Multi-label classification of visceral diseases based on human acupoint networks via local-global graph neural networks

Identifying acupoint-disease associations is of paramount significance. However, existing approaches generally suffer from key limitations: they either overlook potential heterogeneous correlations among samples or fail to fully exploit unstructured individual-specific features. To address these issues, we innovatively construct a novel human acupoint network using mutual information between acupoints, enabling the discovery of previously unexplored latent topological dependencies. A community detection method is then applied to partition the network into structurally consistent local subgraphs. Furthermore, a novel multimodal local-global Acupoint Graph Neural Network (LGA-GNN) is proposed for multi-label classification of visceral diseases. The LGA-GNN framework comprises two core components. The local ACU-GNN module learns structural embeddings from the partitioned subgraphs and identifies key acupoint nodes indicative of visceral diseases. The global Subject-GNN module integrates local embeddings with nonacupoint data (gender, age) to capture heterogeneous inter-individual relationships relevant to disease labels. To enhance the discriminative power of local embeddings, a Feature-Guided Propagation Pooling (FGP-Pooling) mechanism is incorporated into the ACU-GNN to retain diagnostically significant node features. Meanwhile, the Subject-GNN employs a forward information aggregation module to adaptively fuse multi-scale graph features, improving the model’s representation capacity and classification performance. The proposed LGA-GNN is comprehensively evaluated on a large-scale real-world dataset. Experimental results achieve an average AUC of 0.94 and ACC of 0.88 under five-fold cross-validation. The model significantly outperforms state-of-the-art baselines across multiple metrics, demonstrating strong effectiveness and superiority in multi-label disease classification.