robust systems requires this sort of aspects were repeatable and useful in this context, more
iterative testing in real-life situations, enriched with the work is needed to generate clinically meaningful, repeat-
knowledge of specialized medical personnel. Even if the able information. Standardizing measurements to
transition towards the integration of these devices into characterize performance and capacity qualifiers, as de-
clinical practice represents an extra effort on an already scribed by the ICF, will also allow leveraging these data-
strained environment, it has potential at reducing costs sets for the application of more complex analyses
once they become ubiquitous. requiring larger amounts of data (e.g. neural networks).
Hughes and colleagues [120] reported that one of the
main obstacles in the way of adopting these type of tech-
nologies in clinics is the lack of awareness about their Economic impact of stroke and potential benefits from
existence, which calls for better communication and col- wearable devices
laboration between researchers and clinicians. The European Union spends €45 billion on treating
stroke patients every year, with 44% of these costs spent
on direct health care, 22% related to productivity losses
The international classification of functioning, disability and and 35% on informal care of patients [126]. Care after
health (ICF) stroke depends on how involved institutions (govern-
The ICF is an important and well-established tool in clin- ments, healthcare centers, insurance companies, etc.)
ical neuro-rehabilitation and seeks to provide a framework manage their resources [127], which influences the
based on two models of disability, one coming from indi- length of stay in the hospital and the extension of thera-
vidual factors and another from social factors [121]. This peutic care [128]. For instance in the United States,
biopsychosocial model provides standardized grounds for “Medicare” [129] has strict rules for the provision of in-
studying, understanding and addressing disability. Metcalf tensive inpatient rehabilitation therapies (i.e. at least 3 h
and colleagues [122] assessed which of the most fre- per day, 5 to 6 days per week), with an average length of
quently used scales of motor function in stroke patients stay of 15 days, at which point 70% of patients are sent
better fit the framework of the ICF in terms of repeatabil- home [130]. This percentage goes up to 90% after 3
ity and reliability, rating as most reliable those test involv- months, and if patients have not recovered enough to be
ing numerical assessments such as ROM and movement cared for at home by then, they will either receive more
time. Using wearable sensors during regular assessments restricted healthcare coverage from state-based payers
will then improve performance of standardized motor as- (e.g. “Medicaid”) or be sent to nursing homes where they
sessments in the framework of the ICF. will receive limited rehabilitation [130].
A study in Switzerland revealed that 37% of direct at home. Burridge and colleagues [138] discuss the ef-
health care costs after stroke correspond to rehabilita- fectiveness of some home-based rehabilitation systems
tion at the clinic [131]. Using systems like Hocoma’s and show that this approach is feasible and has the po-
“ArmeoSenso” [117] could allow patients to train in tential to improve motor function by training daily at
groups, which besides allowing therapists to tend to home. They also present a new system (the “M-Mark”),
more people simultaneously, could bring enhanced ef- which will allow patients to train at home under
fects of rehabilitation (e.g., [132]), rendering it more different circumstances of daily life (e.g. placing objects
cost-effective. Motor assessments could be made more on a kitchen shelf) while being tracked by IMUs and
agile through wearable sensors, and patients could do it mechanomyography.
without a therapist being present (e.g. at home).
Results from meta-analyses have shown that early
planned and coordinated hospital discharge combined Practical considerations
with home-based rehabilitation yields better results, and There are many aspects to look into for home-based re-
home-based rehabilitation was found to be superior to habilitative training and its corresponding assessments
center-based, as measured by the Bartel Index 6 months and measurements. First, training must be thoroughly
after stroke [133]. Healthcare coverage of home-based and carefully explained to patients and, when applicable,
services can limit the length of therapy provided, but the to their caregivers. An option is to provide center/lab-
use of wearable sensors for home-based therapy could based training for a short amount of time and then in-
grant access to these enhanced benefits while keeping struct patients to train at home [139]. Further, provided
costs low. Extended recovery resulting from home-based equipment must be as simple to use as possible to re-
rehabilitative training (discussed in the next sub-section) duce chance of making mistakes and ensure training ad-
could also increase the level of independence in patients, hesion. An example of how possible mistakes can be
which would decrease costs related to productivity losses reduced in a home-based environment can be found in
and informal care. the work of Durfee et al. [58], like blocking elements not
useful to users (e.g. parts of the keyboard).
Home-based self-application of rehabilitative training Another important aspect to consider is data logging.
Evidence of enhanced recovery related to more extensive One option is to keep all data on the devices and extract
training has been found in stroke patients [12], but high it once the participants give the devices back at the end
costs inherent to provided care, such as patient trans- of their study contribution [31]. Nevertheless, this pre-
portation or the therapy itself (i.e. therapists’ salary, re- sents a risk with longer studies, as devices are lent for
habilitation site, etc.), often limit the therapies’ duration longer periods, and any accident damaging the device
and frequency. On the other hand, training in more fa- would result in loss of all previously gathered data. An
miliar environments, such as at home, improves the ef- alternative would be to relay the data to a protected ser-
fects of training [134]. Training transfer to different ver [61]. This could be challenging whenever partici-
environments, in general, is highly reduced [135], which pants’ homes are located in relatively isolated areas, with
is why training tasks should resemble activities of daily poor internet connection. Mobile broadband modules
life, and take place at locations where they would occur could solve this issue, although constraints from tele-
on a daily basis. communications companies providing the service still
Unsupervised, home-based rehabilitative training has exist. Ultimately, it is most likely best to store data both
the potential to largely improve outcome of rehabilita- on the devices and on a server, in a redundant manner.
tion in patients [136, 137]. Home-based training offers Even though home-based training offers beneficial
many advantages, but reducing contact between trainers possibilities in terms of high-intensity training, other as-
and beneficiaries could impact motivation and engage- pects, such as motivation derived from human inter-
ment, which play a major role in recovery [5]. Thera- action [5] might be lacking. For this reason, taking
pists’ expertise would still be necessary to determine and advantage of virtual conference tools (e.g. “Skype” [140])
adjust therapies, as well as to follow-up on training and could allow therapists to provide feedback and motivate
rehabilitation progress, but contact between therapists patients, as well as to acquire feedback. A recent report
and those under their care could be less frequent. This by Maceira-Elvira and colleagues [141] discusses some
complicates the assessment of training quality and proof the challenges and important aspects to take into ac-
gress evaluation over shorter periods (daily, weekly), count in home-based training. The report highlights the
which might impact on motivation, planning of the importance of remote assistance and proper instructions
intervention and personalized adaption of the treatment provided to users, as well as technical assistance around
strategy [134]. Careful consideration of these potential the clock. Another report by Van de Winckel and col-
threats is paramount to provide effective rehabilitation leagues [142] provides valuable information about the
(generally positive) opinion of six patients enrolled on Availability of data and materials
remotely-monitored home-based training. Not applicable.
Ethics approval and consent to participate

