Regulatory Iteration and Upgrade: In-depth Interpretation of the Guiding Principles for the Review of Artificial Intelligence-Enabled Medical Devices (2026 Revised Draft for Public Comments)

I. Formulation History and Revision Background of the Guiding Principles

China’s regulatory framework for artificial-intelligence-enabled medical devices has evolved from technical review points to official guiding principles and subsequent revisions.

1. Early-stage Exploration: In July 2019, the Center for Medical Device Evaluation (CMDE) released the Review Points for Deep-Learning-Based Decision-Support Medical Device Software, China’s first special regulatory document for AI-powered medical software. It established the core philosophy of full-lifecycle governance and supported the review and market launch of China’s first batch of AI medical devices.
2. First Round of Public Consultation: On June 4, 2021, CMDE opened public comments on the Guiding Principles for the Review of Artificial Intelligence-Enabled Medical Devices. Building on the earlier review points, it expanded general rules covering definitions, product lifecycle, data governance, algorithms and registration submission documentation.
3. Official Release and Implementation: On March 9, 2022, the Guiding Principles for the Review of Artificial Intelligence-Enabled Medical Devices was officially published. Serving as the top-level general guidance for AI medical devices in China, it unified registration submission and review standards for Class II and Class III standalone AI software and AI software components, laying the foundational regulatory framework. Numerous imaging-based AI products completed their registration based on this document.
4. 2026 Revised Draft for Public Comments: Following the roll-out of the 2022 version, generative AI, medical large language models, medical agents, cross-modal image synthesis, multi-modal and multi-disease AI products have developed rapidly. The original guidance lacked clear regulatory provisions for these emerging technologies. Meanwhile, four years of extensive product review practices have accumulated substantial practical regulatory experience, which needs to be codified into formal guidance clauses.

On September 14, 2026, CMDE issued the Guiding Principles for the Review of Artificial Intelligence-Enabled Medical Devices (2026 Revised Draft for Public Comments) for industry-wide public feedback. Comments are due by October 12, 2026. This marks further adaptation of China’s AI medical device regulatory framework to new technological developments.

The revised draft inherits the three core principles of the 2022 edition: risk-based approach, algorithm-specific considerations, and full-lifecycle quality control. It retains core frameworks including product lifecycle management, algorithm update rules and algorithm research reports. Major additions cover dedicated technical considerations for large models, medical agents, cross-modal image synthesis, multi-modal products and multi-disease products. It also refines practical requirements for registration dossiers, instructions for use and product technical requirements, and updates referenced domestic and international standards and regulatory documents to strike a better balance between innovation enablement and risk mitigation.

II. Key Changes: 2022 Official Version vs. 2026 Revised Draft for Public Comments

III. Industry Takeaways

1. Specialized research documentation for emerging-AI products: Products built upon large models, multi-modal inputs, multi-disease outputs or cross-modal image synthesis must complete dedicated algorithm research, risk analysis and warning documentation. Registration strategies for conventional single-modal, single-disease AI solutions cannot be directly adopted.
2. Risk disclosure becomes a mandatory obligation: Risks stemming from black-box algorithms, large-model hallucinations, artifacts of synthesized images and incomplete multi-modal input data must be addressed within risk-management files, algorithm research reports and product instructions for use. Pure pursuit of algorithm performance metrics is insufficient.
3. Standardized use of evaluation datasets: If official evaluation datasets are adopted for performance testing, full dataset metadata shall be recorded in the appendix of product technical requirements. Clear usage boundaries must be maintained between public datasets and formal evaluation datasets.
4. Overall logic for algorithm updates remains consistent with the 2022 version: Distinction is maintained between algorithm-driven updates and data-driven updates. Data-driven updates are judged against statistically-significant performance differences compared with the previous registered version. Where both major and minor updates co-exist, the higher-risk classification shall apply, and major changes require change-of-registration approval.
5. Regulatory stance on continuous / adaptive learning is unchanged: Self-learning functions shall be disabled for clinical release versions. Any product iterations generated via self-learning must undergo full verification and validation. Where applicable, change-of-registration approval must be obtained before clinical deployment.

Note: This article is based on a draft for public comments. Revisions may be made in response to industry feedback prior to official release. Enterprises shall conduct registration activities in accordance with the final officially-published text.