Multiscale Spatiotemporal Neural Network with Multi-Attention Mechanism Using Brain Partitioning for Motor Imagery Recognition

Motor imagery (MI)-based electroencephalogram (EEG) brain-computer interfaces (BCIs) facilitate communication for motorimpaired patients by leveraging artificial intelligence to accurately interpret brain signals. However, EEG signal classification remains challenging due to low signal-to-noise ratio (SNR) and individual variability in brain activity. We propose a novel parallel multi-depth spatial-temporal neural network aimed at enhancing the integration of spatial and temporal features from multichannel EEG signals by leveraging brain functional topography. To improve cortical representations associated with motor imagery, the model incorporates two parallel branches. One branch focuses on inter-channel differences corresponding to contralateral electrode pairs, emphasizing hemispheric disparities, while the other targets the frontal and parietal brain regions. These region-specific enhanced signal representations are then fed into the multi-depth spatial-temporal network for feature extraction and subsequent motor imagery classification. The architecture of the feature extraction network integrates four specialized blocks, ensuring the comprehensive capture of discriminative features that are particularly sensitive to task-relevant frequencies for each MI class. A multi-loss design further optimizes feature integration across networks. Cross-validation results on the BCI Competition IV-2a dataset and High Gamma dataset achieve accuracies of 82.14% and 95.21%, respectively, with kappa values of 0.76 and 0.92, surpassing state-of-the-art methods. These experimental results highlight the significance of parallel spatial-temporal networks based on brain partitioning for MI classification in rehabilitation engineering and real-world BCI applications.