Spatiotemporal Feature-Enhanced Bi-directional Mamba Network for Motor Imagery EEG Decoding

Abstract The accurate decoding of motor imagery electroencephalography (MI-EEG) signals is critical for brain-computer interface(BCI) applications in motor rehabilitation. However, existing methods often suffer from insufficient spatiotemporal featurefusion, high computational complexity, and limited generalization capability. To address these limitations, this paper proposesSTE-BiMamba, a spatiotemporal feature-enhanced bidirectional Mamba model for efficient MI-EEG decoding. The modelincorporates three core components: a multi-scale temporal feature extraction module that employs parallel convolutions withdifferent kernel sizes to simultaneously capture transient event-related desynchronization and long-term rhythmic patterns, athree-view statistical pooling module that extracts highly discriminative spatiotemporal representations from mean, variance,and maximum perspectives, and a bidirectional Mamba global sequence modeling module that efficiently captures bidirectionallong-range dependencies with linear time complexity. Extensive experiments on our self-collected VR-MI dataset and threepublic datasets (BCIC-IV-2a, BCIC-IV-2b, HGD) under both within-subject cross-session and cross-subject settings validate theeffectiveness of STE-BiMamba. On VR-MI, the model achieved 97.50% accuracy (cross-session) and 85.05% (cross-subject),while attaining state-of-the-art results on all public datasets. Compared with Transformer-based models, STE-BiMamba doublesthe inference speed while reducing the number of parameters by a factor of 3.5. STE-BiMamba offers a high-precision andlightweight solution for practical motor imagery BCI applications