The FDA has cleared 1,357 AI medical devices. Three were tested on patient outcomes.

A review of every AI-enabled medical device the FDA has cleared found that only three of 1,357 have ever been tested on whether patients who use them actually live longer, avoid hospitalization, or fare better in any other outcome that matters to a patient — as opposed to whether the device performs its narrower technical task accurately [s1].

What the researchers actually did

Rawan Abulibdeh at the University of Toronto led the analysis, published 19 August in PLOS Digital Health, examining all 1,357 AI-enabled devices the FDA had authorized through 5 December 2025 and tracing how each one was actually evaluated in patients before it reached clearance [s1]. That's a complete census of the FDA's AI-device clearances to that date, not a sample — the finding describes the entire regulated category, not an unrepresentative slice of it.

The funnel, and how narrow it gets

Of the 1,357 cleared devices, only 34 appeared in a registered clinical trial at all [s1]. Of those 34, 12 had posted results and 12 had peer-reviewed publications [s1] — and narrowing further, from that pool, just three devices had been tested specifically on patient-centered outcomes: mortality, stroke, hospitalization, or quality of life [s1]. Abulibdeh's own reaction to that number, quoted in coverage of the study, undercuts any suggestion this was an expected or trivial finding: "We expected the evidence base to be thin, but not this thin. Out of 1,357 AI devices the FDA has cleared for use in patient care, only three have been tested on whether patients actually live longer or better" [s1].

Why a device can clear the FDA without this kind of testing

Most AI-enabled devices reach the market through the FDA's 510(k) clearance pathway, which requires showing a new device is substantially equivalent to an already-legally-marketed device — a standard built around technical performance and safety, not a requirement to demonstrate the device improves patient outcomes in a trial. That's not a loophole unique to AI devices; it's the same pathway most conventional medical devices use, and it predates AI-enabled tools by decades. What the study's finding highlights is that AI devices are being cleared through that same technical-equivalence standard at a rate that's outpaced any parallel expansion of patient-outcome testing specifically for this newer device category.

Who the existing evidence doesn't cover

Beyond the outcome-testing gap, the study flagged who's missing from what evidence does exist: most studies were conducted in well-resourced healthcare systems, and pregnant women, adults over 75, and non-English speakers were systematically excluded from the research base underlying these devices [s1]. That matters independent of the outcomes question — even the handful of studies that exist may not generalize to the full range of patients these devices are cleared to be used on.

What this finding doesn't establish

This study documents an evidence gap; it doesn't establish that the 1,354 devices without patient-outcome testing are ineffective or unsafe. A device can perform its stated technical function reliably — flagging an abnormal scan, calculating a risk score — without that function's real-world effect on patient outcomes ever having been directly measured. The absence of outcome evidence is a statement about what regulators and manufacturers have not yet demonstrated, not proof of harm or failure.

What to watch next

Whether the FDA's ongoing effort to develop lifecycle monitoring standards for AI-enabled devices — an initiative the agency has separately sought public input on — eventually closes this gap by requiring outcome data post-clearance, since Abulibdeh's analysis suggests pre-clearance requirements alone haven't produced it.