FDA Issues Discussion Paper and Request for Comments on Regulatory Considerations for Generative Artificial Intelligence-Enabled Medical Devices

FDA Issues Discussion Paper and Request for Comments on Regulatory Considerations for Generative Artificial Intelligence-Enabled Medical Devices

On April 3, 2025, the U.S. Food and Drug Administration (FDA) Center for Devices and Radiological Health (CDRH) released a discussion paper titled Considerations for the Regulation of Generative Artificial Intelligence-Enabled Medical Devices: Discussion Paper and Request for Comments and opened a public comment period. The paper outlines an initial regulatory framework focused on total product lifecycle oversight, transparency, performance monitoring, and human-machine teaming. It directly affects artificial intelligence medical device companies planning to enter the U.S. market and touches on multiple compliance areas, including registration pathways, technical documentation, clinical evaluation, labeling, and post-market surveillance.

This article is based on public information from FDA CDRH News & Updates. Source: FDA CDRH News & Updates. Specific requirements are subject to the latest official publications of the regulatory authority.

Full Translated Core Content

On April 3, 2025, the U.S. Food and Drug Administration (FDA) Center for Devices and Radiological Health (CDRH) released a discussion paper titled Considerations for the Regulation of Generative Artificial Intelligence-Enabled Medical Devices: Discussion Paper and Request for Comments and opened a public comment period. The FDA stated that the use of generative artificial intelligence in medical devices is developing rapidly, and that its unique adaptive and generative capabilities present new regulatory challenges. Through this paper, the FDA aims to engage in early dialogue with stakeholders to jointly develop an appropriate regulatory framework.

The FDA noted that generative artificial intelligence-enabled medical devices may include devices used for diagnosis, treatment decision support, imaging analysis, document generation, and other functions. Unlike traditional rule-based artificial intelligence, generative models can produce new content, respond to open-ended inputs, and may continue to learn or update as they are used. These characteristics mean that existing regulatory pathways such as 510(k), De Novo, and PMA may need to consider additional technical details where applicable.

In the discussion paper, the FDA proposed a Total Product Lifecycle (TPLC) regulatory approach. For premarket evaluation, the FDA recommends that companies focus on the representativeness of training data, choices in model design, clear definition of intended use, and generalizability across different clinical environments. For continually learning or adaptive models, the FDA is particularly concerned about performance drift risk and stability after retraining.

The paper focuses on transparency. The FDA believes that medical device labeling should clearly disclose to users that the device uses generative artificial intelligence and should provide information on model capabilities, limitations, validation status, and expected user training needs. Transparency not only helps users use the device appropriately but also serves as an important foundation for FDA post-market surveillance.

Regarding post-market surveillance, the FDA recommends that companies establish performance monitoring plans proportionate to device risk, including real-world data collection, predefined performance metrics and thresholds, and corrective actions when performance degradation is identified. For updatable generative artificial intelligence devices, the FDA proposes distinguishing between device updates and changes that require a new submission (change protocol) and discusses the possibility of Predetermined Change Control Plans.

In addition, the paper raises the issue of human-machine teaming, emphasizing that outputs from generative artificial intelligence should be reviewed and confirmed by appropriately qualified healthcare professionals. Device design should support users' critical evaluation rather than encourage over-reliance on automated outputs. The FDA also acknowledges that many generative artificial intelligence-enabled medical devices are still at an early stage and that some questions do not yet have consensus answers. It therefore welcomes feedback from industry, academia, patients, and healthcare providers.

The FDA invited the public to submit comments on the discussion paper by May 20, 2025. The paper does not constitute formal guidance and does not represent a final FDA regulatory decision. It is intended to provide a basis for future guidance development and regulatory policy evolution. The FDA stated that it will consider whether to issue formal guidance based on the feedback received and will continue to advance the development of the regulatory framework for artificial intelligence-enabled medical devices.

Aimeili Regulatory Interpretation

Aimeili Regulatory Affairs believes that although this discussion paper does not have the legal effect of formal guidance, it clearly signals the FDA's core direction for regulating generative artificial intelligence-enabled medical devices. For Chinese medical device companies that are developing or planning to bring generative artificial intelligence products to the U.S. market, the total product lifecycle regulatory approach proposed by the FDA must now be incorporated into product planning, rather than addressed only after formal guidance is issued.