Conclusion Not applicable.
Stroke rehabilitation is an iterative process involving im-
pairment assessment, recovery prognosis, therapy defin- Consent for publication
ition, rehabilitative training and monitoring of functional Not applicable.
changes. Conventional assessments of motor function

Competing interests
face limitations due to several factors, resulting in biased The authors declare that they have no competing interests.
predictions of recovery, which prevent an adequate as-
signment of treatment for patients. Furthermore, limited Author details
1
Defitech Chair in Clinical Neuroengineering, Center for Neuroprosthetics
resources at rehabilitation centers and clinics prevent (CNP) and Brain Mind Institute (BMI), Swiss Federal Institute of Technology
patients from receiving intensive treatment and exten- (EPFL), 9, Chemin des Mines, 1202 Geneva, Switzerland. 2Defitech Chair in
sive attention, frequently reducing the degree up to Clinical Neuroengineering, Center for Neuroprosthetics (CNP) and Brain Mind
Institute (BMI), Swiss Federal Institute of Technology (EPFL Valais), Clinique
which they recover. Wearable sensors show promise re- Romande de Réadaptation, 1951 Sion, Switzerland. 3Clinical Neuroscience,
solving at least some of these problems. Regular assess- University of Geneva Medical School, 1202 Geneva, Switzerland.
ments complemented with this technology can reduce
Received: 2 August 2019 Accepted: 24 October 2019
bias in measurements and estimations, as well as reduce
assessment time for therapists. Short-term rehabilitative
training, offered during the first 6 months after stroke, References
could be prolonged by offering home-based therapies, 1. Feigin VL, Norrving B, Mensah GA. Global burden of stroke. Circ Res. 2017;
120(3):439–48.
designed and monitored remotely by therapists, allowing 2. Vos T, Allen C, Arora M, Barber RM, Brown A, Carter A, et al. Global, regional,
patients to train in a familiar environment. Among the and national incidence, prevalence, and years lived with disability for 310
wide array of sensors available, inertial measurement diseases and injuries, 1990–2015: a systematic analysis for the global burden
of disease study 2015. Lancet. 2016;388(10053):1545–602.
units (IMUs) and electromyography (EMG) offer the 3. Kwakkel G, Kollen BJ, Van der Grond JV, Prevo AJH. Probability of regaining
best balance between unobtrusiveness, robustness, ease dexterity in the flaccid upper limb: impact of severity of paresis and time
of use and data quality. An optimal solution comprising since onset in acute stroke. Stroke. 2003;34(9):2181–6.
4. Bernhardt J, Hayward KS, Kwakkel G, Ward NS, Wolf SL, Borschmann K, et al.
both sensor types is still lacking in the market, but the Agreed definitions and a shared vision for new standards in stroke recovery
collection of studies presented in this review indicate research: the stroke recovery and rehabilitation roundtable taskforce.
that this might be the most promising way to go. Neurorehabil Neural Repair. 2017;31(9):793–9.
5. Langhorne P, Bernhardt J, Kwakkel G. Stroke rehabilitation. Lancet. 2011;
Abbreviations 377(9778):1693–702. Available from:. https://doi.org/10.1016/S0140-
(s)EMG: (surface) Electromyography; (UE)FM: (Upper extremity) Fugl-Meyer; 6736(11)60325-5.
ARAT: Action research arm test; AUC: Area under the curve; BLE: Bluetooth 6. Chen S-Y, Winstein CJ. A systematic review of voluntary arm recovery in
Low-Energy; CE: Conductive elastomer; CPD: Coherent point drift; hemiparetic stroke: critical predictors for meaningful outcomes using the
DTW: Dynamic time warping; ELM: Extreme learning machines; international classification of functioning, disability, and health. J Neurol
FAS: Functional ability score; FMA: Fugl-Meyer Assessment; ICF: International Phys Ther. 2009;33(1):2–13.
Classification of Functioning, Disability and Health; IMU(s): Inertial 7. Poole JL, Whitney SL. Assessments of motor function post stroke: a review.
measurement unit(s); KNN: K-nearest neighbor; LASSO: Least absolute Phys Occup Ther Geriatr. 2001;19(2):1–22.
shrinkage and selection operator; LDA: Linear discriminant analysis; 8. Goldstein LB, Bertels C, Davis JN. Interrater reliability of the NIH stroke scale.
LED: Light-emitting diode; MEMS: Microelectromechanical systems; Arch Neurol. 1989;46:660–2.
MEP: Motor-Evoked Potential; NMF: Non-negative matrix factorization; 9. Fugl-Meyer AR, Jääskö L, Leyman I, Olsson S, Steglind S. The post-stroke
