Abstract
Traditional Chinese Medicine (TCM) clinical informatization focuses on serving user-oriented health knowledge and facilitating online diagnosis. Regularities hidden in clinical knowledge play a significant role in the improvement of the TCM informatization service. However, many regularities can hardly be discovered because of specific data-challenges in TCM prescriptions at present. Therefore, in this article, we propose an end-to-end model, called Semantic-aware Graph Convolutional Networks (SaGCN) model, to learn the latent regularities in three steps: (1) We first construct a heterogeneous graph based on prescriptions; (2) We stack Semantic-aware graph convolution to learn effective low-dimensional representations of nodes by meta-graphs and self-attention; (3) With the learned representations, we can detect regularities accurately by clustering and linked prediction. To the best of our knowledge, this is the first study to use metagraph and graph convolutional networks for modeling TCM clinical data and diagnosis prediction. Experimental results on three real datasets demonstrate SaGCN outperforms the state-of-the-art models for clinical auxiliary diagnosis and treatment.
Introduction
Traditional Chinese Medicine (TCM) relies on a holistic approach where "regularities"—the latent patterns between symptoms, syndromes, and herbs—are key to effective treatment. However, digitizing this knowledge is notoriously difficult due to three primary data challenges. First, prescriptions often exhibit a random structure, recorded in free-text natural language. Second, clinical entities are frequently disordered; an herb listed first might be most related to a symptom mentioned last. Finally, data sparsity arises from personalized prescribing habits, leading to inconsistent records where critical symptoms or herbs might be omitted.
Existing NLP and topic-modeling approaches often fail to capture the multi-layered semantics of how these entities interact. To address this, we propose SaGCN (Semantic-aware Graph Convolutional Networks), an end-to-end deep learning model designed to decode these complex relationships.
Methodology
The SaGCN framework operates in three main stages to transform raw clinical data into actionable insights.
1. Heterogeneous Graph Construction
Prescriptions are transformed into a Heterogeneous Information Network (HIN). Unlike simple graphs, an HIN contains multiple types of nodes (Herbs, Symptoms, Syndromes, Diseases) and diverse link types, capturing the multi-faceted nature of TCM clinical data.
2. Meta-graphs and Self-Attention
To capture complex domain knowledge, SaGCN utilizes meta-graphs—predefined structural templates that describe specific medical relationships (e.g., Herb → Symptom → Disease). Unlike simpler linear "meta-paths," meta-graphs can model non-linear dependencies. Because different paths through the graph carry varying clinical significance, a self-attention mechanism is implemented to dynamically learn which semantic relations are most important for a given node.
3. Semantic-aware Convolution
The model adapts the Graph Convolutional Network (GCN) to be "semantic-aware." This allows the model to aggregate information from a node’s neighbors while respecting their specific medical roles. For instance, it can distinguish between a primary "emperor" herb and an assistant "zuo" herb, a nuance often missed by standard graph models.
Experimental Results
The model was evaluated on three real-world TCM datasets: TCMRel (relation graph), CMD (medical information graph), and CLLT (7,000+ lung tumor prescriptions). SaGCN was tested against state-of-the-art models like ASPEm, GWCN, and HeteroMed across two primary tasks:
- Similarity Discovery (Clustering): SaGCN achieved superior Accuracy and Normalized Mutual Information (NMI) scores. On the CMD dataset, it reached an herb clustering accuracy of 0.7587, significantly outperforming the topic-model baseline TM2P (0.5720).
- Link Prediction (Diagnosis): In predicting herb-symptom pairings, SaGCN achieved a Mean Average Precision (MAP@10) of 0.8745 on the TCMRel dataset, outperforming the next best model (MotifCNN) by nearly 2%.
In a qualitative case study involving lung tumor prescriptions, SaGCN’s herb recommendations matched a doctor’s actual prescriptions with a recall of 0.833. The model also suggested clinically valid alternative herbs that were later verified by medical professionals.
Conclusion and Future Work
SaGCN demonstrates the power of meta-graphs and semantic-aware convolution in uncovering the latent regularities of TCM. Future work will focus on incorporating herb dosage as numerical data, as the current model treats herb presence as binary. Additionally, integrating explicit prior knowledge from medical textbooks into the embedding process could further refine the model's accuracy and its ability to infer traditional roles like Emperor, Minister, Assistant, and Envoy.
Summary updated