Ph.D.

research topic
• Title of the proposed topic: “Novel tensor decompositions for atrial fibrillation
analysis”
• Research axis of the 3IA: IA for integrative computational medicine
• Supervisor (name, affiliation, email): Vicente Zarzoso, I3S Laboratory, UCA, CNRS
[email protected]
• Potential co-supervisor (name, affiliation): Rodrigo Cabral Farias, I3S, UCA, CNRS
• The laboratory and/or research group: Signals, Images, Systems (SIS) Team, I3S Lab.

Apply by sending an email directly to the supervisor.

The application will include :
• Letter of recommendation of the supervisor indicated above
• Curriculum vitæ.
• Motivation Letter.
• Academic transcripts of a master’s degree(s) or equivalent.
• At least, one letter of recommendation.
• Internship report, if possible.

• Description of the topic:

Background
Atrial fibrillation (AF) is the most common sustained arrhythmia encountered in clinical practice,
especially affecting the elderly and held responsible of up to 25% of strokes. With the ageing of the
Western population, this arrhythmia is becoming a major public health concern taking epidemic
proportions: between 6 and 12 million people will be affected in the USA in 2050, and up to 18 million
in Europe in 2060 [MOR17]. AF is indeed “the last great frontier of cardiac electrophysiology” as it
continues to puzzle cardiologists [JAN14]. Physiological signal analysis and machine learning arise a key
tools to improve the understanding and management of this challenging cardiac condition. Despite its
cost and risk of complications, catheter ablation is currently the most attractive therapeutic option in
terms of long-term recurrence rate for the treatment of persistent AF. However, this therapy depends
heavily on the practitioner’s subjectivity, with rather variable protocols and success rates reported by
different centers. The development of robust, widely accepted intervention protocols is still an open
challenge.

The noninvasive analysis of AF can be carried out by processing cardiac signals recorded by electrodes
located on the patient’s skin, giving rise to the surface electrocardiogram (ECG), a low-cost noninvasive
clinical tool in cardiology. Recent research has shown the benefits of arranging ECG data in the form
of three-dimensional arrays or tensors, which are then decomposed using suitable tensor
 3IA Côte d’Azur PhD thesis topic “Novel tensor decompositions for AF analysis”

factorizations. This approach has proven particularly useful for atrial activity extraction from the
surface ECG [MAR19, MAR20, MAR21, GOU20]. Nevertheless, the clinical power of tensor tools and
their usefulness to derive markers of AF characterization and aid in the therapeutic gesture remain
mostly unexplored.

Goals
Taking a step forward towards a patient-tailored management of persistent AF, the present thesis aims
at developing new tensor data analysis techniques to aid in the characterization of this cardiac
disturbance, particularly in the context of the catheter ablation therapy. Emphasis will be laid on the
recent ablation protocol developed at Nice Pasteur University Hospital (CHU) [SEI17] based on multi-
spline catheters. The thesis will consist of three axes:

• Novel quantitative measures of AF complexity: Despite the growing interest in noninvasive

methodologies to assess the complexity of persistent AF, their use in the characterization of
the ablation therapy is still rather limited [MEO18]. To fill this gap, this part of the project aims
to explore the power of tensor techniques to derive novel noninvasive measures of AF
complexity that may accurately reflect the global evolution of the atrial substrate during
catheter ablation procedures. The local complexity of intracardiac atrial electrograms
recorded by multi-spline catheters will also be analyzed. The tensor model structure estimated
by suitable optimization algorithms, such as that recently developed in [GOU20], will be
considered as a starting point.

• Exploiting data multimodality and multiple tensor modeling: Several data modalities, including
invasive (intracardiac electrograms, electroanatomic maps) and noninvasive (ECG,
echocardiography), are acquired in the management of persistent AF patients. This part of the
project will study how to exploit such modalities simultaneously in a robust manner by means
of coupled tensor factorizations. Encouraging preliminary results have recently been obtained
to deal with multiple time segments of a single data modality for blind separation of
nonstationary data (AF ECG) [MAR20]. Extensions to multimodal observations will be
considered in the thesis. Moreover, even in the case of a single modality, signals related to
different cardiac activities may be best modeled by using different constructions of the tensor
from ECG data [GOU20, MAR21]. One of the challenges to be solved in this thesis will be to
propose methods that analyze the different tensors simultaneously. These methods are
expected to result in improved extraction of atrial activity.

