Feature Extraction for Multi-class BCI using Canonical Variates Analysis

To propose a new feature extraction method with canonical solution for multi-class brain-computer interfaces (BCI). The proposed method should provide a reduced number of canonical discriminant spatial patterns (CDSP) and rank the channels sorted by power discriminability (DP) between classes. The feature extractor relays in canonical variates analysis (CVA) which provides the CDSP between the classes. The number of CDSP is equal to the number of classes minus one. We analyze EEG data recorded with 64 electrodes from 4 subjects recorded in 20 sessions. They were asked to execute twice in each session three different mental tasks (left hand imagination movement, rest, and words association) during 7 seconds. A ranking of electrodes sorted by power discriminability between classes and the CDSP were computed. After splitting data in training and test sets, we compared the classification accuracy achieved by linear discriminant analysis (LDA) in frequency and temporal domains. The average LDA classification accuracies over the four subjects using CVA on both domains are equivalent (57.89% in frequency domain and 59.43% in temporal domain). These results, in terms of classification accuracies, are also reflected in the similarity between the ranking of relevant channels in both domains. CVA is a simple feature extractor with canonical solution useful for multi-class BCI applications that can work on temporal or frequency domain.