DC-TSCM: an interpretable dual-channel traditional Chinese medicine syndrome classification model <i>via</i> semantic-structural fusion

Background With the widespread adoption of electronic medical records, massive prescription data can be digitized and systematically stored. This provides a solid foundation for intelligent traditional Chinese medicine (TCM) diagnosis systems. TCM syndrome classification is the core of syndrome differentiation and treatment. Developing an effective classification framework remains a major challenge for intelligent diagnosis systems. Recent progress in natural language processing has introduced new approaches and tools for semantic understanding and knowledge extraction from prescription texts. However, traditional machine learning methods rely on hand-crafted features and struggle to process high-dimensional, sparse, and intricate TCM prescription texts. The single text-based model can capture semantic features but ignore the structural connections in prescription data. The single graph-based model emphasizes structural associations but fails to incorporate rich contextual semantics. Methods To address the challenges, we propose a new dual-channel TCM syndrome classification model (DC-TSCM) in healthcare applications. The text channel extracts deep representations from clinical description and physique detection texts. We developed a TCM differentiation-guided attention fusion module to dynamically learn the optimal weighting between prescription texts. The graph channel constructs a unique TCM differentiation heterogeneous graph and uses hybrid graph neural networks to model the complex semantic associations among clinical entities. Additionally, we extracted 8,280 prescriptions from real electronic medical records, covering 24 different syndrome types. The prescription data were standardized according to clinical diagnostic terminology and divided into training, validation, and test sets in an 8:1:1 ratio. Results Experiments were conducted on a structured multi-label syndrome differentiation dataset. The results indicate that the model achieves superior performance and strong generalization ability in multi-class syndrome classification. Its interpretability is further validated through visualization analysis, including the co-occurrence relationship heat map, confusion matrix, and receiver operating characteristic curve. The dual-channel model achieved an accuracy of 0.8919, precision of 0.9012, recall of 0.8947, and F1-score of 0.8930. Conclusion Overall, DC-TSCM bridges semantic understanding with structural reasoning and incorporates the principles of TCM differentiation. It significantly improves the accuracy of syndrome differentiation and suggests potential applicability beyond TCM, which could be explored in future work. It also provides a robust and interpretable framework for intelligent auxiliary diagnosis systems and lays a foundation for the integration of clinical knowledge with advanced deep learning methodologies.