Multi-Region Temporal–Spatial Attention Network for Motor Imagery EEG Decoding

Accurate decoding of motor imagery (MI) EEG signals is a critical challenge in brain-computer interface (BCI) systems. However, MI-EEG signals exhibit significant spatial heterogeneity across brain regions and distinct temporal dynamics during motor imagery, posing challenges for effective modeling. To address these challenges, this paper proposes a Multi-Region Temporal–Spatial Attention Network (MRTSANet) for MI-EEG classification. First, raw EEG signals undergo bandpass filtering, with channels divided into four brain regions—frontal, parietal, temporal, and occipital—based on functional anatomical structures. Subsequently, independent convolutional neural network branches are designed for each region to extract multi-scale temporal features and intra-regional spatial features. Features from each branch are concatenated and fed into a bidirectional long short-term memory (BiLSTM) network to model temporal dependencies within the motor imagery phase. Finally, a temporal attention mechanism is introduced to adaptively emphasize discriminative temporal segments. The extensive experiments conducted on the publicly available BCI IV 2a dataset achieved a higher accuracy of 80.52% compared to other models. Experimental results demonstrate that the proposed method effectively integrates spatial characteristics of brain regions with temporal dynamic information, enhancing the classification performance of motor imagery EEG signals.