A combination of spatial and spectral filters for mental condition discrimination

It is widely accepted that the common spatial pattern (CSP) analysis method, albeit being very popular in brain-computer interface (BCI) applications as a feature extraction method for binary classification, is vulnerable to artifact. It could underperform when it is exposed to an input whose frequency band is too broad that many interfering frequency components are contained. These drawbacks are closely related to the nature of CSP filters which are based on completely spatial weighting. That is, CSP has no control on the temporal space of brain signals. This work is one attempt to extend CSP by eliminating the undesirable temporal components through spectral filtering. The proposed method in this work retains the simplicity of CSP but derives a number of complex spatial and spectral integrated filters by applying multiple time lags and a regularization term. These filters are data-driven and channel-specific. Their ability to narrow the frequency band of signals so as to enhance feature extraction is demonstrated using a public available dataset, where 4.7% higher mean classification accuracy is achieved.