Cross-domain Spatio-temporal and Multi-scale Residual Attention Mechanisms for Robust EEG Motor Imagery Classification

Motor imagery (MI)-based classification of EEG signals plays a crucial role in developing effective brain–machine interface (BMI) systems used in assistive control and neuro-rehabilitation. However, conventional approaches face challenges due to low signal clarity, subject-wise variability, and insufficient utilization of spatial and temporal signal dependencies. Addressing these concerns, this study proposes a unified spatial–temporal multi-scale attention mechanism (UST-MSAM), which combines cross-domain spatial–temporal attention and dynamic residual multi-scale attention. The design incorporates graph-guided attention layers to extract inter-channel spatial dynamics and utilizes frequency-adaptive attention paths to uncover salient temporal cues across EEG sub-bands. In addition, a hybrid encoder with residual attention refinements is employed to suppress irrelevant signal components and enhance the retention of critical features. The performance of the proposed method was evaluated on two benchmark datasets. For the BCI dataset, UST-MSAM achieved an accuracy of 97.5%, outperforming existing models such as BiLSTM (94.0%), ADBN-FNO (95.7%), and SSTS-Net (96.9%) by margins of 3.5%, 1.8%, and 0.6%, respectively. Similarly, on the PhysioNet dataset, it attained a peak accuracy of 96.4%, marking improvements over BiLSTM (93.0%), ADBN-FNO (94.1%), and SSTS-Net (95.9%). These results confirm consistent gains in classification accuracy, with added improvements in specificity (up to 98.7% for BCI and 96.8% for PhysioNet), precision, recall, and F1-measure across both datasets. The findings demonstrate the potential of UST-MSAM in building robust and generalizable EEG classification systems for real-world neuro-interfacing applications.