PCA: Principal component analysis; RBF: Radial basis functions; RMSE: Root hemiplegic patient- a method for evaluating physical performance.pdf;
mean square error; ROC: Receiver operating characteristic; ROM: Range of 1975.
motion; SVM: Support vector machines; TMS: Transcranial Magnetic 10. Gladstone DJ, Danells CJ, Black SE. The Fugl-Meyer assessment of motor
Stimulation; WMFT: Wolf motor function test recovery after stroke: a critical review of its measurement properties.
Neurorehabil Neural Repair. 2002;16(3):232–40. Available from:. https://doi.
org/10.1177/154596802401105171.
Acknowledgements 11. Cirstea MC, Levin MF. Compensatory strategies for reaching in stroke. Brain.
The authors would like to give special thanks to Una Pale for her valuable 2000;123(5):940–53.
comments concerning the choice for relevant information needed in the 12. Kwakkel G, Van Peppen R, Wagenaar RC, Dauphinee SW, Richards C,
table, particularly relevant for engineers among the readership of this journal. Ashburn A, et al. Effects of augmented exercise therapy time after stroke: a
meta-analysis. Stroke. 2004;35(11):2529–36.
Authors’ contributions 13. Morrison, Melvin M. 1987-12-08, US Patent No. 4711125A. Retrieved from
FCH detected the need for a review in this topic. PME reviewed the https://patents.google.com/patent/US4711125A/en.
literature, wrote the manuscript and prepared the illustrations. TP made 14. Xsens North America Inc. Xsens. 2018 [cited 2018 May 28]. Available from:
important suggestions about the topics to be covered. TP, ACS and FCH https://www.xsens.com/
revised the manuscript and provided valuable comments about the contents 15. Yang CC, Hsu YL. A review of accelerometry-based wearable motion
and narrative. All authors read and approved the final manuscript. detectors for physical activity monitoring. Sensors. 2010;10(8):7772–88.
16. Bellusci G, Dijkstra F, Slycke P. Xsens MTw : Miniature Wireless Inertial
Funding Motion Tracker for Highly Accurate 3D Kinematic Applications. Xsens
This work was supported by the Defitech Foundation (Morges, Project StS). Technol. 2013;(April):1–9.
17. Mbientlab Inc. Wearable technology for motion tracking. 2018 [cited 2018 40. Williamson J, Liu Q, Lu F, Mohrman W, Li K, Dick R, et al. Data sensing and
Jan 24]. Available from: https://mbientlab.com/ analysis: Challenges for wearables. 20th Asia South Pacific Des Autom Conf
18. Gor-García-Fogeda MD, Molina-Rueda F, Cuesta-Gómez A, Carratalá-Tejada ASP-DAC. 2015;2015:136–41.
M, Alguacil-Diego IM, Miangolarra-Page JC. Scales to assess gross motor 41. Wittmann F, Lambercy O, Gassert R. Magnetometer-based drift correction
function in stroke patients: a systematic review. Arch Phys Med Rehabil. during rest in IMU arm motion tracking. Sensors (Switzerland). 2019;19(6):
2014;95(6):1174–83. 13–7.
19. Hester T, Hughes R, Sherrill DM, Knorr B, Akay M, Stein J, et al. Using 42. Gomez C, Oller J, Paradells J. Overview and evaluation of bluetooth low
wearable sensors to measure motor abilities following stroke. Proc - BSN energy: an emerging low-power wireless technology. Sensors (Switzerland).
2006 Int work wearable implant body. Sens Networks. 2006;2006:5–8. 2012;12(9):11734–53.
20. Yu L, Xiong D, Guo L, Wang J. A compressed sensing-based wearable 43. Zhang M, Sawchuk AA. A feature selection-based framework for human
sensor network for quantitative assessment of stroke patients. Sensors activity recognition using wearable multimodal sensors. BODYNETS 2011 -
(Switzerland). 2016;16(2):202. 6th Int ICST Conf Body Area Networks. 2012;1:92–8.
21. Brunnstrom S. Motor testing procedures in hemiplegia. Phys Ther. 1966; 44. Mannini A, Sabatini AM. Machine learning methods for classifying human
46(4):357–75. physical activity from on-body accelerometers. Sensors. 2010;10(2):1154–75.
22. Patel S, Hughes R, Hester T, Stein J, Akay M, Dy J, et al. A novel approach to 45. Zhang Z, Fang Q, Ferry F. Upper limb motion capturing and classification
monitor rehabilitation outcomes in stroke survivors using wearable for unsupervised stroke rehabilitation. IECON Proc (Industrial Electron Conf).
technology. Proc IEEE. 2010;98(3):450–61. 2011:3832–6.
23. Patel S, Hughes R, Hester T, Stein J, Akay M, Dy J, et al. Tracking motor 46. Myronenko A, Song X. Point set registration: coherent point drift. IEEE Trans
recovery in stroke survivors undergoing rehabilitation using wearable Pattern Anal Mach Intell. 2010;32(12):2262–75.
technology. Conf Proc IEEE Eng Med Biol Soc. 2010;2010:6858–61. 47. Berndt D, Clifford J. Using dynamic time warping to find patterns in time
