Hierarchical Window Attention for Motor Imagery EEG Classification

Electroencephalography (EEG) plays a pivotal role in brain-computer interface (BCI). Among the various paradigms, motor imagery (MI) stands out as a spontaneous and valuable paradigm for applications in both cognitive neuroscience and clinical rehabilitation. The rapid propagation of action potentials suggests that crucial neural information may reside within local temporal domains, an aspect that has been underemphasized in previous study. To address this issue, we focus on extracting features from temporal neighborhoods of different scales and propose the utilization of a hierarchical window attention network for MI-EEG classification. Specifically, the temporal spatial embedding (TSE) module transforms EEG signals into token sequence. The local window attention (LWA) module and a hierarchical structure are devised for adaptive learning of local temporal dependencies within non-overlapping windows at various scales. Additionally, we employ a segmentation and reconstruction strategy for data augmentation. Our method outperforms other approaches, yielding an average classification accuracy of 90.84% on the BCI III-3a dataset and 83.14% on the BCI IV-2a dataset. Ablation studies and feature visualization further affirm the effectiveness of these modules.