Overall optimization of CSP based on ensemble learning for motor imagery EEG decoding

The common spatial pattern (CSP) is an effective feature extraction method in motor imagery-based brain-computer interface (BCI) system. However, CSP also has many defects. Existing CSP improvement methods only make partial improvements, without considering the overall optimization of CSP. In this paper, a new ensemble learning algorithm framework is proposed to improve the decoding performance of CSP, in which the regularization, temporal-spatial-frequency joint optimization, and pair number of spatial filters for CSP are comprehensively considered. First, a new temporal-spatial-frequency feature extraction method based on Tikhonov regularization CSP (TRCSP) is proposed, multiple feature subsets with diversity are extracted by TRCSP with different time windows, regularization parameters, and pair numbers of spatial filters. Second, the least absolute shrinkage and selection operator (LASSO) as base classification model is used for feature selection and classification, in which multiple diversified base classification models are trained. Finally, the base classification models with diversity and higher accuracy are used for ensemble model construction using a new integration rule, during which most of the temporal-spatial-frequency information is fully excavated and utilized. The effectiveness of the proposed method is verified by five motor imagery data sets and the average classification accuracy of all data sets is 85.99%. Compared with the existing CSP methods, the proposed method achieved a better classification effect, and with a small amount of calculation, low model complexity, and high robustness.