24. Parnandi AR, Wade E, Mataric M. Motor function assessment using wearable series. Work Knowl Knowl Discov Databases. 1994;398:359–70 Available from:
inertial sensors. Conf Proc IEEE Eng Med Biol Soc. 2010;2010:86–9. http://www.aaai.org/Papers/Workshops/1994/WS-94-03/WS94-03-031.pdf.
25. Del Din S, Patel S, Cobelli C, Bonato P. Estimating fugl-meyer clinical scores 48. Zhang M, Lange B, Chang CY, Sawchuk AA, Rizzo AA. Beyond the standard
in stroke survivors using wearable sensors. Proc Annu Int Conf IEEE Eng clinical rating scales: fine-grained assessment of post-stroke motor
Med Biol Soc EMBS. 2011:5839–42. functionality using wearable inertial sensors. Proc Annu Int Conf IEEE Eng
26. Wang J, Yu L, Wang J, Guo L, Gu X, Fang Q. Automated Fugl-Meyer Med Biol Soc EMBS. 2012:6111–5.
assessment using SVR model. 2014 IEEE Int Symp bioelectron 49. Zhang Z, Fang Q, Gu X. Objective assessment of upper-limb mobility for
bioinformatics. IEEE ISBB. 2014;2014:0–3. poststroke rehabilitation. IEEE Trans Biomed Eng. 2016;63(4):859–68.
27. Repnik E, Puh U, Goljar N, Munih M, Mihelj M, Repnik E, et al. Using Inertial 50. Buke A, Gaoli F, Yongcai W, Lei S, Zhiqi Y. Healthcare algorithms by
Measurement Units and Electromyography to Quantify Movement during wearable inertial sensors: a survey. China Commun. 2015;12(4):1–12.
Action Research Arm Test Execution. Sensors. 2018;18(9):2767 Available 51. Prathivadi Y, Wu J, Bennett TR, Jafari R. Robust activity recognition using
from: http://www.mdpi.com/1424-8220/18/9/2767. wearable IMU sensors. Proc IEEE Sensors. 2014;2014(December):486–9.
28. Lee SI, Adans-Dester CP, Grimaldi M, Dowling AV, Horak PC, Black-Schaffer 52. Strohrmann C, Labruyère R, Gerber CN, van Hedel HJ, Arnrich B, Tröster G.
RM, et al. Enabling Stroke Rehabilitation in Home and Community Settings: Monitoring motor capacity changes of children during rehabilitation using
A Wearable Sensor-Based Approach for Upper-Limb Motor Training. IEEE J body-worn sensors. J Neuroeng Rehabil. 2013;10:83 Available from: http://
Transl Eng Heal Med. 2018;6(May):1-11. dx.doi.org/10.1186/1743-0003-10-83%5Cnhttps://jneuroengrehab.
29. Leuenberger K, Gonzenbach R, Wachter S, Luft A, Gassert R. A method to biomedcentral.com/articles/10.1186/1743-0003-10-83%5Cnhttps://
qualitatively assess arm use in stroke survivors in the home environment. jneuroengrehab.biomedcentral.com/track/pdf/10.1186/1743-0003-10-
Med Biol Eng Comput. 2017;55(1):141–50. 83?site=jneuroengrehab.biomedcentral.com.
30. De Lucena DS, Stoller O, Rowe JB, Chan V, Reinkensmeyer DJ. Wearable 53. Bailey RR, Klaesner JW, Lang CE. Quantifying real-world upper-limb activity
sensing for rehabilitation after stroke : Bimanual jerk asymmetry encodes in nondisabled adults and adults with chronic stroke. Neurorehabil Neural
unique information about the variability of upper extremity recovery. IEEE Repair. 2015;29(10):969–78.
Int Conf Rehabil Robot. 2017;2017:160;2013–8. 54. Leuenberger K, Gassert R. Low-power sensor module for long-term activity
31. Lang CE, Waddell KJ, Klaesner JW, Bland MD. A Method for Quantifying Upper monitoring. Proc Annu Int Conf IEEE Eng Med Biol Soc EMBS. 2011:2237–41.
Limb Performance in Daily Life Using Accelerometers. J Vis Exp. 2017;122:122. 55. Lee S, Wilson KM, Lock BA, Kamper DG. Subject-specific myoelectric pattern
[cited 2017 Aug 18];5567337915(12210). Available from: www.jove.com. classification of functional hand movements for stroke survivors. IEEE Trans
32. Zhou H, Huosheng H, Harris N, Hu H, Harris N, Huosheng H, et al. Neural Syst Rehabil Eng. 2011;19(5):558–66.
Application of wearable inertial sensors in stroke rehabilitation. Eng Med 56. Zhang X, Zhou P. High-density myoelectric pattern recognition toward
Biol Soc 2005 IEEE-EMBS 2005 27th Annu Int Conf. 2005;1(4):6825–8. improved stroke rehabilitation. IEEE Trans Biomed Eng. 2012;59(6):1649–57.
33. Capela NA, Lemaire ED, Baddour N. Feature selection for wearable 57. Donoso Brown EV, Dudgeon BJ, Gutman K, Moritz CT, McCoy SW.
smartphone-based human activity recognition with able bodied, elderly, Understanding upper extremity home programs and the use of gaming
and stroke patients. PLoS One. 2015;10(4):1–18. technology for persons after stroke. Disabil Health J. 2015;8(4):507–13.
34. van Meulen FB, Klaassen B, Held J, Reenalda J, Buurke JH, van Beijnum B-JF, Available from:. https://doi.org/10.1016/j.dhjo.2015.03.007.
et al. Objective Evaluation of the Quality of Movement in Daily Life after 58. Durfee W, Carey J, Nuckley D, Deng J. Design and implementation of a
