Time-Constrained Filter Bank Common Spatial Pattern for Motor Imagery Brain-Computer Interfaces

One of most important tasks or key steps in the designing of an EEG-based BCI system is the optimization of spatio-temporal filters for each subject due to the poor spatial resolution of the EEG recordings, as well as the topographical arrangement and frequency specificity of brain activities. A highly popular technique for the optimization of spatial filters is Common Spatial Pattern (CSP). To address the problem of selecting spectral filtering bands, Filter Bank Common Spatial Pattern (FBCSP) was recently proposed, which improves the performance of CSP significantly. This paper provides a deep insight to the CSP algorithm and proposes a novel method termed Time-Constrained Filter Bank Common Spatial Pattern (TFBCSP). TFBCSP eludes the problem of selecting subjectdependent frequency bands by adopting the filter bank approach as in FBCSP and exploits the short-duration nature of ERD/ERS by imposing a time constraint on the EEG samples whose variance is maximized/minimized. Favorable results were obtained by the proposed method on dataset IVa from BCI Competition III.