A graph-theoretic sensor-selection scheme for covariance-based Motor Imagery (MI) decoding

Abstract

Optimal sensor selection is an issue of paramount importance in brain decoding. When associated with estimates of covariance, its implications concern not only classification accuracy, but also computational efficiency. However, very few attempts have been made so far, since it constitutes a challenging mathematical problem. Herein, we propose an efficient heuristic scheme that combines discriminative learning (from a small training dataset of labelled trials) with unsupervised learning (the automated detection of sensors that collectively maximize the trial discriminability of the induced Covariance structure). The approach is motivated from a complex network modelling perspective. Its efficacy and efficiency are demonstrated experimentally, based on BCI-competition datasets concerning MI-tasks, and compared against popular techniques in the field.