Stroke. Front Bioeng Biotechnol www.frontiersin.org. 2016;3(January):1–11. home stroke telerehabilitation system. Proc 31st Annu Int Conf IEEE Eng
35. Zhang Z, Liparulo L, Panella M, Gu X, Fang Q. A fuzzy kernel motion Med Biol Soc Eng Futur Biomed EMBC. 2009;2009:2422–5.
classifier for autonomous stroke rehabilitation. IEEE J Biomed Heal 59. Chen J, Nichols D, Brokaw EB, Lum PS. Home-based therapy after stroke
Informatics. 2016;20(3):893–901. using the hand spring operated movement enhancer (HandSOME). IEEE
36. Arteaga S, Chevalier J, Coile A, Hill AW, Sali S, Sudhakhrisnan S, et al. Low-cost Trans Neural Syst Rehabil Eng [Internet]. 2017;20010(c):1–1 Available from:
accelerometry-based posture monitoring system for stroke survivors. Proc 10th http://ieeexplore.ieee.org/document/7903687/.
Int ACM SIGACCESS Conf Comput Access - Assets ‘08, vol. 243; 2008. Available 60. Prange-Lasonder GB, Radder B, Kottink AIR, Melendez-calderon A, Buurke JH,
from: http://portal.acm.org/citation.cfm?doid=1414471.1414519 Rietman JS. Applying a soft - robotic glove as assistive device and training
37. Ding ZQ, Luo ZQ, Causo A, Chen IM, Yue KX, Yeo SH, et al. Inertia sensor-based tool with games to support hand function after stroke : preliminary results
guidance system for upperlimb posture correction. Med Eng Phys. 2013;35(2): on feasibility and potential clinical impact. 2017;(Il):1401–1406.
269–76. Available from:. https://doi.org/10.1016/j.medengphy.2011.09.002. 61. Yu L, Xiong D, Guo L, Wang J. A remote quantitative Fugl-Meyer
38. Galiana I, Iii FLH, Howe RD. Wearable Soft Robotic Device for Post-Stroke assessment framework for stroke patients based on wearable sensor
Shoulder Rehabilitation : Identifying Misalignments; 2012. networks. Comput Methods Programs Biomed. 2016;128(88):100–10.
39. Wittmann F, Held JP, Lambercy O, Starkey ML, Curt A, Höver R, et al. Self- Available from:. https://doi.org/10.1016/j.cmpb.2016.02.012.
directed arm therapy at home after stroke with a sensor-based virtual reality 62. Li Y, Zhang X, Gong Y, Cheng Y, Gao X, Chen X. Motor function evaluation
training system. J Neuroeng Rehabil. 2016;13(1):75 Available from: http:// of hemiplegic upper-extremities using data fusion from wearable inertial
jneuroengrehab.biomedcentral.com/articles/10.1186/s12984-016-0182-1. and surface EMG sensors. Sensors (Switzerland). 2017;17(3):582.
63. Ryser F, Tobias B, Held JP, Lambercy O, Gassert R. Fully Embedded Myoelectric 89. De Rossi D, Carpi F, Lorussi F, Scilingo E, Tognetti A. Wearable kinesthetic
Control for a Wearable Robotic Hand Orthosis. Int Conf Rehabil Robot. 2017:615-21. systems and emerging technologies in actuation for upper limb
64. Ives JC, Wigglesworth JK. Sampling rate effects on surface EMG timing and neurorehabilitation. 2009 31st Annu Int Conf IEEE Eng Med Biol Soc 2009.
amplitude measures. Clin Biomech. 2003;18(6):543–52. 2009:6830–3 Available from: inspec:10983776.
65. Noraxon. Movement - Data - People. 2018 [cited 2019 May 13]. Available 90. Lorussi F, Carbonaro N, De Rossi D, Paradiso R, Veltink P, Tognetti A.
from: https://www.noraxon.com/ Wearable textile platform for assessing stroke patient treatment in daily life
66. BTSBioengineering. FreeEMG. 2017 [cited 2019 Jan 30]. Available from: conditions. Front Bioeng Biotechnol. 2016;4(March):28 Available from: http://
https://www.btsbioengineering.com/ www.ncbi.nlm.nih.gov/pubmed/27047939.
67. Turton A, Wroe S, Trepte N, Fraser C, Lemon RN. Contralateral and ipsilateral 91. Carbonaro N, Mura GD, Lorussi F, Paradiso R, De Rossi D, Tognetti A.
EMG responses to transcranial magnetic stimulation during recovery of arm Exploiting wearable goniometer technology for motion sensing gloves. IEEE
and hand function after stroke. Electroencephalogr Clin Neurophysiol - J Biomed Heal Informatics. 2014;18(6):1788–95.
Electromyogr Mot Control. 1996;101(4):316–28. 92. Yuen AC, Bakir AA, Rajdi NNZM, Lam CL, Saleh SM, Wicaksono DHB.
68. Cheung VCK, Piron L, Agostini M, Silvoni S, Turolla A, Bizzi E. Stability of Proprioceptive sensing system for therapy assessment using cotton fabric-based
muscle synergies for voluntary actions after cortical stroke in humans. Proc biomedical microelectromechanical system. IEEE Sensors J. 2014;14(8):2872–80.
Natl Acad Sci U S A. 2009;106(46):19563–8. 93. Nijenhuis SM, Prange-Lasonder GB, Stienen AHA, Rietman JS, Buurke JH. Effects
69. Kim JH. The effects of training using EMG biofeedback on stroke patients of training with a passive hand orthosis and games at home in chronic stroke:
