An LLM Method for Understanding Traditional Chinese Medicine: Mechanism Exploration and Innovative Application

Large language models (LLMs) show promise in medical knowledge representation but struggle with dynamic clinical workflows and personalized treatment in complex systems like Traditional Chinese Medicine (TCM). We propose an efficient and novel LLM framework for TCM mechanism exploration and clinical application, combining incremental domain-specific pre-training, multi-task supervised fine-tuning, and Chain-of-Thought (CoT) reasoning. Our two-stage approach-"Understanding and Inheritance" followed by "Exploration and Innovation"-uniquely leverages a heterogeneous database of 100,538 records from 19 TCM physicians to model the core TCM principle of "different treatments for the same disease". Six downstream tasks assess clinical capabilities, including personalized prescription generation (Task 3). After incremental pre-training, the model improves BLEU-4 by 1,313% over baseline, reaching 41.26-43.21 after fine-tuning. We quantify physician-specific variations and formally validate the decisive role of basic formulas-removing them causes a 23.9% performance drop. Cross-school evaluations confirm robust generalization, with 22.8 BLEU-4 on external data. CoT annotation boosts performance by 20% using only 10% labeled data, demonstrating high data efficiency. The model captures TCM's "different treatments for the same disease" principle and preserves school-specific diagnostic logic. This work advances intelligent TCM inheritance and paves the way for AI-driven personalized medicine.