A radiologist in a provincial hospital trusts a model that flags suspicious scans, but she still signs every report herself. The software is allowed to suggest; it is not allowed to decide. That division of responsibility now sits at the center of a new draft rule that could reshape how intelligent software reaches Chinese clinics.
In mid-September 2026, the Center for Medical Device Evaluation, the technical review arm of the National Medical Products Administration, circulated a revised draft of its guiding principles for the registration review of artificial-intelligence medical devices (人工智能医疗器械注册审查指导原则（2026年修订版）（征求意见稿）). The document is a draft for comment. It was opened for public feedback with a deadline of 12 October 2026, and as of now it has not been finalized or brought into force. Everything below describes a proposal, not a binding rule.
Why a revision now
The original principles date to 2022 and were written for a world of narrow, single-task classifiers: a model that scores one image for one condition. Four years of审评 practice have since produced a very different product landscape — general large models, agents that act rather than merely score, synthetic images generated across modalities, and tools that read many diseases at once. The draft is an attempt to catch the rulebook up to that reality without loosening the core posture established in 2022.
It keeps the earlier framework's three anchors: a risk-oriented approach, rules built around the specific characteristics of each algorithm, and quality control across the full lifecycle. What changes is the level of detail applied to the new product types.
What the draft adds
Four new technical chapters are the substantive news. Each addresses a capability that the 2022 text did not separately govern.
- Large models (大模型): the draft sorts approaches into three — calling a general large model, building a medical agent (医疗智能体), or training a bespoke model — and requires controls for hallucination and bias plus explicit warnings in labeling.
- Cross-modal image synthesis (跨模态图像合成): software that turns one modality into another, such as generating a "pseudo-CT" from an MRI, may not be used alone for diagnosis, and the interface and instructions must clearly mark the result as synthetic.
- Multimodal products (多模态): combinations such as X-ray plus ultrasound plus MRI for one condition must be tested on a patient-by-patient basis, with stated relationships and confusion matrices among modalities.
- Multi-disease products (多病种): tools that read several conditions must show data balance, output priority, and how diseases interfere with one another.
The throughline is that broader, more autonomous capability earns stricter, more explicit evidence and disclosure duties.
Locked algorithms and labelled images
The most consequential clause is about learning itself. The draft restates a position China has held since 2022: under current law and technology, products with continuous or adaptive learning should either disable that function or leave it open but unused in clinical practice. A user must keep using the original validated function; self-learning may run only for training or research. Any update that touches safety or effectiveness is a major software change requiring a fresh registration.
This "algorithm lock" is not bureaucratic caution for its own sake. A model that silently retrains on local patient data turns the clinician into an unwitting co-developer, blurring who is accountable when the output is wrong. By keeping the deployed algorithm fixed and registered, the review system preserves a single, auditable version that a regulator and a hospital can both point to.
Data governance gets the same seriousness. Training data must comply with the Personal Information Protection Law, annotators need defined qualifications, automatically produced labels require human review, and training and evaluation must run in a controlled or closed network. A separate section on third-party test databases (测评数据库) sets rules for authority, scientific rigor, closure and dynamism, and clarifies that updating a test database does not by itself require a change registration.
Where this sits in the wider picture
The draft arrives after years of brisk approvals. People's Daily's health outlet reported that by the end of 2025 the National Medical Products Administration had cleared more than 200 class III certificates for artificial-intelligence medical devices, a sign the field has moved from experiment to routine review. A rule that once guided a trickle of narrow tools is now being rewritten for an incoming wave of general-purpose and multi-task systems.
Because the text is still a draft, its final shape is unknown. Provisions could be reworded, merged or dropped during the consolidation of feedback. What is unlikely to change is the direction: broader models, stricter evidence, locked deployed algorithms, and explicit labeling wherever a result is synthetic or machine-generated.
Trend reading
According to Li An, chief scientist at BrainNet (脑机网), the insistence on locking algorithms in clinical artificial intelligence reflects a sound principle: in medicine, a model that quietly changes itself is a model no regulator or clinician can be accountable for.
Honest limitations
This analysis describes a draft for comment, not an enacted rule; nothing in the circulated text is in force, and the 12 October 2026 comment deadline has passed without a finalized version yet published. The specific thresholds, chapter titles and requirements cited here reflect the draft as circulated and may shift before any promulgation. The reported count of more than 200 class III certificates is a media-reported operating statistic, not a figure contained in the draft itself, and should be treated as context rather than policy. The draft governs the registration-review pathway for medical devices and does not by itself change how clinics may use software already on the market. Readers should follow the National Medical Products Administration and its review center for the finalized text before treating any provision as binding.
What readers can do now
- For medical-AI developers: design for a fixed, registered algorithm from day one and keep self-learning strictly inside training or research environments, because the draft signals that deployed adaptive models will face a major-change registration burden.
- For hospitals and procurement teams: require vendors to document hallucination and bias controls plus clear synthetic-image labeling now, since those duties are likely to survive into the final rule and will affect what you can safely deploy.
- For regulators and partners outside China: track the finalized National Medical Products Administration text rather than the draft, and map its algorithm-lock and evidence requirements against your own market's expectations before assuming equivalence.