LLM-SDaT: A knowledge-informed LLM framework for syndrome differentiation in TCM.

Area of Science:

• Artificial Intelligence
• Traditional Chinese Medicine
• Computational Linguistics

Background:

• Large language models (LLMs) show promise for Traditional Chinese Medicine (TCM) applications.
• Current TCM AI methods lack standardized data and structured knowledge integration, limiting accuracy.

Purpose of the Study:

• To develop LLM-SDaT, a knowledge-informed framework for parameter-efficient fine-tuning of LLMs in TCM.
• To improve precision and interpretability in TCM syndrome differentiation and treatment planning.

Main Methods:

• Introduced TCMSD100 (100-syndrome clinical corpus) and TCMSDaT100 (knowledge-integrated dataset).
• Implemented a two-stage fine-tuning framework using LoRA (Low-Rank Adaptation).
• Stage 1: Syndrome differentiation on TCMSD100. Stage 2: Treatment recommendation aligned with TCMSDaT100.

Main Results:

• Achieved an 85.19% F1-score in syndrome classification.
• Significantly outperformed existing baselines and general-purpose LLMs.
• Demonstrated superior performance in generating clinically coherent and personalized treatment recommendations.

Conclusions:

• Integrating structured knowledge via parameter-efficient adaptation enhances LLM performance in TCM.
• LLM-SDaT offers a scalable pathway for interpretable TCM decision-support systems.
• Publicly available datasets and code facilitate further research and development.