Extraction of spatially sparse common spatio-spectral filters with recursive weight elimination

The common spatial pattern(CSP) technique linearly combines the channels to filter the neural signal spatially. It operates on data that is band-pass filtered between particular cutoff frequency for all subjects and channels. On the other hand the common spatio-spectral pattern (CSSP) method extends the the traditional CSP technique that combines spectral filtering with the original spatial filtering by using the temporally delayed version of the original data. All recording channels and the delayed versions are combined when extracting the variance as input features for a brain machine interface. This linear combination increases the number of channels extensively and results in overfitting and robustness problems of the constructed system in presence of low number of training trials. To overcome this problem, we proposed spatially sparse CSSP method in which only a subset of all available channels and its temporally shifted versions are linearly combined when extracting the features. We utilized three different versions of the recursive weight elimination (RWE) technique to select a subset of electrodes for spatio-spectral projections. We evaluate the performance of the proposed method to distinguish between the movements of the first three fingers of the hand using electrocorticogram (ECoG) signals of the brain computer interface competition 2005. We observed that spatially sparse CSSP filter outperforms both original CSP and CSSP filter and results in improved generalization in classification.