From a registration pathway perspective, companies should determine the product's intended use and degree of automated decision-making as early as possible to judge whether the product is an artificial intelligence-assisted decision-making device or a traditional diagnostic device. This will directly affect whether the 510(k), De Novo, or PMA pathway is appropriate. The FDA emphasizes in the paper that intended use should be "clearly defined and not overly broad." Companies should avoid using generalized clinical claims to mask the uncertainty of generative models. We recommend that intended use statements clearly limit the patient population, use scenario, input data type, and the role of output results (assistive versus automated).

In technical documentation and clinical evaluation, the FDA is likely to focus on the representativeness of training data, model calibration, and robustness validation for generative artificial intelligence. Companies should prepare complete data management documentation, including data sources, collection methods, labeling rules, data distribution, and bias analysis. They should also conduct sufficient robustness testing of device performance across different populations, device endpoints, and environments. Clinical evaluation is not limited to retrospective validation; the FDA is likely to require prospective data to demonstrate the clinical acceptability of generative artificial intelligence outputs, especially in high-risk decision-making scenarios.

From a labeling and quality system perspective, transparency requirements mean that instructions for use, user interfaces, and training materials all need dedicated design. Companies need to establish a labeling management process that continuously updates disclosures of model capabilities and limitations. At the same time, if a product has continual learning or model update functions, a clear change control procedure must be established within the quality management system to distinguish "permitted updates" from "significant changes requiring a new submission." Companies should also consider whether to adopt the FDA-recommended predetermined change control approach to manage anticipated changes.

Post-market maintenance is another critical area. The FDA expects companies to establish real-world performance monitoring plans, including key performance indicators, monitoring frequency, anomaly trigger thresholds, and response mechanisms. Chinese companies often underinvest in post-market surveillance when going global. We recommend establishing or entrusting a local agent in the U.S. market to create 24/7 complaint and adverse event feedback channels and to periodically analyze data for performance drift. Finally, companies should actively use the window period to submit feedback to the FDA or participate in public meetings to influence future guidance direction, while closely monitoring international harmonization progress to avoid duplicative efforts across U.S., Chinese, and European artificial intelligence regulatory requirements.

Business Impact

For manufacturers, distributors, contract research organizations, and regulatory affairs teams, the discussion paper is an early warning of expected U.S. evidence expectations. Products using generative artificial intelligence will face greater scrutiny of intended use, data governance, model robustness, transparency, change control, and post-market performance monitoring. Companies should reassess U.S. market entry strategies, registration pathways, and resource allocation. A product that may appear to be a conventional software device could require deeper clinical evidence, stronger cybersecurity and software lifecycle documentation, and a more mature quality management system. Early gap analysis can reduce the risk of submission delays, clinical evidence deficiencies, and post-market compliance issues.

FAQ

What is the most difficult part of overseas registration for artificial intelligence-enabled medical devices?

The difficulty is usually not simply completing an application form. It typically lies in risk classification, algorithm evidence, clinical evaluation, software lifecycle management, and post-market change management. Companies must demonstrate that the device's intended use, performance, and risk controls are appropriately supported by data and documentation.

Can an artificial intelligence-enabled medical device be submitted using ordinary software documentation?

It is not advisable to simply apply ordinary software documentation. Companies should first confirm whether the product qualifies as Software as a Medical Device (SaMD), whether it has artificial intelligence-assisted decision-making attributes, and whether it affects diagnosis or treatment. The registration documentation should then be designed accordingly.

Content Review and Applicability Boundaries

Content author: Aimeili Regulatory Editorial Team. Professional review: Aimeili Medical Device International Registration Project Team. Source principle: priority is given to official regulatory authorities, international organizations, standards organizations, and publicly available regulatory materials; industry media and project experience are used only as supplementary judgment. Applicability boundary: This article is intended for preliminary understanding, documentation preparation, and project planning. It does not replace the formal requirements, testing conclusions, or legal opinions of the target country's regulatory authority.

Contact Aimeili

Need to assess a registration pathway? Please provide the product name, intended use, target countries, and existing documentation basis. Aimeili can assist in determining the registration pathway, documentation gaps, and compliance risks.