Correlation-Based Regularized Common Spatial Patterns for Classification of Motor Imagery EEG Signals

Common Spatial Patterns (CSP) is a powerful and common method for effective feature extraction and dimensionality reduction in Brain-Computer Interface (BCI) applications. However CSP has some shortcomings, particularly, it is sensitivity to noise and outlier data which results in lower classification accuracy. In this paper, we propose a regularized version of the original CSP (Corr-CSP), in which the objective function is penalized by a properly designed penalty term which encourages decorrelation between the data from two classes in such a way that the resulting objective function has still straightforward solution through Eigen value decomposition. Furthermore, we have used three different datasets from the BCI Competition BCI database in order to evaluate the performance of the proposed approach and compare it to the original CSP. The simulation results show on the average 4% of improvement in terms of classification accuracy for the proposed Corr-CSP approach.