Abstract (EN)

Electroencephalography (EEG) is the most common non-invasive technique for measuring brain activity. It uses electrodes to capture the electric fields with high temporal resolution. However, analysing this data with machine learning presents several challenges, the most prominent one being the high inter/intra-subject variability (i.e. BCI inefficiency). One solution to tackle this challenge is to explore alternative features that capture additional types of information and help to better discriminate the subjects' mental state. For instance, complementary features that reflect interactions between brain areas, rather than relying solely on local measurements as proven efficient [1]. Functional Connectivity (FC) studies the interaction between pairs of electrodes in the spectral domain, embedding information in robust but distributed structures. To uncover those structures, graph-based approaches, such as Graph Neural Networks (GNN), demonstrated their ability to analyse complex networks, such as the interconnected nature of brain signals [2]. This study proposes a GNN classifier method that combines FC information from various estimators like the Riemannian Covariance, Instantaneous and Imaginary Coherence with traditional feature space such as raw EEG features and spectral phase for each electrode for Brain-Computer Interfaces decoding. Preliminary results obtained with the High Gamma Dataset [3] show that the algorithm could successfully classify MI movements, but further experiments are necessary to evaluate the robustness of the intra-subject and cross-subject methods.