Multi-classification for EEG Motor Imagery Signals using Auto-selected Filter Bank Regularized Common Spatial Pattern

Motor Imagery (MI) is a critical topic in Brain-Computer Interface (BCI). Due to the low signal-to-noise ratio, it is not easy to accurately classify motor imagery signals, especially for multiple classification tasks. Common Spatial Pattern (CSP) is a spatial transformation method that can effectively extract spatial features of EEG signals. However, the covariance matrix is inaccurate due to the small training data size,. Thus, in this paper, a regularization parameter auto-selection algorithm is proposed to automatically adjust the ratio of the covariance matrix calculated by other subjects’ data based on the mutual information. It can be used to tackle the problem of an inaccurate mixed covariance matrix caused by fixed regularization parameters.To illustrate the merits of the proposed Auto-selected Filter Bank Regularized Common Spatial Pattern (AFBRCSP), we used the ten folds cross-validation accuracy and Kappa as the evaluation metrics to evaluate two data sets (BCI4-2a and BCI3a data set). Both data set include four mental classes. By using BCI4-2a data set, we found that the mean accuracy of AFBRSP is 77.31% and the Kappa is 0.6975, which is higher than Filter Bank Regularized Common Spatial Pattern (FBRCSP) by 5.67% and 0.0756, respectively. By using BCI3a data set, the proposed AFBRCSP improved the accuracy by 8.34% and the Kappa by 0.1111 compared with FBRCSP where the mean accuracy of AFBRCSP is 80.56%, and the kappa is 0.7407. The overall Kappa obtained by the proposed method is also higher than some state-of-the-art methods, implying that the proposed method is more reliable.