Wavelet coherence based channel selection for classifying single trial motor imagery

Multi-channel electroencephalography (EEG) recordings require excessive computation and sometimes engender outliers, which make brain computer interface (BCI) systems inefficient. Thus, optimal channel selection becomes a key factor for developing a more comfortable BCI. This study emphasized on a time-frequency (T-F) coherence method, called as Wavelet Coherence (WC), for selecting lesser number of channels. The selected sets of channels were then used to classify two motor imagery (MI) tasks, i.e., right hand (RH) and right foot (RF). The data was collected from publicly available dataset IVa from BCI Competition III. Common spatial pattern (CSP) with and without regularization were applied as preprocessing techniques. While the classification accuracy is 90% using available 118 channels for subject ay, we have achieved higher classification accuracy of 93% using only 24 channels using CSP with regularization. Interestingly, the achieved classification accuracy for subject av is 67% using 4 channels only, that outperform the classification accuracy (i.e., 61%) achieved using 118 channels.