Multi-Task EEG Signal Classification Using Correlation-Based IMF Selection and Multi-Class CSP

In the context of motor imagery (MI)-based brain-computer interface (BCI) systems, a great amount of research has been studied for attaining higher classification performance by extracting discriminative features from MI-based electroencephalogram (EEG) signals. In this study, we propose an innovative approach for classifying multi-class MI-EEG signals, which consists of a signal processing technique based on empirical mode decomposition (EMD) and multi-class common spatial patterns (MCCSP). Specifically, after applying the EMD, we propose selecting the best intrinsic mode functions (IMF) as the substitution to the original EEG signal for the next stage of processing. The metric we used for the selection is based on the cross-correlation of each decomposed IMF with the original signal. Next, we extend the CSP algorithm to the MCCSP to be utilized as the feature extractor. We applied our technique to the BCI competition IV (2a). Results revealed that the proposed technique improved classification accuracy significantly compared to the original case when applying MCCSP directly to the original EEG channel data. Moreover, the K-nearest neighbor (KNN) achieved the highest mean classification accuracy rate of 91.28%. Our findings suggest that a promising elevated classification accuracy of 96.71% can be achieved by raising the feature dimension through MCCSP. Compared to state-of-the-art algorithms, the performance of the proposed method is highly convincing and motivating for future studies.