Abstract
Objectives: Existing motor imagery electroencephalography (MI-EEG) decoding approaches are constrained by their reliance on sole representations of brain connectivity graphs, insufficient utilization of multi-scale information, and lack of adaptability. Methods: To address these constraints, we propose a novel Local–Partition–Global Graph learning Network (LPGGNet). The Local Learning module first constructs functional adjacency matrices using partial directed coherence (PDC), effectively capturing causal dynamic interactions among electrodes. It then employs two layers of temporal convolutions to capture high-level temporal features, followed by Graph Convolutional Networks (GCNs) to capture local topological features. In the Partition Learning module, EEG electrodes are divided into four partitions through a task-driven strategy. For each partition, a novel Gaussian median distance is used to construct adjacency matrices, and Gaussian graph filtering is applied to enhance feature consistency within each partition. After merging the local and partitioned features, the model proceeds to the Global Learning module. In this module, a global adjacency matrix is dynamically computed based on cosine similarity, and residual graph convolutions are then applied to extract highly task-relevant global representations. Finally, two fully connected layers perform the classification. Results: Experiments were conducted on both the BCI Competition IV-2a dataset and a laboratory-recorded dataset, achieving classification accuracies of 82.9% and 87.5%, respectively, which surpass several state-of-the-art models. The contribution of each module was further validated through ablation studies. Conclusions: This study demonstrates the superiority of integrating multi-view brain connectivities with dynamically constructed graph structures for MI-EEG decoding. Moreover, the proposed model offers a novel and efficient solution for EEG signal decoding.