Independent component analysis in a motor imagery brain computer interface

Abstract

There are a lot of scientific papers reporting a significant increase in classification accuracy after applying independent component analysis (ICA) for cleaning EEG data. Most of them, however, are focused on multidimensional data, recorded from a dense matrix of electrodes. When there are enough EEG channels, the benefits of ICA are straightforward — some of the components returned by ICA algorithm reflect artifacts disturbing the true brain activity and it is enough to detect and remove them to improve the signal quality. The question is what to do when data are recorded only from a few-channel EEG. The paper presents the results of the experiment that was performed in order to test our strategy for applying ICA for a 4-channel EEG data recorded for motor imagery brain computer interface. Five subjects, untrained in motor imagery paradigm took part in the experiment. According to our results the mean classification accuracy increased after applying ICA from 67% to 76% (for 10-second time window) and from 66% to 77% (for reduced 7-second time window).