GMDNet: Graph-Augmented Multimodal Dual-Path Network for EEG Motor Imagery Decoding

Motor Imagery (MI) decoding technology based on electroencephalogram (EEG) has broad application prospects in brain-computer interface and other related fields. However, existing methods suffer from two main limitations: (1) Processing the raw EEG signal holistically neglects the distinct information carried by different intrinsic frequency sub-bands, limiting the extraction of fine-grained, mode-specific features; (2) Most methods' reliance on standard convolutional networks inadequately captures the complex spatial correlations among electrode channels, failing to fully model the underlying functional brain connectivity. To overcome these limitations, we propose a Graph-Augmented Multimodal Dual-Path Network named GMDNet, which achieves a comprehensive fusion of spatio-temporal features. The first path, Local Modal Analysis, employs Multi-variable Variational Mode Decomposition (MVMD) to deconstruct the signal into multiple frequency bands, capturing detailed modal dynamics. The second path, Global Structural Embedding, utilizes a Graph Convolutional Network (GCN) to explicitly model the spatial topology of the electrode array. By adaptively learning the weights between different EEG electrode channels, this global path captures deep correlations across different brain regions. Evaluated on the BCI Competition IV 2a dataset show that the model achieves an accuracy of 80.79% in the four-category cross-session MI task, outperforming other leading methods. Moreover, the spatial features extracted by the model are highly consistent with the functional connectivity patterns of the brain during MI tasks. Our work demonstrates that GMDNet provides a powerful and robust framework for EEG decoding, holding significant promise for both BCI applications and broader neurophysiological analysis.