Seperability of four-class motor imagery data using independent components analysis

This paper compares different ICA preprocessing algorithms on cross-validated training data as well as on unseen test data. The EEG data were recorded from 22 electrodes placed over the whole scalp during motor imagery tasks consisting of four different classes, namely the imagination of right hand, left hand, foot and tongue movements. Two sessions on different days were recorded for eight subjects. Three different independent components analysis (ICA) algorithms (Infomax, FastICA and SOBI) were studied and compared to common spatial patterns (CSP), Laplacian derivations and standard bipolar derivations, which are other well-known preprocessing methods. Among the ICA algorithms, the best performance was achieved by Infomax when using all 22 components as well as for the selected 6 components. However, the performance of Laplacian derivations was comparable with Infomax for both cross-validated and unseen data. The overall best four-class classification accuracies (between 33% and 84%) were obtained with CSP. For the cross-validated training data, CSP performed slightly better than Infomax, whereas for unseen test data, CSP yielded significantly better classification results than Infomax in one of the sessions.