Tri-branch spatio-temporal learning network with dendritic aggregation based dynamic GCN for multi-lead ECG classification

Electrocardiogram (ECG) is a widely used non-invasive diagnostic tool for cardiovascular disease detection. Its complex spatio-temporal properties pose challenges for effective representation learning. Existing methods mainly focus on spatio-temporal information extraction, but generally ignore the cooperative relationship among temporal, spatial, and their joint spatio-temporal characteristics, which weakens the model’s ability to learn comprehensive ECG representations. To address it, a Tri-Branch Spatio-Temporal Learning Network (Tri-STLN) is proposed for classification of multi-lead ECG signals by learning complementary temporal, spatial, and spatio-temporal representations. Specifically, the Tri-STLN comprises three parallel branches to extract temporal, spatial and spatio-temporal information, respectively. In the temporal and spatial branches, a Dendritic Aggregation based Dynamic Graph Convolutional Network (DA-DGCN) module is proposed. This module integrates the structured representation learning capability of Dendritic Neural Model (DNM) with the adaptive learning capability of DGCN, and incorporates prior constraints into the edge-generation process. Therefore, it can effectively capture temporal or spatial information by modeling the multi-lead structure and dynamic characteristics of ECG signals. In the spatio-temporal branch, a Spatio-Temporal Hypergraph Neural Network (ST-HGNN) module is developed, which constructs a unified hypergraph to capture higher-order spatio-temporal relationships. In addition, a disentangled representation learning (DRL) strategy is incorporated to enhance the complementarity among the three branch features, thereby improving the effectiveness of the global representation. Experiments on two public multi-label 12-lead ECG datasets demonstrate that the Tri-STLN outperforms all the compared algorithms, suggesting its effectiveness for the multi-label ECG classification tasks.