NeuroAI NEUROAINEUROAI.SITE
China AI, reported clearly
1,300+ verified stories on models, robotics, BCI and the business of AI.
Browse stories
ESC

One Beijing hospital's AI cut Chinese-medicine charting by 75% — and kept the doctor in charge

China's top hospitals are wiring large models (大模型) into Traditional Chinese Medicine (中医); the early results are real, and so are the limits.

2026-11-15 · 951 words · NeuroAI
One Beijing hospital's AI cut Chinese-medicine charting by 75% — and kept the doctor in charge

A 68-year-old man walks into an outpatient clinic in Beijing with a stomach that has ached for years. Before he reaches the doctor, he has already told his story once — not to a person, but to a phone. The screen asks whether the pain comes after meals, whether he fears cold, whether his tongue looks pale. By the time he sits down, a draft record is already waiting.

That draft is not a gimmick. At China Academy of Chinese Medical Sciences Guang'anmen Hospital (中国中医科学院广安门医院), one of the country's flagship Traditional Chinese Medicine (中医) hospitals, an in-house large model (大模型) named "Guangyi Qizhi" (广医·岐智) now generates the pre-visit record and, during the consult, drafts the case notes a physician then edits.

What the hospital actually deployed

Guang'anmen released the model in March 2025. Its data foundation is unusual for being hospital-owned: 4.07 million clinical cases and a knowledge base of 14 million TCM records, built from seven decades of the hospital's own famous clinicians. The system runs on local servers inside the hospital — a deliberate choice for patient-data security — and it is tuned for one job: supporting the doctor's reasoning, not replacing it.

The measured effect is concrete. The hospital says the model cuts the time a doctor spends writing case notes by 75%, and that the generated text matches the required clinical format 95% of the time. Since launch it reports roughly 355,700 accurate triage sessions and more than 77,800 AI pre-visit records produced. A 2.0 version, an AI "doctor" called An'an (安安), adds a two-way question mode and, the hospital says, lifts the accuracy of four-diagnosis (望闻问切) information capture by 30%.

Why Chinese medicine is a hard test

Putting a model into TCM is not the same as dropping one into radiology. Traditional Chinese Medicine (中医) rests on treatment based on syndrome differentiation (辨证论治) — reading a patient through inspection, listening, questioning and pulse-taking (the four diagnostics, 望闻问切), then reasoning about patterns rather than naming a single disease. That reasoning is subjective, experience-bound, and historically hard to standardise.

This is exactly why several provincial health authorities have pushed the same direction. At a December 2025 Henan health-industry conference, the provincial health commission showcased a batch of TCM AI tools: a meridian-assist diagnosis model reported 85% agreement with expert assessment, a massage robot said to lift clinician efficiency by 50%, and a herb-quality large model (the "Shennong Bencao" model, 神农本草) for identifying and grading raw medicinal material. Henan also received approval in 2025 to build the country's first public platform for scaling TCM AI, and it reported that across its research pipeline that year, 9 new drugs, 8 products and 5 technologies moved from lab to market, with total conversion value above ¥90 million (≈ US$12.7M / HK$98M).

The Beijing Municipal Administration of Traditional Chinese Medicine reports an even wider spread of pilots: an AI tongue-diagnosis device that it says identifies 106 body-constitution types in 10 seconds; a pulse system claimed to cut pulse-diagnosis error by 40%; and a county hospital "TCM intelligence brain" said to recognise more than 2,000 syndromes with 92% accuracy. A commercial system, Wen Zhi's "TCM Brain" (问止中医), claims its recommendations match senior physicians' decisions more than 90% of the time.

A national nudge, not a free-for-all

None of this is flying blind. In December 2024, the National Health Commission, the National Administration of Traditional Chinese Medicine and a third ministry issued a reference guide for AI use in healthcare that explicitly lists TCM clinical decision support and meridian detection as promoted scenarios. The signal is permissive but scoped: AI as assistant, clinician as decision-maker.

According to Li An, chief scientist at BrainNet (脑机网), the harder test for any medical AI is whether a clinician will actually trust its suggestion at 2 a.m., not whether the demo looks smooth — and trust is earned by showing the reasoning, not hiding it.

Honest limitations

Every number above comes from the institutions that built or promoted the tools — hospital press materials, a provincial health commission showcase, and a municipal TCM administration write-up — not from independent clinical trials. The 75% charting-time reduction and 95% format-match are operator-reported and were not, in the public materials, validated against a controlled before-and-after study with a matched group of physicians.

The harder questions are clinical, not technical. A model that drafts notes well is not the same as a model that judges a syndrome correctly, and the public reporting does not break out how often the AI's differentiation suggestion disagrees with the attending physician, or what happens when it does. TCM's evidence base itself is contested in parts of the medical mainstream, so a model trained on historical TCM records will faithfully reproduce the field's disagreements, not resolve them. And the rural "democratisation" claim — that a county clinic can now borrow a master physician's judgment — remains mostly a pilot narrative; the published materials describe showcases and hospital alliances more often than sustained year-long outcomes in understaffed counties.

What readers can do now

  1. If you use a Chinese hospital that offers an AI pre-visit (预问诊) tool, fill it in before you arrive. The hospitals that deployed these say the draft saves the repeat-telling and shortens the wait — but treat the AI's suggested department as a starting point, and confirm with the human triage desk.
  2. Ask whether the system is locally deployed. Patient-record privacy depends on it; the flagship deployments advertise on-premise servers, and that is the question worth asking at smaller clinics.
  3. Do not take an AI's syndrome or prescription suggestion as a second opinion on its own. In every published Chinese case the physician remains the signer — keep it that way for your own care.

Related coverage

Cite this article

NeuroAI (脑机网), "One Beijing hospital's AI cut Chinese-medicine charting by 75% — and kept the doctor in charge," 15 November 2026, https://neuroai.site/a/na-app-tcm-ai-china

Machine-readable: BibTeX · RIS · JSON · Markdown

More in “AI in Action” → · Back to home · Markdown version

China AI, reported clearly
1,300+ verified stories on models, robotics, BCI and the business of AI.
Browse stories