An automated method for relevant frequency bands identification based on genetic algorithms and dedicated to the Motor Imagery BCI protocol

This paper presents an automated method for relevant frequency bands identification to be used in a left/right hand motor imagery based Brain Computer Interface system. The adopted optimization method aimed at maximizing the ratio between the mutual information and the error rate obtained using a Regularized Linear Discriminant Analysis (RLDA) based classifier and band-specific amplitude modulated envelopes as features. The search problem was handled by a genetic algorithm starting from an initial population determined on the basis of a-priori mu and beta relevant frequency bands identified by means of a standard power spectral density analysis between the idle and the left/right imagery data subset.