upper extremity functions. J Phys Ther Sci. 2017;29(6):1085–8. a pilot randomised controlled trial. Clin Rehabil. 2017;31(2):207–16.
70. Reaz MBI, Hussain MS, Mohd-Yasin F. Techniques of EMG signal analysis: detection, 94. Prabhakaran S, Zarahn E, Riley C, Speizer A, Chong JY, Lazar RM, et al. Inter-
processing, classification and applications. Biol Proced Online. 2006;8(1):11–35. individual variability in the capacity for motor recovery after ischemic stroke.
71. Zardoshti-Kermani M, Wheeler BC, Badie K, Hashemi RM. EMG feature Neurorehabil Neural Repair. 2008;22(1):64–71.
evaluation for movement control of upper extremity protheses. IEEE Trans 95. Duncan PW, Zorowitz R, Bates B, Choi JY, Glasberg JJ, Graham GD, et al.
Rehabil Eng. 1995;3(4):324–33. Management of Adult Stroke Rehabilitation Care: A Clinical Practice
72. Phinyomark A, Phukpattaranont P, Limsakul C. Feature reduction and Guideline, vol. 36; 2005. p. 100–43. Available from: http://stroke.ahajournals.
selection for EMG signal classification. Expert Syst Appl. 2012;39(8):7420–31. org/cgi/doi/10.1161/01.STR.0000180861.54180.FF
Available from:. https://doi.org/10.1016/j.eswa.2012.01.102. 96. Krakauer JW, Marshall RS. The proportional recovery rule for stroke revisited.
73. Pizzolato S, Tagliapietra L, Cognolato M, Reggiani M, Müller H, Atzori M. Ann Neurol. 2015;51:845–7.
Comparison of six electromyography acquisition setups on hand movement 97. Hope TMHH, Friston K, Price CJ, Leff AP, Rotshtein P, Bowman H. Recovery
classification tasks. PLoS One. 2017;12(10):1–17. after stroke: not so proportional after all? PLoS One. 2018;11(2):1–22
74. Palermo F, Cognolato M, Gijsberts A, Müller H, Caputo B, Atzori M. Available from: http://dx.doi.org/10.1101/306514%0Ahttps://www.biorxiv.
Repeatability of grasp recognition for robotic hand prosthesis control based org/content/early/2018/05/15/306514.
on sEMG data. IEEE Int Conf Rehabil Robot. 2017:1154-9. (Submitted). 98. Sadarangani GP, Jiang X, Simpson LA, Eng JJ, Menon C. Force Myography for
75. Rainoldi A, Melchiorri G, Caruso I. A method for positioning electrodes Monitoring Grasping in Individuals with Stroke with Mild to Moderate Upper-
during surface EMG recordings in lower limb muscles. J Neurosci Methods. Extremity Impairments: A Preliminary Investigation in a Controlled
2004;134(1):37–43. Environment. Front Bioeng Biotechnol. 2017;5:42 [cited 2017 Aug 18]; Available
76. Hermens HJ, Freriks B, Disselhorst-Klug C, Rau G. Development of from: http://journal.frontiersin.org/article/10.3389/fbioe.2017.00042/full.
recommendations for SEMG sensors and sensor placement procedures. J 99. Ahsan MR, Ibrahimy MI, Khalifa OO. Others. {EMG} signal classification for
Electromyogr Kinesiol. 2000;810(1):333–49. human computer interaction: a review. Eur J Sci Res. 2009;33(3):480–501.
77. De Luca CJ, Donald Gilmore L, Kuznetsov M, Roy SH. Filtering the surface EMG 100. Shull PB, Jiang S, Zhu Y, Zhu X. Hand Gesture Recognition and Finger Angle
signal: movement artifact and baseline noise contamination. J Biomech. 2010; Estimation via Wrist-Worn Modified Barometric Pressure Sensing. IEEE Trans
43(8):1573–9. Available from:. https://doi.org/10.1016/j.jbiomech.2010.01.027. Neural Syst Rehabil Eng. 2019;PP(c):1 Available from: https://ieeexplore.ieee.
78. Atzori M, Müller H. Electromyography low pass filtering effects on the org/document/8668695/.
classification of hand movements in amputated subjects. SCIEI Int Conf 101. Atzori M, Gijsberts A, Castellini C, Caputo B, Hager AGM, Elsig S, et al.
Digit Signal Process. 2015;3(2):19–21. Electromyography data for non-invasive naturally-controlled robotic hand
79. Posic. Encoders. [cited 2019 Sep 27]. Available from: https://www.posic.com/ prostheses. Sci Data. 2014;1:1–13.
EN/products.html 102. Thalmic Labs. MYO. 2016 [cited 2018 May 28]. Available from: https://www.
80. Brokaw EB, Holley RJ, Lum PS. Hand Spring Operated Movement Enhancer myo.com/
(HandSOME) Device for Hand Rehabilitation after Stroke. 2010 Annu Int 103. Masson S, Fortuna FS, Moura FS, Soriano DC. Integrating Myo Armband for
Conf Ieee Eng Med Biol Soc. 2010;19(4):5867–70. the Control of Myoelectric Upper Limb Prosthesis. XXV Congr Bras Eng
81. Chen T, Lum PS. Hand rehabilitation after stroke using a wearable, high Biomédica - CBEB. 2016;2016(October):1–4.
DOF, spring powered exoskeleton. Proc Annu Int Conf IEEE Eng Med Biol 104. Mendez I, Hansen BW, Grabow CM, Smedegaard EJL, Skogberg NB, Uth XJ,
Soc EMBS. 2016;2016:578–81. et al. Evaluation of the Myo armband for the classification of hand motions.
