AI models learn to read ancient Chinese medicine texts and support clinical reasoning

Large language models are being drawn into one of medicine’s most demanding knowledge domains: traditional Chinese medicine, a system built on multi-source data, implicit knowledge representation, and syndrome-based diagnostic and therapeutic reasoning. A review published in the journal Chinese Medicine on 8 October 2026 organizes the scattered research landscape into a single workflow, tracing how digitized classical texts, electronic medical records, and multimodal diagnostic data are being converted into training corpora for TCM-specific language models, and what still stands between these systems and routine clinical use.

The review, led by Chenyue Li and colleagues with senior authorship including Yuqun Zeng and Jianlan Zheng, frames the field around a pipeline that runs from data construction through model development, evaluation, and application. Its central argument is that no single component is sufficient on its own: a closed-loop ecosystem integrating data, models, evaluation, and safety is presented as the prerequisite for cautious and appropriate use of large language models in TCM research and practice. The authors describe their work as a synthesis of published literature rather than new experimental findings, and they emphasize that existing studies remain fragmented.

The data foundation is described in detail. The review categorizes available TCM data into five primary types: ancient TCM classics, modern TCM knowledge, electronic medical records, scientific research literature, and multimodal diagnostic information. Ancient classics, written or printed predominantly in classical Chinese over several millennia, are characterized by concise syntax, implicit semantics, and strong contextual dependency, and they exist as paper manuscripts, woodblock prints, digitized versions, and electronic scans that include hand-drawn herb illustrations and acupuncture diagrams. The Zhonghua Yidian database is cited as an example of a comprehensive digital repository, containing millions of structurally annotated entries across approximately 5000 classical texts.

Electronic medical records carry particular weight because they capture authentic real-world clinical data from routine TCM practice, including individual medical cases and longitudinal patient cohorts. According to the review, EMR systems now cover 92% of tertiary TCM hospitals in China and accumulate over 5 billion structured TCM diagnosis and treatment records annually. These records preserve the sequential clinical reasoning process, from symptom presentation and syndrome differentiation to prescription formulation and dynamic adjustment over time, which the authors identify as essential raw material for modeling TCM decision-making. Multimodal repositories of tongue images, facial complexion data, pulse waveforms, and phonation signals, often paired with symptom annotations and syndrome labels, extend this foundation beyond text.

Turning heterogeneous raw material into machine-readable corpora requires a multi-stage pipeline. Format normalization and optical character recognition customized for classical Chinese and medical terminology come first, but the review stresses that rare characters, variant glyphs, and archaic expressions mean OCR-generated texts frequently contain errors, making manual proofreading and expert curation by trained TCM scholars indispensable. Cleaning procedures then correct typographical and OCR-induced errors, normalize terminology and measurement units, remove contradictory records, and deduplicate overlapping content that arises from multiple editions of classical texts. Entity extraction follows, progressing from early BERT and ERNIE-based named entity recognition to architectures such as BERT-BiLSTM-CRF and frameworks like TCMERE, which combine domain-adaptive models with graph convolutional networks to handle entity nesting and relationship overlap. Extracted entities are then mapped to standardized terminologies and structured ontologies such as GFO-TCM and HerbKG to resolve the semantic ambiguity produced by thousands of years of regional variation in naming.

Training data are grouped into five functional categories: large textual corpora for pre-training, question-answering pairs, dialogue data, image or audio-text pairs, and instruction-following data. The TCMCorpus dataset, with 3.4 billion tokens drawn from textbooks and academic literature, exemplifies the pre-training layer, while the Huatuo-26M dataset aggregates 26 million question-answer pairs from diverse medical knowledge bases. Dialogue corpora such as the BianQue Corpus, with over 2.4 million conversations, and CMtMedQA, with approximately 70,000 conversations, are highlighted as particularly important for modeling the iterative questioning central to TCM syndrome differentiation. The review notes a practical complication: clinical responses extracted directly from real doctor-patient conversations are typically concise, context-dependent, and stylistically heterogeneous, which limits their usefulness for instruction tuning, so self-instruction and LLM-assisted construction strategies constrained by templates, ontologies, or knowledge graphs and verified by experts are increasingly adopted.

