DFGLM-TCM: an integrated knowledge-and experience-driven large language model system for Traditional Chinese Medicine practice
TL;DR
DFGLM-TCM, a modular large language model-based service system that separately models these two knowledge types through task-oriented components and coordinates them within a unified multi-task architecture, highlights the value of separating general TCM knowledge from practitioner-specific experience while coordinating task-oriented training and multi-task services within a unified system.
Abstract
Traditional Chinese Medicine (TCM) clinical practice depends on both codified theoretical knowledge and practitioner-specific experience, which differ in their data sources, reasoning patterns, and scope of generalization. We developed DFGLM-TCM, a modular large language model-based service system that separately models these two knowledge types through task-oriented components and coordinates them within a unified multi-task architecture. The knowledge-oriented component integrated a curated 22-GB TCM corpus containing approximately 3 million structured entries, a manually validated knowledge graph with more than 200,000 entities, and retrieval-augmented generation for literature retrieval and general TCM question answering. The experience-oriented component was trained using approximately 5000 authentic outpatient records from a senior TCM practitioner and 2000 expert-reviewed augmented cases to provide practitioner-specific diagnostic and prescription references. Through standardized interfaces and role-based access, the system supports knowledge retrieval, question answering, structured consultation, and prescription reference for clinicians, patients, and students. In an expert-rated evaluation of 100 TCM knowledge questions, DFGLM-TCM achieved the highest descriptive mean score among the evaluated models (4.48 ± 0.76). In 454 independent pulmonary-nodule cases, the experience-oriented component achieved a prescription-consistency score of 8.95 ± 0.63, exceeding that of the model additionally trained with general TCM knowledge (7.52 ± 0.44). A one-month assessment involving 35 physicians at three primary-care institutions suggested favorable short-term usability and acceptance. These findings highlight the value of separating general TCM knowledge from practitioner-specific experience while coordinating task-oriented training and multi-task services within a unified system, and support further evaluation of DFGLM-TCM as an auxiliary reference tool in broader clinical settings.