Smooth Spatial Filter for Common Spatial Patterns

Common Spatial Patterns CSPs is a popular feature extraction algorithm for Brain-Computer Interface BCI. However, the standard CSP spatial filters completely ignore the spatial information of EEG electrodes. To solve this problem, two smooth Regularized CSP RCSP algorithms are proposed in this paper, which are Spatially RCSP with a Gaussian Prior GSRCSP and Spatially RCSP with a Feature-Associations Modeling Matrix MSRCSP respectively. Then these algorithms are compared with the standard CSP and Spatially RCSP SRCSP, an existing smooth CSP, in an experiment on EEG data from three publicly available data sets from BCI competition. Results show that GSRCSP outperforms other algorithms in classification accuracy and MSRCSP needs least training time. Besides, the spatial filters obtained by GSRCSP and MSRCSP are smoother than the standard CSP and SRCSP and are more interpretable neuro-physiologically.