82. Chen J, Lum PS. Spring operated wearable enhancer for arm rehabilitation IEEE Int Conf Rehabil Robot. 2017:1211–4.
(SpringWear) after stroke. Proc Annu Int Conf IEEE Eng Med Biol Soc EMBS. 105. CTRL-Labs. Neural Interface Platform. 2019. [cited 2019 Oct 5]. Available
2016;2016:4893–6. from: https://www.ctrl-labs.com/
83. Lim CK, Chen IM, Luo Z, Yeo SH. A low cost wearable wireless sensing 106. OYMotion. OYM. 2019 [cited 2019 May 13]. Available from: http://oymotion.
system for upper limb home rehabilitation. 2010 IEEE Conf robot autom com/site/
mechatronics. RAM. 2010;2010:1–8. 107. Beange KHE, Chan ADC, Graham RB. Evaluation of wearable IMU
84. Noitom. Perception Neuron. 2018 [cited 2019 Jul 6]. Available from: https:// performance for orientation estimation and motion tracking. MeMeA 2018.
www.noitom.com/solutions/perception-neuron 2018 IEEE Int Symp Med Meas Appl Proc. 2018;3528725544:1–6.
85. Lee S, Lee YS, Kim J. Automated evaluation of upper-limb motor function 108. Delsys. Wearable sensors for movement sciences. 2019 [cited 2019 Jan 24].
impairment using Fugl-Meyer assessment. IEEE Trans Neural Syst Rehabil Available from: https://www.delsys.com/
Eng. 2018;26(1):125–34. 109. Cometa. Cometa Systems. 2015 [cited 2019 Jan 30]. Available from: https://
86. Tognetti A, Lorussi F, Bartalesi R, Quaglini S, Tesconi M, Zupone G, et al. www.cometasystems.com/
Wearable kinesthetic system for capturing and classifying upper limb 110. Ltd B. Biometrics. 2015 [cited 2019 May 13]. Available from: http://www.
gesture in post-stroke rehabilitation. J Neuroeng Rehabil. 2005;B:8. biometricsltd.com/
87. Giorgino T, Lorussi F, De Rossi D, Quaglini S. Posture classification via 111. Yang G, Deng J, Pang G, Zhang H, Li J, Deng B, et al. An Iot-Enabled Stroke
wearable strain sensors for neurological rehabilitation. Annu Int Conf IEEE Rehabilitation System Based On Smart Wearable Armband And Machine
Eng Med Biol - Proc. 2006:6273–6. Learning. IEEE J Transl Eng Heal Med. 2018;6(May):1-10.
88. Giorgino T, Tormene P, Lorussi F, De Rossi D, Quaglini S. Sensor evaluation 112. Herrmann S, Buchenrieder K. Fusion of myoelectric and near-infrared signals
for wearable strain gauges in neurological rehabilitation. IEEE Trans Neural for prostheses control. Proc 4th Int Conv Rehabil Eng \& Assist Technol.
Syst Rehabil Eng. 2009;17(4):409–15. 2010:1–4.
113. Guo W, Yao P, Sheng X, Zhang D, Zhu X. An enhanced human-computer 138. Burridge JH, Lee ACW, Turk R, Stokes M, Whitall J, Vaidyanathan R, et al.
interface based on simultaneous sEMG and NIRS for prostheses control. Telehealth, wearable sensors, and the internet: will they improve stroke
2014 IEEE Int Conf Inf Autom ICIA. 2014;2014(July):204–7. outcomes through increased intensity of therapy, motivation, and
114. Paleari M, Luciani R, Ariano P. Towards NIRS-based hand movement adherence to rehabilitation programs? J Neurol Phys Ther. 2017;41:S32–8.
recognition. 2017. Int Conf Rehabil Robot. 2017:1506–11 Available from: 139. Olney SJ, Nymark J, Brouwer B, Culham E, Day A, Heard J, et al. A
http://ieeexplore.ieee.org/document/8009461/. randomized controlled trial of supervised versus unsupervised exercise
115. Sadri B, Goswami D, Martinez R. Rapid Fabrication of Epidermal Paper-Based programs for ambulatory stroke survivors. Stroke. 2006;37(2):476–81.
Electronic Devices Using Razor Printing. Micromachines. 2018;9(9):420 140. Microsoft. Skype. 2018 [cited 2018 Jul 17]. Available from: https://www.
Available from: http://www.mdpi.com/2072-666X/9/9/420. skype.com/en/home/
116. Bareket L, Inzelberg L, Rand D, David-Pur M, Rabinovich D, Brandes B, et al. 141. Maceira-Elvira P, Popa T, Schmid A-CHFC. Feasibility of home-based, self-
Temporary-tattoo for long-term high fidelity biopotential recordings. Sci applied transcranial direct current stimulation to enhance motor learning in
Rep. 2016;6(December 2015):1–8. middle-aged and older adults. Brain Stimul. 2019. In press.
117. Hocoma. ArmeoSenso. 2019. [cited 2019 Sep 29]. Available from: https:// 142. Van de Winckel A, Carey JR, Bisson TA, Hauschildt EC, Streib CD, Durfee WK.
www.hocoma.com/solutions/armeo-senso/ Home-based transcranial direct current stimulation plus tracking training
118. Cometa. EMG-Easy-Report. 2015 [cited 2019 Sep 29]. Available from: https:// therapy in people with stroke: an open-label feasibility study. J Neuroeng
www.cometasystems.com/products/emg-easy-report Rehabil. 2018;15(1):83 Available from: https://jneuroengrehab.biomedcentral.
