MATHEMATICAL ENGINEERING TECHNICAL REPORTS Spectrally weighted Common Spatial Pattern algorithm for single trial EEG classification

We propose a simultaneous spatio-temporal filter optimization algorithm for the single trial ElectroEncephaloGraphy (EEG) classification problem. The algorithm is a generalization of the Common Spatial Pattern (CSP) algorithm, which incorporates non-homogeneous weighting of the cross-spectrum matrices. We alternately update the spectral weighting coefficients and the spatial projection. The cross validation results of our SPECtrally-weighted CSP (SPEC-CSP) algorithm on 162 EEG datasets show significant improvements when compared to wide-band filtered CSP and similar accuracy as Common Sparse Spectral Spatial Pattern (CSSSP), however, with far less computational cost. The proposed method is at the same time highly interpretable and modular because the temporal filter is parameterized in the frequency domain. The possibility of incorporating any prior filter opens up the applicability of the method far beyond brain signals.