Bhattacharyya distance and confidence map based feature selection for Common spatial patterns algorithms in brain computer interface

A novel feature selection methodology based on Bhattacharyya distance and confidence map is presented and illustrated with electroencephalogram (EEG) signal classification problem. Although Common spatial pattern (CSP) is a mostly used algorithm for classification of EEG in brain-computer interface (BCI), which has poor frequency selectivity. To address this problem, a constant-bandwidth Butterworth filters bank was utilized for frequency decomposition. Then, our novel feature selection methodology was used for new CSP features ranking and selection. We compare our method with the existing approaches, the results on 4 subjects showed that the new method outperforms the other two existing approaches based on conventional CSP and Common Spatio-Spectral Pattern (CSSP), the proposed algorithm yielded lowest test error rate of 2.2±1.8% in subject 1 and will be a up-and-coming signal processing tool for developing BCI and improving the efficiency of the classification method in low-resolution EEG input and small-dataset conditions.