FDA Has Authorized 41 AI Models to Retrain After Clearance. Six Say How They'll Prove It Went Well.

By Sentir Health · 07/21/2026

A Predetermined Change Control Plan exists for one reason: to let a company change its device without filing a new submission each time. So the most direct question you can ask of the 70 AI/SaMD devices that hold one is what change they asked permission for.

The answer, more often than anything else, is retraining. By our classification, 41 of the 70 hold pre-authorization to retrain or update the model itself. Not to add a scanner vendor, not to swap a component. To change the weights.

Now the second question. Of those 41, how many state an explicit real-world performance-monitoring commitment in their cleared documentation?

Six.

Retraining is the change they ask for

This is worth sitting with, because it says something about where the industry believes its products are going. Compatibility expansions and input/output changes are useful, and plenty of plans cover them. But the single largest group of AI/SaMD PCCPs, more than the next three types combined, is plans that pre-authorize changing the model. The June cohort alone added four: Brainomix, Siemens, Mighty Oak Medical, and Tyto Care all cleared devices in June 2026 with retrain-type plans.

These are companies telling FDA, in writing, that the version of the model that cleared is not the version they intend to be running in two years.

What a retrain plan actually promises

The cleared documents are specific about the deal. Therapixel's MammoScreen plan is a good public example. It pre-authorizes support for new mammography manufacturers, and for manufacturers beyond the first tier that support requires retraining the model. The plan names its acceptance criteria in advance: standalone AUC on the new manufacturer's images must be non-inferior to performance on Hologic images, with the lower bound of the 95 percent confidence interval allowed to sit no more than two points below. Updates go out with a two-week advisory notice, and users can opt out during the notice window.

That is what a retrain authorization looks like: named modifications, pre-specified statistical gates, a defined release procedure. We walked through the required Modification Protocol elements in The PCCP Inflection Point, and the burden they create is the same for every one of the 41. When you exercise the plan, you must demonstrate, on demand, that the change met the criteria you wrote down. FDA does not require you to monitor continuously in the meantime, and does not ask for interim reports. It asks for the demonstration when you act.

Which is exactly what makes the monitoring commitments in the cleared documents so interesting. Nobody is required to state one. So the six that did are telling you something real about what they built. The silence of the other 35 is more ambiguous, and worth reading carefully.

The gap is not closing

When we published our read of the 2026 landscape in June, the tracker held 64 AI/SaMD PCCPs, and 8 stated a monitoring commitment. Five weeks later the tracker holds 70. Six new AI/SaMD plans cleared in that window.

The number of monitoring committers is still 8.

Every one of the six new plans, including all four retrain-type plans, is silent on real-world performance monitoring. That does not mean these companies are flying blind. As we noted in The PCCP Inflection Point, almost everyone tracks model performance somehow: scripts, dashboards, a notebook someone runs each month. What the silence in the cleared documents records is a bet. The bet is that when a retrain happens and the demonstration comes due, FDA will accept whatever form that internal tracking took.

The permission is public. The proof is private.

Here is the asymmetry that we think defines the next two years of this market.

Everything a company is authorized to change is public. A competitor can read your modification list. A hospital AI governance committee can read your acceptance criteria. A plaintiff's expert can read the performance gates you promised FDA, down to the confidence interval. The tracker exists because these documents are public records, and we are not the only ones reading them.

What nobody outside the company can see is whether the proof exists. Whether there is a baseline captured before the model started moving. Whether the acceptance criteria running in the company's evaluation code still match the wording FDA cleared. Whether the performance record would survive the departure of the engineer who set it up. The 41 retrain plans are precisely the plans that will someday be exercised, because retraining is why they were written. Each exercise converts a public promise into a private demonstration, produced under scrutiny, from evidence that either exists by then or does not.

A baseline is the one piece you cannot reconstruct afterward. When the first exercised retrain PCCP gets audited, or the first post-retrain performance question comes from a health system rather than a regulator, the cohort sorts into two groups: companies with a record, and companies with a story. Nothing in the public documents says which group a given company is in, and that is the problem. Evidence a team generates about its own model is self-reported, however careful the team is, and there is no independent name behind it when a hospital buyer or an auditor asks who vouches for the numbers. FDA may well accept internal tracking when the demonstration comes due. The buyers asking earlier and more often are not FDA.

If you hold one of the 41

Three things are worth doing this quarter, whether or not a retrain is scheduled.

Capture the baseline now, against the endpoints your plan actually names. If your gates are written in AUC by manufacturer, a dashboard of request volumes will not defend you.

Check that the acceptance criteria in your running evaluation match the cleared plan's wording. These drift apart quietly, and the gap surfaces at the worst possible moment.

Keep the record somewhere that outlives your current team, in a form you would be comfortable handing to a reviewer. Evidence a team generates about its own model, held only by that team, is the weakest version of the truth it can hold.

You can see where your device sits, and who else holds a retrain authorization in your clinical panel, in the Sentir PCCP Tracker, browsable by category, clinical panel, PCCP type, or company.

Sentir Health is the independent system of record for the performance of FDA-cleared AI. We keep the baseline and the evidence your PCCP requires you to produce on the day you exercise it. Learn more or book a call.

Methodology note: Counts come from FDA 510(k) and De Novo records via the Sentir PCCP Tracker, current as of 2026-07-21. "Retrain-permitting" is our editorial label for plans whose primary authorized change is model retraining or a model update, based on the cleared summary. A small number of AI/SaMD plans are statement filings whose contents are not public; those are not counted as retrain-permitting. PCCP type, category, and monitoring-commitment labels are editorial classifications based on public documents, not official FDA designations.