• Dealing with high dimensionality: Computational difficulties linked to the high dimensionality
of the underlying tensor approach have been tackled in [GOU20] by undersampling the
measured AF ECG signals. Although this approach has been successful, other approaches that
do not discard part of the samples can also be considered. For example, by using tensor
versions of principal component analysis [DEL00] to reduce the dimensionality of the data prior
to its analysis. An issue with such an approach is that the resulting tensor cannot be analyzed
with the same methods as in [GOU20]. One of the objectives of this thesis will be to propose
specialized tensor decomposition methods that can be applied for AF ECG signal analysis after
tensor dimensionality reduction. Extensions of the methods will also be considered in the
multimodal setting, where dimensionality reduction is applied jointly to the tensors arising
from different modalities.

The above goals will require the development of novel tensor decomposition models and
algorithms in the challenging scenarios of non-stationary, multimodal, noisy observations with
possibly missing data, often encountered in AF ablation procedures. Recent results [MAR20,
MAR21] show the performance gains brought forth by adapted tensor decompositions in this

2
 3IA Côte d’Azur PhD thesis topic “Novel tensor decompositions for AF analysis”

context. The thesis will deepen this fruitful line of research. The impact of these theoretical results
will likely extend beyond the medical application motivating this work.

Collaborations
This interdisciplinary thesis will be carried out in close collaboration with cardiologists from Nice CHU
and Monaco Princess Grace Hospital (CHPG), and could attract industrial partners.

Pre-requisites
Prospective candidates must hold an MSc diploma or equivalent in data science, signal processing or
applied mathematics and will require a strong background in theoretical as well as computational
aspects of linear algebra, optimization and signal processing. Previous experience with biomedical
signals and familiarity with cardiac electrophysiology will also be interesting assets.

Bibliography

[DEL00] L. De Lathauwer, B. De Moor, J. Vandewalle, “A multilinear singular value decomposition”, SIAM journal
on Matrix Analysis and Applications, vol. 21, no. 4, 2000, pp. 1253-1278.

[GOU20] J. H. M. Goulart, P. Marinho, R. C. Farias, V. Zarzoso and P. Comon, “Alternating group lasso for block-
term tensor decomposition with application to ECG source separation”, IEEE Transactions on Signal Processing,
vol. 68, 2020, pp. 2682-2696.

[JAN14] C. T. January, L. S. Wann, J. S. Alpert, et al., “2014 AHA/ACC/HRS guideline for the management of
patients with atrial fibrillation: A report of the American College of Cardiology/American Heart Association Task
Force on practice guidelines and the Heart Rhythm Society”, Circulation, vol. 64, no. 21, pp. 2246-2280, Dec.
2014.

[MAR19] P. Marinho and V. Zarzoso, "Block term decomposition of ECG recordings for atrial fibrillation analysis:
temporal and inter-patient variability", Journal of Communication and Information Systems, vol. 34, no. 1, 2019,
pp. 111-110.

[MAR20] P. Marinho, V. Zarzoso, and C. A. R. Fernandes, "Coupled tensor model of atrial fibrillation ECG", in:
Proc. EUSIPCO- 2020, 28th European Signal Processing Conference, Amsterdam, The Netherlands, Jan. 18-21,
2021.

[MAR21] P. M. R. de Oliveira, J. H. de M. Goulart, C. A. R. Fernandes and V. Zarzoso, "Blind Source Separation in

Persistent Atrial Fibrillation Electrocardiograms Using Block-Term Tensor Decomposition with Löwner
Constraints", IEEE Journal of Biomedical and Health Informatics, in press.

[MEO18] M. Meo, T. Pambrun, N. Derval, C. Dumas-Pomier, S. Puyo et al., “Noninvasive assessment of atrial
fibrillation complexity in relation to ablation characteristics and outcome”, Frontiers in Physiology, vol. 9, Jul.
2018.

[MOR17] C. A. Morillo, A. Banerjee, P. Perel, D. Wood, and X. Jouven, “Atrial fibrillation: the current epidemic”,
Journal of Geriatric Cardiology, vol. 14, no. 3, pp. 195–203, Mar. 2017.

[SEI17] J. Seitz, C. Bars, G. Théodore, S. Beurtheret, N. Lellouche, et al., "AF ablation guided by spatiotemporal
electrogram dispersion without pulmonary vein isolation. A wholly patient-tailored approach", Journal of the
American College of Cardiology, vol. 69, no. 3, pp. 303–321, Jan. 2017.