Combined Common Spatial Pattern and spectral filtering for EEG-based BCIs

Signal processing for EEG-based BCIs is complicated by spatial blurring effects between electrodes due to complex volume conduction characteristics of the brain, skull, and scalp. Although spatial filters are increasingly being used to mitigate these effects, results have been limited. In this work, we investigate the use of the Common Spatial Pattern (CSP) filtering technique to spatially filter signals from a 64-channel offline EEG data set comprising two classes of imagined movements. Despite CSP methods having already been demonstrated in BCI algorithms, there has been no study of how this technique can be used as a de-noising stage prior to feature extraction and classification. We propose the use of spectral analysis subsequent to CSP-based spatial filtering and demonstrate that this signal processing combination provides superior class separation to existing methods. Our results demonstrate that CSP-filtered signals exhibit strong separation between signal classes at frequency peaks at 4, 12, and 20Hz and at various electrode locations across the scalp.