Kivorent News

Wednesday, November 18, 2026

Health techExplainer

How the FDA lets AI medical devices change after clearance

A predetermined change control plan lets a device maker get FDA sign-off on future software changes, such as retraining an AI model, before it makes them. Here is what goes in one and what it cannot cover.

Docket, covering health regulation and AI in care
, 3 min read

3parts the FDA expects in every change control plan

What to know

  • A predetermined change control plan (PCCP) is reviewed with a device's marketing submission and describes changes the maker intends to make later.
  • Changes made as the authorized plan describes do not each need a new 510(k), De Novo or PMA submission.
  • Changes outside the plan, including most changes to what the device is for, still go through the usual FDA review.

Most medical devices are fixed once they reach the market. A change that could affect safety or performance usually means the maker files a new submission with the FDA and waits. That rule fits a hip implant. It fits poorly with software, and worst of all with AI models, which can get better each time they are retrained on new data. A predetermined change control plan, or PCCP, is the FDA's answer: the maker describes future changes up front, and the agency reviews them before they happen.

Where the authority comes from

Congress gave the FDA explicit power to approve these plans in the Food and Drug Omnibus Reform Act, passed at the end of 2022 as part of a year-end spending law. It added a section to the Food, Drug, and Cosmetic Act that lets a plan be part of a premarket approval (PMA) or a 510(k) clearance, and the FDA applies the same approach to De Novo requests. For the differences between those three routes, see our explainer on clearance, approval and De Novo.

The FDA finalized its main guidance for AI-enabled device software in December 2024, after a draft in 2023. It also has published draft guidance applying the same idea to devices in general, and in 2023 it set out shared guiding principles for these plans with regulators in Canada and the United Kingdom.

What goes in a plan

The guidance asks for three parts.

  • A description of modifications. A specific list of the changes the maker plans, such as retraining a model on more images, adjusting a decision threshold, or supporting a new scanner model as an input. Vague promises to "improve the algorithm" do not qualify.
  • A modification protocol. How each change will be developed, tested and rolled out: what data will be used, how it will be kept separate from test data, what performance targets the new version must meet, and how users will be told about the update.
  • An impact assessment. The maker's analysis of the benefits and risks of the planned changes, including how they could interact with one another.

The FDA reviews the plan alongside the rest of the submission. If the agency agrees, the plan becomes part of the device's authorization. A change made exactly as the plan describes, and meeting its test criteria, can then ship without a new submission. The maker still has to document it under its quality system and update labeling where the plan calls for that.

What a plan cannot do

A plan is a fence, not a blank check. Changes that fall outside the authorized plan follow the normal rules, which often means a new 510(k) or PMA supplement. The FDA's guidance also says plans are generally not the place for changes to a device's intended use, such as moving from helping a radiologist triage scans to making a diagnosis on its own, or extending to a new patient population the original evidence did not cover. A plan also does not exempt the device from the agency's other rules, such as adverse event reporting and recalls.

Why it matters

Clinicians and patients increasingly rely on AI tools that sit inside imaging software, monitoring systems and wearables. Without a plan, a maker must choose between freezing a model that could get better and paying for a fresh review each time it changes. A plan lets updates move faster while keeping the testing rules visible to the FDA in advance. It does not guarantee that updated versions help patients: a recent study of AI device updates of devices first cleared from 2015 to 2021 found that most algorithm updates reported no clinical or organizational benefit in public FDA records. A plan sets the test a change must pass, so the bar is only as useful as the tests the maker writes into it. For how strong evidence is built in the first place, see our explainer on randomized controlled trials. Recent AI clearances on our site include Bracco and Avicenna.AI's aneurysm software.

This article is general information about how technology works. It is not medical advice. Talk to a clinician about your own health.

FDA and health regulationDigital health and AI in careAI law and policy

Written by an AI correspondent under standards set by editor Jacob Knight. Spotted an error? Tell us and we will fix it.