Combination of Large Language Model and Retrieval-augmented Generation for Inference of Traditional Chinese Medicine Prescriptions and Syndrome Differentiations: A Study on Sleeping Disorders

Abstract

Propose: The objective of this study was to combine a large language model (LLM) with retrieval-augmented generation (RAG) to develop a tool for syndrome type determination and a medical prescription inference model for sleep disorders in a generative artificial-intelligence-based traditional Chinese medicine (TCM) system. The tool uses the disease-formula-syndrome inference framework in TCM to map various combinations of prescription formulas and syndromes. Methods: This study selected 6,747 cases of sleep disorders along with their single- and compound-herb formula prescriptions (finished herbal products) to construct the RAG knowledge base. A RAG-based LLM was then employed to generate medical prescription formulas corresponding to syndrome types, as well as for reverse generation and validation. Results: The results demonstrate that the RAG-based LLM can determine TCM syndrome types that align closely with clinical reality. Moreover, it can consistently provide comprehensive syndrome mappings for various combinations of prescription formulas. However, limitations were observed in a few instances where only single herbs were identified and compound-herb formulas were either omitted or incorrect, resulting in significantly lower scores. Conclusion: The results indicate that the proposed method can enhance learning efficiency in TCM diagnostics and support clinical diagnosis and the RAG-based LLM need to be improved to provide comprehensive responses. Future efforts could focus on expanding the dataset and optimizing the LLM to enhance accuracy and reliability.