An Efficient Framework for Cross-Session EEG Classification Based on Harmony Search Common Spatial Pattern

Motor Imagery (MI) decoding is fundamental to Brain-Computer Interfaces (BCI), playing a crucial role in rehabilitation and robotic control. The effectiveness of these systems depends significantly on the precise extraction of features from EEG signals, which is essential for translating user intent into commands for external devices. Traditional methods often fail to accommodate individual differences when selecting optimal frequency bands for Common Spatial Pattern (CSP) analysis. To address this limitation, we developed the Harmony Search Common Spatial Pattern (HSCSP) approach, which utilizes the Harmony Search (HS) optimization algorithm to dynamically refine frequency band selection. This method efficiently identifies optimal bands by systematically evaluating the impact of different selections on classification accuracy. Subsequently, features are extracted using Linear Discriminant Analysis (LDA) from the CSP-processed EEG data, maximizing inter-class distinction and intra-class compactness. Cross-session experiments on the BCI Competition IV 2a dataset validate the efficacy of our approach, achieving an accuracy of 73.9%, significantly outperforming existing benchmark methods. This adaptive feature significantly enhances HSCSP’s utility across diverse subjects, highlighting its superiority in personalizing BCI applications.