PresRecRF: Herbal prescription recommendation via the representation fusion of large TCM semantics and molecular knowledge.

Background: Herbal prescription recommendation (HPR) is a hotspot in the research of clinical intelligent decision support. Recently plentiful HPR models based on deep neural networks have been proposed. Owing to insufficient data, e.g., lack of knowledge of molecular, TCM theory, and herbal dosage in HPR modeling, the existing models suffer from challenges, e.g., plain prediction precision, and are far from real-world clinics.

Purpose: To address these problems, we proposed a novel herbal prescription recommendation model with the representation fusion of large TCM semantics and molecular knowledge (termed PresRecRF).

Study Design And Methods: PresRecRF comprises three key modules. The representation learning module consists of two key components: a molecular knowledge representation component, integrating molecular knowledge into the herb-symptom-protein knowledge graph to enhance representations for herbs and symptoms; and a TCM knowledge representation component, leveraging BERT and ChatGPT to acquire TCM knowledge-enriched semantic representations. We introduced a representation fusion module to effectively merge molecular and TCM semantic representations. In the herb recommendation module, a multi-task objective loss is implemented to predict both herbs and dosages simultaneously.

Results: The experimental results on two clinical datasets show that PresRecRF can achieve the optimal performance. Further analysis of ablation, hyper-parameters, and case studies indicate the effectiveness and reliability of the proposed model, suggesting that it can help precision medicine and treatment recommendations.

Conclusion: The entire process of the proposed PresRecRF model closely mirrors the actual diagnosis and treatment procedures carried out by doctors, which are better applied in real clinical scenarios. The source codes of PresRecRF is available at https://github.com/2020MEAI/PresRecRF.