An efficient domain adaptation framework for cross-subject EEG classification

Noninvasive brain-computer interface (BCI) have attracted great attention for exploration and rehabilitation in bioinformatics and biomedicine. Recently, electroencephalogram (EEG) has been widely adopted for BCI due to its low-cost and convenience. However, for the classification of EEG based on motor imagery (MI) or event-related potentials (ERP), most existing methods suffered from the variability of samples across subjects. To tackle this issue, the cross-subject scenario based on domain adaptation methods has been proposed for MI-EEG or ERP-EEG classification. To meet a large amount of samples during cross-subject classification, recently methods struggled with a large amount of computation costs and the complex optimization and regularization with parameters. To solve this problem, this paper proposed an efficient domain adaptation framework to exploit intra-subject structure of EEG representations without parameters tuning and select the discriminative samples. Meanwhile, the fastest domain adaptation methods, such as correlation alignment and geodesic covariance alignment, can be simply embedded in such framework. Empirical studies on four bench mark MI/ERP-EEG datasets have revealed the feasibility and effectiveness of the proposed framework, which achieves a great classification performance improvement with impressive efficiency.