FDA AI Guidance for Medical Devices: What to Submit

Clinician reviewing AI analysis of a brain scan on a tablet beside a submission file

On 7 January 2025, FDA issued the draft guidance Artificial Intelligence-Enabled Device Software Functions: Lifecycle Management and Marketing Submission Recommendations. FDA’s digital health guidance list still shows it as Draft. The PCCP guidance for AI-enabled device software functions is final: first issued on 4 December 2024, with the current version dated 18 August 2025.

One is final, one is a draft, and the submission plan must reflect which is which. This article covers what the draft asks for in a submission, how the PCCP works, the performance monitoring plan, and where the GMLP and transparency principles fit.

Status and scope

Document Status Date
AI-enabled device software functions: lifecycle management and marketing submission recommendations Draft, “Not for Implementation” 7 January 2025
PCCP for AI-enabled device software functions Final First issued 4 December 2024; current version 18 August 2025
Good Machine Learning Practice (GMLP) Guiding principles October 2021
Transparency for machine learning-enabled medical devices Guiding principles Builds on GMLP

Source: FDA, Guidances with Digital Health Content; FDA GMLP and transparency pages.

The draft “Contains Nonbinding Recommendations”. A different approach is allowed if it meets the applicable statutes and regulations.

Under Section II of the draft:

  • an AI-enabled device software function (AI-DSF) implements one or more AI models;

  • a “marketing submission” means a 510(k), De Novo, PMA, HDE or BLA;

  • some recommendations may also apply to an IDE.

Section IV reserves the term “validation” for confirming the intended use, not for training or tuning the model.

In our view, use the draft as a map of the submission, not a rulebook. FDA says it is non-binding, and “should” means recommended. Our article on software as a medical device covers whether software qualifies as a device in the first place.

What the draft asks for

Section What FDA recommends Source
Device Description A statement that AI is used; inputs and outputs; users; environment; workflow; degree of automation Section V
User Interface and Labeling How the device and its AI output are presented to users, and labelling content Section VI
Risk Assessment Risks of the AI-DSF Section VII
Data Management Data sources and sites, time period, limitations, quality assurance, dataset size, diversity, synthetic data, test data independent of development data Section VIII
Model Description and Development The model and how it was developed Section IX
Validation Performance evidence, including subgroup analyses Section X
Device Performance Monitoring A plan for monitoring performance after authorisation Section XI
Cybersecurity Cybersecurity considerations Section XII
Public Submission Summary Information for the public summary Section XIII

Source: draft guidance (PDF).

Appendices A to F hold supporting material, including an example model card.

The draft defines bias as “a potential tendency to produce incorrect results in a systematic, but sometimes unforeseeable way”. It asks for evidence across groups such as race, ethnicity, sex and age. Subgroups need enough statistical power only where a subgroup claim is made.

The draft refers to sections of Part 820 as they stood before the amended Part 820 (the QMSR) took effect on 2 February 2026. Those references should not be read as the current text.

In our view, data management and subgroup choices are made when the data are collected, not when the submission is written.

The PCCP

Section 515C of the FD&C Act, added by section 3308 of the Food and Drug Omnibus Reform Act of 2022 (enacted 29 December 2022), allows FDA to authorise a predetermined change control plan. According to the PCCP guidance (PDF), it applies to 510(k), De Novo and PMA submissions and has three components:

  • Description of Modifications: the planned changes.

  • Modification Protocol: the methods for developing, validating and implementing those changes.

  • Impact Assessment: the benefits and risks of the changes.

Changes outside the authorised PCCP need a new marketing submission where one would otherwise be required. Labelling should state that the device has an authorised PCCP. Our article on what a De Novo review is covers one of the routes a PCCP can be part of. Our guide to writing a predetermined change control plan covers each component in detail.

The performance monitoring plan

Pathway Monitoring plan under the draft (Section XI)
510(k) Generally not required, unless it is a special control
De Novo May be made a special control
PMA May be made a condition of approval

Source: draft guidance, Section XI.

The draft places the plan in the Risk Management File. It does not replace the reporting and correction requirements of 21 CFR Parts 803 and 806.

In our view, plan the PCCP and the monitoring plan together: one defines the permitted changes, the other detects when change is needed.

GMLP, transparency and the AI device list

Good Machine Learning Practice is a set of 10 principles published by FDA, Health Canada and the MHRA in October 2021. IMDRF published a final document with 10 principles in January 2025.

The transparency principles, from the same three regulators, build on GMLP principles 7 and 9 and are framed around who, why, what, where, when and how. Both are principles, not guidance or law. Our article on AI and ML standards for medical devices covers the standards side.

FDA’s AI-Enabled Medical Devices list shows authorised devices, identified mainly through AI terms in their summaries or classification. It is not comprehensive and is updated periodically. FDA says it will explore tagging devices with LLM-based functionality.

In the EU, the same software can also be a high-risk AI system under the AI Act; our article on high-risk AI systems explains that.

A hypothetical example

Hypothetical example: a Dutch start-up has an image-analysis SaMD and plans a 510(k) (the pathway is assumed).

  1. Description: it states AI use, inputs, outputs, users, environment and workflow (Section V).

  2. Data: its test data come from sites not used in development and are kept away from the developers (Section VIII).

  3. Subgroups: it pre-specifies subgroup analyses by sex and age (Section X).

  4. Labelling: it adds a model card, based on the example in the appendices.

  5. Monitoring: it places a performance monitoring plan in the Risk Management File (Section XI).

  6. PCCP: it adds a PCCP for retraining, with all three components.

  7. Feedback: it asks FDA for feedback through a Q-Submission.

Conclusion

The draft maps what a submission for an AI-enabled device should contain, but it is still a draft. The PCCP guidance is final and is the route for planned model changes without a new submission.

In our view, build the file in the draft’s section order, design the data and subgroups early, plan the PCCP and monitoring plan together, and check the draft’s status on FDA’s guidance list before you submit.

If you need help preparing an AI-enabled device for FDA, our software lifecycle management team can help, or contact MedQAIR.

Frequently asked questions

Is FDA’s AI guidance final?

The lifecycle management and marketing submission guidance is still a draft. The PCCP guidance for AI-enabled device software functions is final.

Is a PCCP mandatory?

No. A PCCP is optional. It lets FDA authorise planned changes in advance, so they do not need a new marketing submission.

Is a performance monitoring plan required?

Under the draft, generally not for a 510(k) unless it is a special control. For a De Novo it may be a special control, and for a PMA a condition of approval.

What is GMLP?

Good Machine Learning Practice: 10 guiding principles published by FDA, Health Canada and the MHRA in October 2021. They are principles, not guidance or law.

Does the draft guidance apply to software inside a device?

Yes. It covers AI-enabled device software functions, which can be standalone software or part of a hardware device.

Latest Blogs

November 20, 2026

On 7 January 2025, FDA issued the draft guidance Artificial Intelligence-Enabled Device Software Functions: Lifecycle Management and Marketing Submission Recommendations.

November 19, 2026

Since 2 February 2026, FDA's Quality Management System Regulation (21 CFR Part 820) has incorporated ISO 13485:2016 by reference. Even

November 18, 2026

Since 2 February 2026, 21 CFR Part 820 has been the Quality Management System Regulation (QMSR), incorporating ISO 13485:2016 by

logo

Unlock Your Quick Guide to AI Act
Compliance!

Explore AI-enabled SaMD requirements with our easy step-by-step guide.

Cookies help us improve your experience on our website. By using our site, you consent to the use of cookies as described in this policy.