Multi-level hierarchical dynamic graph convolutional networks for motor imagery EEG analysis

Xin Deng, Daijiang Xu, Huaxiang Huo, Xinyi Hong, Bin Liu

Abstract: The human brain network consists of tightly connected modular nodes, including local functional areas and global functional connections. Graph Convolutional Networks (GCNs) have shown impressive capabilities in learning the topological relationships among electroencephalogram (EEG) channels during motor imagery tasks. However, existing GCN methods predominantly focus on a single-level functional connection mode, overlooking the diverse connection patterns across different functional modules and neglecting local functional dependencies. To address this limitation, this paper proposes a novel approach called the Multi-level Hierarchical Dynamic GCN (MHD-GCN) method, which aims to explore the multi-layer dynamic information among EEG channels within distinct functional modules. Firstly, the spatiotemporal attention module is employed to capture critical sequence fragments and spatial position information within EEG signals. Secondly, a multi-level wavelet convolutional block is utilized to capture distinct spectral features from corresponding subbands. Lastly, a pre-constructed topological space is leveraged for representation learning, comprising two branches dedicated to exploring global dynamic information and extracting local functional information, respectively. In each branch, the multi-level adjacency matrix is employed to enhance the expressive power of GCN, enabling the acquisition of discriminative features. Extensive experiments on publicly available datasets show that our proposed method performs better than state-of-the-art approaches on the Physionet dataset. Moreover, interpretability analysis is performed on the model, revealing active brain regions and important electrode pairs associated with movement. These findings provide valuable insights and practical guidance for future research in motor imagery tasks.