MANET

Brain-Computer Interfaces (BCIs) consist of acquiring, analysing, and translating brain signals into commands for control or communication. These systems present many clinical applications including control of wheelchairs, brain-based communication, and treatment of neurological disorders. Despite being beneficial for patients, controlling a BCI system is a learned skill that a non-negligible proportion of the users cannot develop even after training sessions. This phenomenon, called "BCI inefficiency" constitutes a strong limitation to the BCI diffusion. The “BCI inefficiency” concept is useful to understand why some users cannot interact well with BCI technologies. Nevertheless, it relies on the assumption that users are expected to reach an accuracy level within a finite time; and it infers that it is inherently on the user. Therefore, instead of considering the user as a cause of the observed variability, it appears essential to develop BCI systems that aim at considering the user's specificity. The underlying challenge here is to extract the most relevant information to discriminate properly the user's mental state, referred as “features”. My goal is to develop methods to assess and to improve BCI performance by considering the subjects' specificity. Aware that the BCI training relies on mutual-learning schemes between the users and the machines, I will propose a framework that tackles the "BCI inefficiency" phenomenon from both the neural decoder and the users’ sides. I first propose to enrich the decoded features by considering combination of multimodal and heterogeneous information to develop innovative classification tools (WP1). Then, I will generalize this framework by introducing new features based on the identification of reliable and subject-specific neurophysiological markers of BCI performance (WP 2). Finally, I will conduct an experimental validation through online BCI experiments (WP 3).