Supervised Relevance Analysis for Multiple Stein Kernels for Spatio-Spectral Component Selection in BCI Discrimination Tasks

Brain-Computer Interfaces (BCI) allow direct communication from the human brain to machines by analyzing sensorimotor activity. Within the most common paradigms of cognitive neuroscience for BCI, the Motor Imagery one relies on imaging movement of hands, feet, and tongue, usually. Standard BCI systems make use of electroencephalographic signals due to its high temporal resolution, portability, and easiness to implement. Recently, the filter-banked analysis of common spatial patterns has proved to be among the most discriminative approaches. Nonetheless, such an approach results in high dimensional spaces yielding to overtrained systems, or ill-posed spatial covariance matrices, when the number of EEG sensors increases. In this work, we propose a spatio-spectral relevance analysis, termed Multiple Stein Kernels (MSK), to simultaneously select channels and bands in a supervised criterion, so reducing the spatial and frequency dimension. We test our approach in three known BCI datasets with a varying number of subjects and channels. The attained classification results prove that MKS not only improves the classification accuracy and reduces the dimension of the representation space, but also select frequency bands that are physiologically interpretable.