Soutenance de thèse: Igor CARRARA (INRIA)

Devant le jury composé de :
Marco CONGEDO, Directeur de Recherche, Gipsa-Lab - Grenoble Alpes University
Michael TANGERMANN, Associate Professor, Radboud University
Marie-Constance CORSI, Chargé de Recherche, NERV Lab - Paris Brain Institute
Sylvain CHEVALLIER, Professor, LISN - Univ. Paris-Saclay
Fabien LOTTE, Directeur de Recherche, LaBRI - Inria Centre at the University of Bordeaux
Théodore PAPADOPOULO, Directeur de Recherche, CRONOS - Centre Inria D’université Côte d’Azur

Summary :
Electroencephalography (EEG) non-invasively measures the brain’s electrical activity through electromagnetic fields generated by synchronized neuronal activity. This allows for the collection of multivariate time series data, capturing a trace of the brain electrical activity at the level of the scalp. At any given time instant, the measurements recorded by these sensors are linear combinations of the electrical activities from a set of underlying sources located in the cerebral cortex. These sources interact with one another according to a complex biophysical model, which remains poorly understood. In certain applications, such as surgical planning, it is crucial to accurately reconstruct these cortical electrical sources, a task known as solving the inverse problem of source reconstruction. While intellectually satisfying and potentially more precise, this approach requires the development and application of a subject-specific model, which is both expensive and technically demanding to achieve. However, it is often possible to directly use the EEG measurements at the level of the sensors and extract information about the brain activity. This significantly reduces the data analysis complexity compared to source-level approaches. These measurements can be used for a variety of applications, including monitoring cognitive states, diagnosing neurological conditions, and developing brain-computer interfaces (BCI). Actually, even though we do not have a complete understanding of brain signals, it is possible to generate direct communication between the brain and an external device using the BCI technology. This work is centered on EEG-based BCIs, which have several applications in various medical fields, like rehabilitation and communication for disabled individuals or in non-medical areas, including gaming and virtual reality.

Despite its vast potential, BCI technology has not yet seen widespread use outside of laboratories. The primary objective of this PhD research is to try to address some of the current limitations of the BCI-EEG technology. Autoregressive models, even though they are not completely justified by biology, offer a versatile framework to effectively analyze EEG measurements. By leveraging these models, it is possible to create algorithms that combine nonlinear systems theory with the Riemannian-based approach to classify brain activity.
The first contribution of this thesis is in this direction, with the creation of the Augmented Covariance Method (ACM). Building upon this foundation, the Block-Toeplitz Augmented Covariance Method (BT-ACM) represents a notable evolution, enhancing computational efficiency while maintaining its efficacy and versatility. Finally, the Phase-SPDNet work enables the integration of such methodologies into a Deep Learning approach that is particularly effective with a limited number of electrodes. Additionally, we proposed the creation of a pseudo online framework to better characterize the efficacy of BCI methods and the largest EEG-based BCI reproducibility study using the Mother of all BCI Benchmarks (MOABB) framework. This research seeks to promote greater reproducibility and trustworthiness in BCI studies.

In conclusion, we address two critical challenges in the field of EEG-based brain-computer interfaces (BCIs): enhancing performance through advanced algorithmic development at the sensor level and improving reproducibility within the BCI community.

Keywords : Brain Computer Interfaces, EEG/MEG data, Brain activity, Riemannian Geometry, Machine Learning, Deep Learning.