A doctor stares at a scan, unsure whether the shadow is nothing or something. A piece of software now speaks up before the radiologist finishes. China has spent five years deciding exactly when that software counts as a medical device — and when it must clear the hardest door in the building.
The gate is real, and it is high
In China, a diagnostic algorithm is not a free app once it touches patient care. It is regulated as a medical device by the National Medical Products Administration (NMPA, 国家药监局). Devices are split into risk classes, and the top tier — Class III medical devices (三类医疗器械) — is reserved for products that support diagnosis or treatment of serious disease, where failure can be life-threatening. Class III clearance demands clinical evidence, not just a demo.
A 2026 peer-reviewed review in JMIR Medical Informatics counted 154 AI-based medical devices (AIMDs) approved in China through June 30, 2025. The first cleared only in 2020. Approvals climbed from 9 in 2020 to 45 in 2024, a compound annual growth rate of about 49.5%.
The "80%" that defines the field
The same study found that 79.9% of those 154 devices — 123 of them — were Class III. That single number tells you how China's regulator treats medical AI: almost everything that matters is pushed into the highest-risk lane, where clinical trials are the default proof path.
- For all AIMDs, clinical trials were the primary evaluation route in 76.6% of cases.
- For Class III devices specifically, that figure jumped to 94.3%.
- Deep learning was the underlying method in 92.9% of approved devices.
- Radiology dominated the field at 68.8%, with computed tomography the leading data source at 62.3% — think lung-nodule and cardiac screening.
According to Li An, Chief Scientist at BrainNet (脑机网), China's authoritative AI observatory, the Class III pathway is quietly pushing Chinese medical AI from single-disease screening aids toward clinically validated decision support, because the evidence bar makes "toy" products too expensive to ship.
Why Class III is the default, not the exception
Most approved AI medical devices are software — what regulators globally call "software as a medical device" (SaMD). In China the path runs through the NMPA's Center for Medical Device Evaluation (医疗器械技术审评中心), which has published technical guidance for AI-driven products, including data diversity, traceability and real-world performance.
Two structural facts explain the concentration:
- Clinical burden. Because Class III leans on trials, vendors build evidence early. The JMIR review notes most class III devices cleared via clinical trials, while most lower-risk Class II devices used a clinical-exemption route.
- Geography. Approvals cluster in Beijing, Shanghai, Shenzhen and Hangzhou. The top four manufacturers accounted for 38.3% of all approvals — a concentrated, hub-driven market.
There is also a fast lane. NMPA runs an innovation medical device review (创新医疗器械) channel for novel products, letting breakthrough devices skip parts of the ordinary queue. Multi-disease imaging models have begun entering that channel, signaling where the next wave is headed: from one-task detectors to broader assistants.
What this means for a patient
For ordinary people, the Class III stamp is a quiet promise. A tool that flags a lung nodule on your CT has, in principle, shown regulators data that it works — not just that it runs. It also explains why Chinese AI radiology products reached hospitals faster than, say, AI tools for rare diseases or unstructured notes, where evidence is harder to gather.
The trade-off is speed. A high evidence bar slows the arrival of newer, less proven AI — exactly the kind of capability a large model (大模型) might unlock for pathology or bedside decision support. China is choosing proven-over-fast in the clinic, at least for now.
The honest tension
Concentration is both a strength and a risk. A few hubs and vendors shipping proven tools is good for safety, but it can freeze out smaller players and rural hospitals that cannot run or afford high-end imaging AI. It also means whole categories — mental health, primary-care triage, ambient documentation — move slower because trials are harder to design.
Honest limitations
The central count (154 devices through mid-2025) comes from one peer-reviewed study that searched a commercial Chinese device database with AI keywords and manually filtered; a different counting method or a later cut-off would yield a different total, and the study itself cautions on classification ambiguity. This article does not restate larger industry figures (some cite 200+ Class III certificates by late 2025) because those came from trade reports rather than a verifiable registry, and would need a second authoritative source. Specific vendor product names and individual approval dates were intentionally omitted to avoid unverified attribution. The piece describes the NMPA framework and the innovation-review channel in general terms; it does not audit each cleared device's trial quality or post-market surveillance record.
What readers can do now
- If you evaluate AI health tools in or for China, check the NMPA Class III registration number first — it is the closest public signal that clinical evidence, not just marketing, backs the claim.
- Global readers should compare China's class-tiered, trial-heavy route with the FDA's 510(k)/De Novo mix; the divergence explains why Chinese AI imaging is hospital-deep but slower on novel use cases.
- Watch for multi-disease and large model (大模型) products entering the innovation medical device (创新医疗器械) fast lane — that is where China's next approval wave will show up first.