Composite kernel support vector machine based performance enhancement of brain computer interface in conjunction with spatial filter

For Motor imagery Brain Computer interface, a large number of electrodes are placed on the scalp to acquire EEG signals. However, the available number of samples from a subject's EEG is very less. In such a situation, learning models which use spatial features obtained using common spatial pattern (CSP) method suffer from overfitting and leads to degradation in performance. In this paper, we propose a novel three phase method CKSCSP which automatically determines a minimal set of relevant electrodes along with their spatial location to achieve enhanced performance to distinguish motor imagery tasks for a given subject. In the first phase, electrodes placed on brain scalp are divided among five major regions (lobes) viz. frontal, central, temporal, parietal and occipital based on anatomy of brain. In the second phase, stationary-CSP is used to extract features from each region separately. Stationary-CSP will handle the non-stationarity of EEG. In the third phase, recursive feature elimination in conjunction with composite kernel support vector machine is used to rank brain regions according to their relevance to distinguish two motor-imagery tasks. Experimental results on publically available datasets demonstrate superior performance of the proposed method in comparison to CSP and stationary CSP. Also, Friedman statistical test demonstrates that the proposed method CKSCSP (μ≠0) outperforms existing methods.