AI System Splits TCM Knowledge From Expert Experience to Aid Prescribing

Researchers in China have built a large language model system that deliberately keeps two very different kinds of Traditional Chinese Medicine expertise apart: the codified knowledge found in classical texts and textbooks, and the prescribing habits of a single senior practitioner. The system, called DFGLM-TCM, is described in a study published in the journal Chinese Medicine, and its modular design produced a striking result: adding general TCM knowledge to the expert-trained prescribing model actually made it less consistent with the expert’s own prescriptions.

The team, led by researchers including Ruikang Zhong, Dianna Liu and Kaiwen Hu, argued that TCM practice resists one-size-fits-all automation because clinical decisions draw on both standardized theory and context-dependent interpretation. Senior practitioners accumulate diagnostic and therapeutic patterns over decades that remain largely tacit, difficult to formalize in textbooks, and traditionally transmitted through apprenticeship with limited scalability. Rather than blending these knowledge types into a single model, as most prior TCM-oriented language models have done, DFGLM-TCM trains separate task-oriented components and coordinates them through standardized interfaces within a unified service architecture.

The knowledge-oriented component rests on substantial data infrastructure. The researchers assembled a curated TCM corpus of roughly 22 gigabytes, containing about 3 million structured text entries drawn from classical medical texts, modern textbooks, exercise collections, peer-reviewed literature, clinical guidelines and educational materials. Duplicate files and redundant editions were manually reviewed and excluded before ingestion, and optical character recognition followed by text cleansing and expert sampling supported quality control. Alongside the corpus, the team built a knowledge graph with more than 200,000 entities across eight categories, modeling relationships among diseases, symptoms, prescriptions and medicinal substances. Entity and relation extraction from 107 case-based classical texts, anchored by complete manual annotation of five foundational works including the Treatise on Febrile Diseases and Shennong’s Classic of Materia Medica, achieved a reported extraction accuracy of 86.7 percent before expert verification. The graph was implemented in a Neo4j database and combined with vector indexing to power retrieval-augmented generation.

The experience-oriented component took a different technical path. It was fine-tuned on approximately 5,000 authentic outpatient records from Professor Kaiwen Hu, a nationally recognized TCM practitioner with over three decades of experience in herbal medicine for tumor prevention and treatment, plus about 2,000 expert-reviewed augmented cases generated from extracted symptom-pattern-prescription-herb reasoning chains. Using full-parameter supervised fine-tuning of a GLM4-32B backbone, with diagnostic information encoded as reasoning chains, the model was trained to reproduce the practitioner’s prescribing logic rather than broad factual knowledge. Twenty percent of the dataset was held out for validation monitoring during training.

To test whether general knowledge would help or hurt expert-pattern modeling, the researchers ran a controlled comparison. Two models sharing the same backbone and optimization settings differed only in training data: one fine-tuned exclusively on the expert-derived dataset, the other further trained with 16,021 incremental general TCM knowledge samples. On the same evaluation cohort, the expert-only model achieved an overall prescription-consistency score of 8.95 plus or minus 0.63, while the incrementally trained model scored 7.52 plus or minus 0.44. The general-knowledge model generated prescriptions faster, at 9.72 versus 23.54 seconds on average, but tracked the practitioner’s patterns less closely. The authors cautioned that this balance should be carefully considered and does not constitute evidence of clinical efficacy or comprehensive safety.

The prescription evaluation itself was prospective and disease-specific. The team collected 454 pulmonary-nodule cases exclusively for testing, with CT-confirmed nodules, complete medical histories, informed consent, and no comorbid conditions affecting the chief complaint. Data included structured questionnaires, standardized tongue and face images, chest CT scans and radiology reports. No evaluation case overlapped with the training records. For each case, the system and the practitioner generated prescriptions independently and blindly; anonymized outputs were scored by two associate chief TCM physicians using a rubric developed by five senior clinicians, with disagreements adjudicated by a chief physician. Inter-rater agreement on the manually assessed compatibility item was excellent, with an ICC of 0.968. The system’s prescriptions had a median of 24 herbs, exceeded 90 percent coverage of the practitioner’s commonly used herbs, and passed both predefined safety checks on herb incompatibility and toxic-herb usage. The authors emphasized these scores measure prescribing-pattern consistency, not therapeutic benefit.

Knowledge question answering was assessed separately. Two TCM experts curated 100 questions spanning classical theories, syndrome concepts, treatment principles and clinical specialties, with standardized reference answers. DFGLM-TCM achieved the highest descriptive mean of 4.48 plus or minus 0.76, though DeepSeek-V3.2 at 4.41 and GPT-4o at 4.36 were close behind. The authors noted that general-purpose models performing comparably suggests domain-specific adaptation may not confer a uniform advantage as commercial LLMs improve. Overall inter-rater reliability across all responses was high, with an ICC of 0.888.

The full system was deployed as a layered application with role-based access for clinicians, patients and students. Clinicians can enter structured records, review AI-generated syndrome differentiation and practitioner-specific prescription references, and query the knowledge base. Patients complete QR-code symptom questionnaires before consultation, and a dialogue engine guides multi-round structured information collection. Students access classical texts linked to a visualized knowledge graph. The researchers stressed the system is intended as auxiliary reference support under physician oversight, not a replacement for clinical decision-making.

A one-month demonstration at three primary-care hospitals offered preliminary usability data. Questionnaires from 35 physicians showed 77.14 percent agreed the system could serve as a daily clinical reference, 85.72 percent said the knowledge module supported their understanding of TCM theory, and 71.43 percent considered the prescription outputs suitable as auxiliary references. Yet perceptions varied: only 57.14 percent reported the system did not substantially increase outpatient workload, and 88.57 percent said further optimization was necessary before large-scale deployment. The authors acknowledged these were self-reported perceptions without a control group or objective outcome measures.

The study’s limitations are considerable and the authors laid them out plainly. The knowledge evaluation was grounded in textbooks that were part of the training corpus, so generalizability to unseen sources remains untested. The experience component reflects a single practitioner and a single disease setting, with no patient follow-up, treatment-response data or objective clinical outcomes. Safety assessment covered only predefined rubric items, omitting patient-specific contraindications, dose appropriateness, herb-drug interactions and adverse events. The primary-care assessment involved 35 physicians over one month relying on subjective reports. The team called for multi-practitioner data, external validation datasets, longer multicenter prospective studies with objective outcomes, and strengthened contraindication and interaction screening before any broader implementation. Within those bounds, the work offers a concrete architectural argument: when modeling both what a medical tradition knows and what one expert does, keeping the two apart may matter more than merging them.

Subject of Research: A modular large language model system for Traditional Chinese Medicine that separately models general domain knowledge and practitioner-specific clinical experience

Article Title: DFGLM-TCM: an integrated knowledge-and experience-driven large language model system for Traditional Chinese Medicine practice

Article References: Zhong, R., Liu, D., Dai, Y., Li, Z., Song, Z., Li, Z., Zhang, L., Chen, S., Su, Y., Sun, Y., Zhang, W., Jia, P., Jiang, M., Gao, L., & Hu, K. (2026). DFGLM-TCM: an integrated knowledge-and experience-driven large language model system for Traditional Chinese Medicine practice. Chinese Medicine, 21(1), Article 215. https://doi.org/10.1186/s13020-026-01512-y

Image Credits: AI Generated

DOI: 10.1186/s13020-026-01512-y

Keywords: Traditional Chinese Medicine, large language models, retrieval-augmented generation, knowledge graph, prescription consistency, clinical decision support, fine-tuning, primary care, pulmonary nodules, expert evaluation, digital health, DFGLM-TCM