FDA cleared 1,357 AI devices. Three were tested on patients.

Clearance is not proof.

A systematic review published in PLOS Digital Health on August 19, 2026, examined every AI and machine-learning medical device authorized by the FDA through December 5, 2025.

The count reached 1,357 devices.

Of those, 34 were linked to a registered prospective clinical trial.

Twelve posted results.

Twelve reached peer-reviewed publication.

Three were evaluated for outcomes that matter to a patient: death, major complications, or hospital readmission.

Three, out of 1,357, survived the funnel.

The evidence funnel

Most AI devices reach the market through the FDA's 510(k) pathway.

That pathway asks a manufacturer to show substantial equivalence to a device already on the market.

It does not ask for proof that the new tool improves what happens to the patient on the table.

Radiology accounts for 1,059 of the 1,357 cleared devices, or 78% of the total.

Fewer than 1% of radiology devices have a registered prospective trial behind them.

Cardiovascular and neurology devices, categories that include many surgical planning and guidance tools, fare only slightly better, with trial rates of 9.5% and 9.7%.

Anesthesiology's zero

Twenty-two AI devices have been cleared for anesthesiology.

None of them has a registered prospective trial.

That is not a rounding error. That is a category with zero.

Where trials do exist, they tend to be small. Nearly three-quarters of the 34 registered trials enrolled fewer than 500 patients, and a quarter enrolled fewer than 100.

Most were run at a single site in the United States, which limits what they can say about performance elsewhere.

Pediatric patients were almost universally excluded from enrollment.

A three-phase fix, unadopted

The study's authors propose a staged model, borrowed from how drugs get tested.

Phase 0 would require retrospective validation on diverse datasets before a device is even cleared.

Phase 1 would require a prospective study of at least 500 patients, embedded in a real clinical workflow, before wide release.

Phase 2 would require a multi-center outcome trial of at least 2,000 patients, powered to detect differences across demographic subgroups.

None of these three phases is part of FDA's requirements for a Class II AI device.

The infrastructure to run them exists. The requirement to use it does not.

For a surgeon deciding whether to trust an AI-guided planning tool, clearance alone cannot answer that question.

---

The AI Claims Audit was built for exactly this gap, separating what a device is cleared to do from what it has actually been shown to do in patients.

Source: PLOS Digital Health