MIBC-GCN: a topology-optimized EEG classification model

Graph Convolutional Neural Networks (GCNs) have achieved remarkable success in motor imagery (MI) recognition and decoding. As an endogenous neural activation approach that closely approximates voluntary movement, MI serves as a critical technological foundation for intention-driven rehabilitation in patients with spinal cord injury (SCI). In GCNs, the adjacency matrix plays an essential role by providing the necessary foundational structural information for graph learning and reasoning; however, traditional construction methods mainly rely on simple trial-level functional connectivity, leading to redundant and unstable connections that hinder classification accuracy and practical brain–computer interfaces (BCIs) application. To address this, MIBC-GCN, a graph convolutional model, has been proposed. It integrates the adjacency matrices constructed via normalized mutual information (NMI) and betweenness centrality (BC), combined with an adaptive graph convolutional block. On the collected dataset, the proposed method MIBC-GCN improved classification accuracy by 12% over fully connected mutual information–based matrices while reducing training time. This approach utilizes user-specific adjacency matrices and maintains robust performance through cross-participant tunable parameters across datasets with varying channel counts. Optimal results were achieved with first-order graphs (k = 1), highlighting the method's potential for personalized BCI interaction paradigms and providing an algorithmic foundation and for the online control of personalized neurorehabilitation devices.