Kivorent News

Thursday, November 19, 2026

Health techExplainer

When a software update to a medical device needs a new FDA 510(k)

FDA rules say a cleared device needs a new 510(k) when a change could significantly affect safety or effectiveness. Here is how the agency's software-change guidance turns that phrase into questions, and why retraining an AI model often trips them.

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

2017year the FDA issued its software-change 510(k) guidance

What to know

  • Under 21 CFR 807.81(a)(3), a cleared device needs a new 510(k) when a change could significantly affect its safety or effectiveness, or when its intended use changes in a major way.
  • The FDA's software-change guidance asks whether an update adds or changes a risk that could cause significant harm, or changes clinical performance.
  • A maker that decides no new 510(k) is needed must still document why; the FDA can disagree at inspection.

A 510(k) clearance covers the device as it was when the FDA looked at it. Software, though, keeps moving: bug fixes, new screens, a faster engine, a retrained model. Most of those updates never go back to the agency. Some must. The line between the two is drawn by one sentence of regulation and a guidance document that turns it into a flowchart, and the maker, not the FDA, does the first reading.

The sentence that decides

The rule sits at 21 CFR 807.81(a)(3). A device already on the market needs a new 510(k) when a change "could significantly affect" its safety or effectiveness, or when its intended use changes in a major way. Notice the verb. The test is whether the change could matter, not whether it did. A maker that retrains a model and finds it performs better has not escaped the question; it has answered a different one.

Nobody files anything to make this call. The company works through the question itself and, if it concludes no new submission is needed, writes down why. That memo, usually called a letter to file, stays in the company's records until an FDA investigator asks for it. Our explainer on warning letters covers what happens when the investigator reads it and disagrees.

How the guidance turns it into questions

In 2017 the FDA published "Deciding When to Submit a 510(k) for a Software Change to an Existing Device," a companion to its general guidance on device changes. It walks a maker through a sequence of questions. Roughly:

  • Is the change made only to strengthen cybersecurity, with no other effect on the device? Then a new 510(k) is generally not needed.
  • Is it only meant to restore the device to the specifications it was cleared with, as a bug fix does? Same answer, as long as nothing else changes.
  • Does it introduce a new risk, or change an existing one, in a way that could cause significant harm? Does it need a new risk control, or change one that guards against significant harm?
  • Could it significantly affect how the device performs clinically, or change what it is for?

A yes to any of the later questions usually points to a new submission. The guidance also tells makers to look at changes together: several small updates, each harmless on paper, can add up to one that is not.

Why AI models keep tripping it

A conventional software update changes code a developer wrote and can read. Retraining a machine learning model changes behavior learned from data, and that behavior can shift in places nobody tested. New training data, a different architecture or a different training objective can change where the model fails, not just how often. That maps straight onto the guidance's questions about new risks and clinical performance.

The FDA made the point in a warning letter to O.N. Diagnostics, the maker of VirtuOst vertebral fracture assessment software, after a March 2026 inspection. According to Orrick's reading of the letter, the agency said a version that replaced a landmarking step with a new machine learning model needed its own 510(k), and that moving the product from desktop software to a web platform was a separate reason, because it widened the device's exposure to cyberattack. Orrick's analysis, published Oct. 9, is the source here; we have not read the letter itself.

The way around it

The cleaner route for AI is to ask in advance. A predetermined change control plan lets a maker describe planned changes, such as retraining on new data within set limits, and get them reviewed with the original submission. Changes inside the plan then skip a new filing. Changes outside it land back at 807.81.

What the rules do not require is telling users. A cleared device can be updated many times with no public record, which is why studies of AI device updates, like the one we covered last week, have to dig for them. When a new 510(k) is filed, its public summary is the best window into what changed; here is how to read one.

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

Sources

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.