Optimizing Common Spatial Pattern for a Motor Imagerybased BCI by Eigenvector Filteration

One of the fundamental criterion for the successful application of a brain-computer interface (BCI) system is to extract significant features that confine invariant characteristics specific to each brain state. Distinct features play an important role in enabling a computer to associate different electroencephalogram (EEG) signals to different brain states. To ease the workload on the feature extractor and enhance separability between different brain states, the data is often transformed or filtered to maximize separability before feature extraction. The common spatial patterns (CSP) approach can achieve this by linearly projecting the multichannel EEG data into a surrogate data space by the weighted summation of the appropriate channels. However, choosing the optimal spatial filters is very significant in the projection of the data and this has a direct impact on classification. This paper presents an optimized pattern selection method from the CSP filter for improved classification accuracy. Based on the hypothesis that values closer to zero in the CSP filter introduce noise rather than useful information, the CSP filter is modified by analyzing the CSP filter and removing/filtering the degradative or insignificant values from the filter. This hypothesis is tested by comparing the BCI results of eight subjects using the conventional CSP filters and the optimized CSP filter. In majority of the cases the latter produces better performance in terms of the overall classification accuracy.