Feature Selection for Brain-Computer Interface

Abstract

The implementation of a realistic Brain-Computer Interface (BCI) for non-trained subjects requires a classifier able to detect subject-specific electroencephalogram (EEG) patterns. The classifier should be simple enough (i.e. few EEG features, few EEG channels) to enable the mutual adaptation between the BCI and the user. However the classifier should still capture subject-specific EEG patterns. Hence, a feature selection algorithm is introduced as a new method to detect subject-dependent feature and channel relevance for mental task discrimination. This algorithm employs a new formulation of principal component analysis that accommodates the group structure in the dataset. Five subjects were submitted to an EEG experiment that instructed them to perform movement imagery tasks. A linear discriminant classifier used the selected features to discriminate EEG responses to left vs. right hand movement imagery performance with 18% cross-validation error. The discrimination of EEG responses to tongue vs. feet movement imagery performance achieved 26% cross-validation error. Besides the low classification error attained, the selected features often included the frequency-space EEG patterns generally suggested for movement imagery task discrimination. Moreover, each subject’s characteristic EEG patterns were also detected and enhanced classifier performance and customization.