Guiding Principles for Registration Review of AI Medical Devices (2026 Revision)

On September 14, 2026, the Center for Medical Device Evaluation of the National Medical Products Administration issued a notice soliciting public comments on the Guiding Principles for Registration Review of Artificial Intelligence Medical Devices (2026 Revision) (Draft for Comments). The deadline for comments is October 12, 2026.

The public consultation includes two attachments: Attachment 1 is the Guiding Principles for Registration Review of Artificial Intelligence Medical Devices (2026 Revision) (Draft for Comments), and Attachment 2 is the comment feedback form. The guidance aims to guide applicants in establishing the lifecycle process for AI medical devices and preparing registration submissions, while standardizing technical review requirements and providing reference for system inspections of AI medical devices and quality management software. It is an important part of the digital health guidance system and aligns with documents on medical device software, medical device cybersecurity, mobile medical devices, medical device usability engineering, and the appendix for standalone software under the production quality management specification.

In terms of scope, the document applies to registration submissions for AI medical devices, including Class II and Class III AI standalone software and medical devices containing AI software components, including in vitro diagnostic medical devices. It applies to registration submissions for self-developed software; off-the-shelf software components are handled by reference, and external software environments are not covered. It can also serve as a reference for system inspections of AI medical devices. If quality management software uses AI to achieve its functions or intended use, it may also refer to this guidance.

On key concepts, an AI medical device refers to a medical device that is based on “medical device data” and uses AI technology to achieve its intended medical purpose. Types include AI standalone software and AI software components; from the use perspective, assisted decision-making and non-assisted decision-making; from the function perspective, processing, control, and safety functions; from the algorithm perspective, supervised and unsupervised learning, model-based and data-based algorithms, and white-box and black-box algorithms; from the maturity perspective, mature and novel types. Algorithm updates fall within software updates and are divided into algorithm-driven updates and data-driven updates. Updates affecting safety or effectiveness are generally major software updates and require a change registration; updates not affecting safety or effectiveness are minor software updates controlled through the quality management system. Software version naming rules should generally cover both algorithm-driven and data-driven updates.

The basic principles include: algorithm characteristics-based evaluation, risk orientation, and full lifecycle quality control. The lifecycle process mainly includes requirements analysis, data collection, algorithm design, verification and validation, and update control. Data collection emphasizes compliant, sufficient, and diverse sources; scientific and reasonable data distribution; and adequate, effective, and accurate quality control, covering data acquisition, data curation, data annotation, and dataset construction. Algorithm design focuses on algorithm selection, training, performance evaluation, and interpretability of black-box algorithms. Verification and validation include software verification, software validation, and clinical evaluation. Update control requires appropriate verification and validation activities for both algorithm-driven and data-driven updates, and consideration of rollback mechanisms.

Technical considerations cover cybersecurity and data security, mobile computing and cloud computing, usability engineering, stress testing and adversarial testing, third-party databases and evaluation databases, white-box algorithms, ensemble learning, transfer learning, reinforcement learning, federated learning, generative adversarial networks, cross-modal image synthesis, multimodal products, multi-disease products, large models, continuous/adaptive learning, AI algorithm programming frameworks, and AI chips. For large models, the document discusses three approaches: calling a general large model through an interface, developing a medical agent based on a general large model, and self-developing a model using a large-model technical route. For continuous/adaptive learning, it states that under the current legal framework and technical level, the self-learning function should be turned off, or if enabled, must not be put into use and may only be used for algorithm training or medical research.

For registration submissions, the algorithm research report applies to the initial and subsequent release of an AI algorithm or algorithm combination, including basic algorithm information, algorithm risk management, algorithm requirements specification, data quality control, algorithm training, algorithm verification and validation, algorithm traceability analysis, and conclusions. The algorithm update research report applies only to subsequent releases. For products with moderate or serious software safety levels, novel types generally require an algorithm research report on a per-algorithm basis; mature types may only need basic algorithm information. Minor-level products may only need basic algorithm information. For change registration, an algorithm update research report or algorithm research report should be submitted according to the algorithm update. The instructions for use should specify product limitations and necessary warnings based on the comprehensive algorithm performance evaluation.