• Citer

Machine Learning for Electrophysiological Decoding: From Single-Trial EEG State Classification to EMG-Based Hand Kinematics Estimation

par Simar, Cédric
Président du jury Lenaerts, Tom
Promoteur Bontempi, Gianluca
Co-Promoteur Chéron, Guy
Publication Non publié, 2026-09-08

Thèse de doctorat

• Accès en ligne
• Détails
• Contenu
• Statistiques

Résumé :

Electrophysiological signals provide measurable windows onto the neural and muscularprocesses underlying perception, intention, and movement. Electroencephalography(EEG) records brain electrical activity at the scalp, whereas surface electromyography(sEMG) captures activity associated with muscle recruitment. Traditional analysesreveal robust population-level patterns, but averaging can conceal information specificto a person or moment. Machine learning models can be trained to identify informativepatterns within these trial-to-trial and inter-individual differences, enabling predictionsfrom individual recordings rather than only from population averages. Yet, predictiveaccuracy alone does not guarantee physiological plausibility or generalization. Thisthesis investigates how physiological knowledge and machine learning can be combinedto decode EEG and sEMG for scientific analysis and neurotechnology.The thesis progresses from single-trial EEG classification to interpretable individualpredictions, and from discrete EMG gesture recognition to continuous finger-kinematicsestimation. Covariance matrices, which describe how signals vary together, and Rieman-nian geometry provide a shared mathematical framework across these distinct biosignals.In an interactive hyperscanning paradigm, Riemannian covariance pipelines discrimi-nated neutral, festive, and violent conditions from preparatory EEG activity in actorsand observers. Riemannian covariance pipelines were then applied to visual responsestraditionally studied through grand averages. Single-trial responses to checkerboard andthree-dimensional navigation images could be distinguished across participants, whilescalp- and source-level analyses offered complementary perspectives on generalizationand individual interpretation. A local saliency method was subsequently developedto map classifier evidence back to sensor-covariance coordinates. The resulting mapshighlighted posterior scalp patterns consistent with visual processing and supportedthe physiological assessment of individual predictions. This method could be used as anonline quality-control step for BCI and EEG decoding systems.The EMG part examines how forearm muscle activity can be transformed into rep-resentations of hand movement. A virtual-reality paradigm enabled the synchronizedacquisition of sEMG and hand poses during guided bimanual gestures. The resultsshowed that sustained tonic activity recorded while holding a posture was as informa-tive for gesture recognition as the phasic activity produced during movement. Also,Riemannian covariance pipelines combined with a physiologically informed decompo-sition into move and hold components improved gesture recognition. The final partof the thesis moves beyond predefined gestures toward continuous, high-dimensionalfinger-motion estimation. It evaluates a lightweight recurrent regression model thatuses sequences of multi-band covariance features to simultaneously estimate multiplefinger joint angles from recent sEMG activity. The model achieved lower average errorsthan the evaluated baselines on two datasets, was compatible with embedded inference,and enabled real-time control of a physical robotic hand.Across these contributions, physiological expertise and predictive modeling emergeas complementary: domain knowledge guides model construction and interpretation,while machine learning tests whether electrophysiological patterns remain informativefor individual observations and unseen participants. Overall, this thesis advanceselectrophysiological decoding by moving, on the one hand, from averaged descriptionsto interpretable individual predictions and, on the other, from simple gesture recognitionto continuous, real-time estimation of complex finger kinematics. It identifies adaptivelearning, time-resolved interpretation, and realistic validation as priorities for robustscientific and assistive applications.