The motor imagery EEG classification method combining common spatial pattern and ensemble learning

• Published in: 2021 6th International Conference on Communication, Image and Signal Processing (CCISP) pp. 361 - 366
• Main Authors: Du, Chenxiao, Shi, Chenyun, Huang, Huang, Wu, Xiaopei
• Format: Conference Proceeding
• Language: English
• Published: IEEE 01-11-2021
• Subjects: Brain-Computer Interface; Brain-computer interfaces; Built-in self-test; Common spatial pattern; Electroencephalography; Ensemble learning; Motor imagery; Signal processing; Spatial filters; Training
• DOI: 10.1109/CCISP52774.2021.9639289

The performance of motor imagery EEG (MI-EEG) based brain-computer interface (BCI) is limited by the nonstationarity of EEG signal and small-sized training datasets. Although the ensemble learning (EL) using multiple base classifiers was suggested to be promising to solve this problem, the mislabeled trials in MI-EEG datasets will bring great challenges to design a set of qualified base classifiers in the ensemble framework. In this paper, the novel EL method with multiple base classifiers was studied in the context of MI-classification. In view of the importance of spatial filter in BCI construction, we focused on the spatial filter design based on a small number of training samples. We proposed a MI classification method that combines common spatial pattern (CSP) with ensemble learning (EL-CSP). We experimentally demonstrated that the base classifier based on a small number of training samples meets the requirements of EL. Thirty MI datasets of six subjects were employed to validate the proposed method in this study. The experimental results showed that the accuracy of EL-CSP obtained relative increment of 9.5% during self-testing, 1.2% during session-to-session transfer, and 1.7% during subject-to-subject transfer compared to the traditional classification methods based on CSP, respectively. Moreover, the proposed method has higher stability.