On the modeling side, the review traces an evolution from early pre-trained models like BERT and ERNIE, used for entity recognition and clinical note classification, through general medical LLMs such as MedGPT, to dedicated TCM-specific models cataloged in its tables. Four structural features define current TCM-LLMs. Dual-language pre-training strategies and models like SiKuBERT address the linguistic gap between ancient texts and modern language. Knowledge-augmented generation, including retrieval-augmented generation connected to clinical guidelines, pharmacological databases, and classical texts, grounds outputs in retrieved documents rather than learned patterns alone, and integration with structured knowledge graphs enables multi-hop reasoning over symptom-herb associations. Advanced models are also being trained to explicitly generate the reasoning chain from symptoms through syndrome differentiation and pathogenesis analysis to formula and herb selection, outputting step-by-step explanations rather than bare recommendations. Finally, multimodal models such as ShizhenGPT integrate visual understanding with textual knowledge, aiming to approximate the four diagnostics of inspection, auscultation and olfaction, inquiry, and palpation.

Training proceeds through pre-training on massive mixed-domain corpora, supervised fine-tuning on curated instruction-response datasets covering question answering, syndrome differentiation, and formula interpretation, and preference alignment. Reinforcement learning from human feedback, in which expert practitioners rank model-generated responses, remains widely adopted, while direct preference optimization and Kahneman-Tversky optimization have emerged as alternatives that bypass explicit reward modeling, an advantage the review links to the scarcity and cost of expert-labeled preference data in TCM. Multimodal alignment adds a further stage, using adapter networks to project image and signal features into the language model’s shared embedding space. A separate section on computing requirements argues that, relative to general-purpose foundation models, TCM-LLMs typically require more moderate resources: existing systems fall into three tiers, from lightweight 6-7B parameter models trained on 4 to 8 GPUs to high-end 14B-32B architectures requiring 32 or more GPUs, with training times of roughly 12 to 75 hours and near-universal adoption of parameter-efficient methods such as LoRA and QLoRA.

Evaluation receives extended treatment because general-purpose benchmarks are deemed insufficient for medical accuracy, diagnostic reasoning, and safety. The review proposes five dimensions: knowledge accuracy, covering factual consistency and entity correctness for herbs, acupoints, and formulas; reasoning ability, assessing the plausibility of the diagnostic chain from symptom to syndrome to treatment principle; generation quality, including fluency and terminological consistency; clinical usability, measured through expert physician ratings and retrospective case-control comparisons in a decision-support context rather than autonomous decision-making; and safety assessment, which quantifies medically unsafe hallucinations and evaluates privacy preservation and sensitivity to humanistic aspects of care. Specialized benchmark datasets derived from professional qualification exams, classical texts, and clinical case studies provide the standardized basis for these measurements.

Application scenarios described in the literature span medical consultation support, where models like HuatuoGPT ask clarifying questions mimicking the inquiry component of the four diagnostics; clinical diagnosis and treatment, where LLMs recommend classical herbal formulas, generate novel herb combinations, and suggest formula modifications based on a patient’s symptoms, constitution, and comorbidities; TCM examinations and professional assessment; knowledge-oriented question answering over classical and modern sources; and scientific research assistance, including literature retrieval, hypothesis generation, and knowledge graph construction. The review is explicit about the limits. It identifies limited interpretability, safety concerns, and the lack of standardized evaluation frameworks as key challenges, notes that recent regional regulations in China explicitly prohibit AI systems from independently generating medical prescriptions without physician oversight, and cautions that although retrieval-augmented generation has improved factual grounding, the risk of model hallucination remains. Pulse waveforms are sensitive to sensor architecture, applied pressure, and calibration, and electronic-nose outputs vary with sampling protocols and ambient conditions, so routine clinical deployment of multimodal systems remains constrained by the limited availability of standardized, clinically validated acquisition hardware. The authors conclude that LLMs are positioned not to replace the human practitioner but to become a partner within a closed-loop ecosystem, provided that future models justify each step of the disease-symptom-syndrome-formula-herb-effect chain with evidence and that every generated recommendation remains traceable to its source.

Subject of Research: A review of how large language models are being developed and applied to traditional Chinese medicine knowledge and clinical practice

Article Title: Large language models linking traditional Chinese medicine knowledge and clinical practice

Article References: Li, C., Zhang, X., Zhu, L., Zhu, B., Chu, C., Li, X., Yao, Y., Zeng, Y., & Zheng, J. (2026). Large language models linking traditional Chinese medicine knowledge and clinical practice. Chinese Medicine, 21(1), Article 227. https://doi.org/10.1186/s13020-026-01489-8

Image Credits: AI Generated

DOI: 10.1186/s13020-026-01489-8

Keywords: large language models, traditional Chinese medicine, artificial intelligence, classical texts, electronic medical records, multimodal diagnosis, syndrome differentiation, retrieval-augmented generation, knowledge graphs, model evaluation, hallucination, clinical decision support