China’s New AI Medical Device Draft Guidelines: What the 2026 NMPA Drafts Mean for Manufacturers

China is updating how it regulates artificial intelligence in healthcare. In September 2026, the Center for Medical Device Evaluation (CMDE) of the National Medical Products Administration (NMPA) opened public consultation on two drafts: a revised guideline on AI medical devices and a new guideline on clinical evaluation of AI multi-disease auxiliary decision-making devices. If you build or sell a China AI medical device, these drafts will shape your next submission. 

What Did CMDE Release? 

• Draft Registration Review Guideline for AI Medical Devices (2026 Revision): published 14 September 2026, comments due 12 October 2026. 
• Draft Clinical Evaluation Registration Review Guideline for AI Multi-Disease Auxiliary Decision-Making Devices: released 15 September 2026, comments due 14 October 2026. 

Both comment windows have closed. CMDE has not announced dates for the final versions, so treat each CMDE draft guideline as a strong signal of where review expectations are heading. 

Draft Guideline on AI Medical Devices (2026 Revision) 

This core NMPA AI guideline revises the version first issued in 2022 and applies to Class II and Class III standalone AI software, as well as devices that contains AI software components. It keeps the lifecycle approach covering requirements, data collection, algorithm design, verification and validation, and update control. 

• Higher risk means stronger proof. If the software helps doctors decide on diagnosis, triage, or treatment, it needs more clinical evidence than software that only saves time or improves images. 
• Data rules  
  – Adding extra data to the training and tuning sets is allowed. 
  – Adding extra data to the test set is generally not allowed, except for adversarial testing. 
  – Any outside evaluation database should be trusted, real-world, and varied. 
• Update rules
  – Check every algorithm or data update to see if it is a major change. 
  – Keep a rollback option so you can go back to the earlier version. 
• New areas covered
  – Large models and medical agents: manage risks like hallucination and bias and add warnings to the manual. 
  – Products using several data types, and products covering several diseases. 
  – Generative adversarial networks (GANs). 
  – Creating an image in one modality from another (cross-modal image synthesis). 
  – Continuous learning, where the software keeps learning after release. 

Draft Guideline on Clinical Evaluation of AI Multi-Disease Devices 

A multi-disease auxiliary decision-making device gives decision support for several diseases within one product, for example, software that flags more than one condition from the same imaging study. This creates questions a single-disease product never faces. How do the diseases relate? Can one find a mask or mimic another? Does performance hold for every disease, not just on average? 

• Risk and intended use set the rules. Software that helps doctors make decisions (diagnosis, triage, treatment) needs stronger clinical evidence than software that only speeds up work or improves image quality. 
• Clear rules on data. You can add more data to the training and tuning sets, but you generally cannot do this for the test set. The only exception is adversarial testing. Any outside evaluation database you use should be trusted, real-world, and varied. 
• Controlled updates. Every algorithm or data update must be checked to see whether it counts as a major change. The guideline also recommends a rollback option, so you can return to the earlier version if needed. 
• New topics covered. The draft now includes:
  – Large models and medical agents (you must manage risks like hallucination and bias, and put warnings in the manual) 
  – Products using multiple data types, and products covering multiple diseases 
  – Generative adversarial networks (GANs) 
  – Creating an image in one modality from another (cross-modal image synthesis) 
  – Continuous learning, where the software keeps learning after release 

The new draft is CMDE’s dedicated answer for AI clinical evaluation of these products. It sits alongside the 2023 guideline on clinical evaluation of AI-assisted detection software and the June 2026 draft on AI-assisted diagnosis devices, which together form China’s clinical evidence framework for multi disease AI software. 

What Should Manufacturers Do Now? 

• Run a gap assessment of your technical file against both drafts. 
• Define your intended use clearly, since auxiliary decision-making claims drive clinical evidence needs. 
• For multi-disease products, plan evidence for each disease and document how diseases interact. 
• Keep development and trial data strictly separate, and strengthen algorithm change management 
• Track CMDE for the final texts and adjust your NMPA device registration timeline. 

Getting these right early matters. AI medical device regulation in China is moving toward deeper scrutiny of data, algorithms and clinical proof, and late changes are expensive. 

Quick FAQs 

• Are these guidelines mandatory?
  – Not yet. Both are consultation drafts. CMDE guidelines are technical guidance for applicants and reviewers rather than binding law, but reviewers use them to judge submissions. 
• Do they apply to foreign manufacturers?
  – Yes, any AI device seeking NMPA registration in China, imported or domestic, will be reviewed against the final versions. 

How NexorTest Can Help

NexorTest supports AI and software device makers across the full life-cycle. Our team offers NMPA medical device registration consulting, Software as a Medical Device (SaMD) regulatory consulting, cybersecurity and software lifecycle testing, and clinical study support. Talk to NexorTest to prepare your AI device for China’s evolving requirements.