In Motor-Imagery based Brain-Computer Interfaces (MI-BCIs) MI tasks modulate EEG activity in the α and β frequency bands (8-30 Hz). Data driven methods are often used to select features in those bands during the system calibration, with little consideration for the resulting human performances. Indeed, the system is learning without taking into account any neurophysiological constraints or user related constraints.

One of these methods is the one that selects the frequency band that better discriminate which task the user is doing (B. Blankertz, 2007). Our study suggests that online MI-BCI performances correlate with the characteristics of the frequency band, selected using machine learning. This raises questions regarding the causality link direction: could we use this frequency band to predict online performances? Could we improve machine learning algorithms with constraints on the band to be selected? Is this correlation due to a covert confounding factor (e.g., mental strategy)?

We are currently running experiments using a modified version of the frequency band selection algorithm (i.e, a version with constraints) in order to better understand the relation between the algorithm and online BCI performances. In addition to the previous algorithm, we are also using common spatial pattern (CSP), a spatial method to extract features from the EEG signals. We are currently identifying characteristics of those CSP such as the number of foci or their position.

The objective of classification is to translate the extracted features into commands to best evaluate the user’s intent. Among the various classification algorithms that exist, we are using the linear discriminant analysis (LDA). We believe that it can also be interesting to extract characteristics of the LDA classifier to better understand BCI performances. One of those characteristics could be the sparsity of the LDA features weight. Finally, as BCIs are “communication” systems, we found it interesting to try to extract features from the feedback the user is given.