119. Noraxon. myoMuscle. 2019 [cited 2019 Sep 29]. Available from: https:// com/articles/10.1186/s12984-018-0427-2.
www.noraxon.com/our-products/myomuscle/
120. Hughes AM, Burridge JH, Demain SH, Ellis-Hill C, Meagher C, Tedesco-Triccas
L, et al. Translation of evidence-based Assistive Technologies into stroke
Publisher’s Note
Springer Nature remains neutral with regard to jurisdictional claims in
rehabilitation: Users’ perceptions of the barriers and opportunities. BMC
published maps and institutional affiliations.
Health Serv Res. 2014;14(1):124.
121. WHO. Towards a common language for functioning, disability and health:
ICF. Int Classif. 2002;1149:1–22 Available from: http://www.who.int/
classifications/icf/training/icfbeginnersguide.pdf.
122. Metcalf C, Adams J, Burridge J, Yule V, Chappell P. A review of clinical upper
limb assessments within the framework of the WHO ICF. Musculoskeletal
Care. 2007;6(October 2007):150–4.
123. Escorpizo R, Stucki G, Cieza A, Davis K, Stumbo T, Riddle DL. Creating an
Interface between the international classification of functioning, disability
and health and physical therapist practice. Phys Ther. 2010;90(7):1053–63.
124. Geyh S, Cieza A, Schouten J, Dickson H, Frommelt P, Omar Z, et al. ICF Core
Sets for stroke. J Rehabil Med Suppl. 2004;36(44):135–41.
125. De Baets L, van der Straaten R, Matheve T, Timmermans A. Shoulder
assessment according to the international classification of functioning
by means of inertial sensor technologies: a systematic review. Gait
Posture. 2017;57(June):278–94. Available from:. https://doi.org/10.1016/j.
gaitpost.2017.06.025.
126. Wilkins E, Wilson L, Wickramasinghe K, Bhatnagar P, Leal J, Luengo-
Fernandez R, et al. European Cardiovascular Disease Statistics 2017 edition.
Eur Hear Network, Brussels. 2017;192:180. Available from: http://www.
ehnheart.org/images/CVD-statistics-report-August-2017.pdf.
127. Johnson BH, Bonafede MM, Watson C. Short- and longer-term health-care
resource utilization and costs associated with acute ischemic stroke. Clin
Outcomes Res. 2016;8:53–61.
128. Moon L, Moise P, Jacobzone S. Stroke care in OECD countries: a comparison
of treatment, costs and outcomes in 17 countries. Labour. 2003:74.
129. Government USF. Medicare. [cited 2019 Oct 9]. Available from: https://www.
medicare.gov/index
130. Conroy BE, Dejong G, Horn SD. Hospital-based stroke rehabilitation in the
United States. Top Stroke Rehabil. 2009;16(1):34–43.
131. Mahler MP, Züger K, Kaspar K, Haefeli A, Jenni W, Leniger T, et al. A cost
analysis of the first year after stroke - early triage and inpatient rehabilitation
may reduce long term costs. Swiss Med Wkly. 2008;138(31–32):459–65.
132. Renjilian DA, Perri MG, Nezu AM, McKelvey WF, Shermer RL, Anton SD.
Individual versus group therapy for obesity: effects of matching participants
to their treatment preferences. J Consult Clin Psychol. 2001;69(4):717–21.
133. Walker MF, Sunnerhagen KS, Fisher RJ. Evidence-based community stroke
rehabilitation. Stroke. 2013;44(1):293–7.
134. von Koch L, Wottrich AW, Holmqvist LW. Rehabilitation in the home versus
the hospital: the importance of context. Disabil Rehabil. 1998;20(10):367–72
Available from: http://www.tandfonline.com/doi/full/10.3109/096382
89809166095.
135. Grossman R, Salas E. The transfer of training: what really matters. Int J Train
Dev. 2011;15(2):103–20.
136. Langan J, DeLave K, Phillips L, Pangilinan P, Brown SH. Home-based
telerehabilitation shows improved upper limb function in adults with
chronic stroke: a pilot study. J Rehabil Med. 2013;45(2):217–20.
137. Dobkin BH. A rehabilitation-internet-of-things in the home to augment
motor skills and exercise training. Neurorehabil Neural Repair. 2017;31(3):
217–27.

Jurnal 1 C

Uploaded by

Copyright:

Available Formats

Jurnal 1 C

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Jurnal 1 C

Uploaded by

Copyright:

Available Formats

Maceira-Elvira et al.

Journal of NeuroEngineering and Rehabilitation (2019) 16:142

REVIEW Open Access

Wearable technology in stroke

Introduction differently, the rehabilitative process after stroke de-

In addition to thorough standardized assessment Wearable devices used in stroke patients

interesting possibilities have been explored in every-day-

Implementation (requirements and challenges)

in the healthy person.

“GTX+” (ActiGraph®, Pensacola,

during motor assessments, is often approached by extract-

Ethics approval and consent to participate

changes. Conventional assessments of motor function

You might also like