JMIR Publications

ABSTRACT

The accelerating loss of senior Traditional Chinese Medicine (TCM) practitioners is causing irreversible erosion of tacit clinical expertise—knowledge grounded in sensory intuition and dynamic, embodied mentorship. Conventional preservation methods fail to capture this interactive, context-sensitive practice. To address this gap, we developed an AI-driven digital mentor system that authentically simulates face-to-face master–apprentice teaching for structured capture and transmission of irreplaceable TCM tacit knowledge. We constructed a fine-grained knowledge base of 50,000 structured entries from the medical corpus of Master Xichun Zhang using automated semantic chunking and multimodal extraction. To evaluate retrieval performance, we created TCM-RAGBench, a domain-specific benchmark comprising 600 question–answer pairs. Our multimodal system integrates a domain-adapted Retrieval-Augmented Generation (RAG) framework—featuring task-specific strategies for diagnostic reasoning, herbal formula recommendation, and procedural guidance—with digital twin technologies (voice cloning and lip synchronization). Seven large language models (LLMs) were benchmarked on the Chinese Medical Benchmark for optimal selection. Technically, fine-grained knowledge structuring improved retrieval recall (Recall@1) from 59.30% to 72.00% on TCM-RAGBench. The system also demonstrated expert-level proficiency, achieving 93.95% and 96.65% accuracy on the national TCM practitioner and pharmacist licensing examinations in China, respectively. To assess educational efficacy, we recruited 24 TCM students for a single-blind, randomized controlled trial. Through a six-month learning period with periodic assessments, the 12 participants who utilized the "AI-driven digital mentor system" developed in this study demonstrated significant learning gains in both objective test scores and clinical reasoning abilities evaluated by three blinded experts. Statistical data (p < 0.05) revealed that the group using the digital mentor system not only significantly outperformed the traditional learning group in overall scores (by 18.75%，p <0.001)), but more importantly, exhibited particularly pronounced advantages in complex clinical reasoning tasks(+20% accuracy) and in the acquisition of tacit clinical expertise across multiple dimensions.This study presents a reproducible framework for digitizing medical expertise through cognitive alignment and embodied interaction. Our findings demonstrate that high-fidelity digital mentorship can effectively enhance clinical reasoning and offer a scalable paradigm for modernizing medical education while safeguarding intangible medical heritage.