Modeling symptom-acupoint interactions via a heterogeneous graph learning framework for intelligent acupoint recommendation

Objective Acupuncture prescriptions involve complex compatibility mechanisms grounded in multi-symptom and multi-acupoint interactions, embodying millennia of clinical experience. Despite growing interest in computational acupoint recommendations, significant challenges persist due to sparse clinical data and the insufficient modelling of symptom-acupoint relationships, posing considerable hurdles to effective prediction. Methods We introduced an acupoint compatibility prediction framework with graph neural networks (GNN) and fine-tuned bidirectional encoder representations from transformers (termed GNN-BERT-Attention). The heterogeneous feature interaction learning mechanism was introduced to model symptom-acupoint interactions through heterogeneous graph construction, capturing semantic features and relational patterns in a unified space, which alleviated the sparsity of data. Neural collaborative filtering is utilised via label-aware fusion to iteratively refine the confidence of predictions, while Focal Loss and randomised augmentation strategies enhance robustness against imbalanced label distribution. Results Comprehensive experiments demonstrate the superiority of the proposed GNN-BERT-Attention model over State-of-the-Art (SOTA) baselines in precision, recall, ranking-based metrics and robustness. Ablation studies validate the effectiveness of each architectural module, and hyperparameter tuning confirms models’ stability. A web-based demonstration system further validates clinical applicability, enabling real-time, interpretable acupoint recommendations. Conclusion This study contributes to enhancing the performance of acupoint prediction, ultimately benefiting the efficiency and precision of acupuncture treatment while providing a theoretical foundation for optimising prescriptions and advancing evidence-based traditional Chinese medicine interventions.