Customized Linear Discriminant Analysis for Brain-Computer Interfaces

This study presents a procedure to customize mental task discrimination for a specific human subject. Three male subjects, between 20 and 30 years old, were submitted to 4-5 sessions. Each session was composed of 4 blocks of 20 trials. Two block types were implemented. One required that the subject perform feet and tongue movements. The other block required the subject to perform left and right arm movements. Subjects were instructed to perform motor imagery as well as actual movements. In order to avoid previous assumptions on preferable channel locations and frequency ranges, 105 (21 electrodestimes5 frequency ranges) electroencephalogram (EEG) features were extracted from the sessions' data. A linear discriminant analysis (LDA) approach was applied to the feature set. The dimensionality of the multivariate data set was reduced through a discriminant stepwise procedure. Only the variables which best discriminated between groups, for a specific subject, were used. Those features predicted group membership during online feedback sessions with error lower than 12%, in each subject best performance. Classification errors for training data were very low and were neglected