FDA-Cleared Medical AI and Patient Outcomes: What Clinics Should Ask

In short: A systematic review published in PLOS Digital Health in August 2026 looked at AI medical devices cleared by the US FDA before 5 December 2025. Of at least 1,357 devices, only 34 were linked to registered prospective trials, 12 had results posted, 12 had peer-reviewed publications and just 3 assessed patient outcomes such as death, hospitalisation or quality of life. That does not mean the tools are unsafe. It means "FDA-cleared" tells you a tool met a regulatory bar, not that it has been shown to improve patient health. A clinic buying AI should ask for more.

Key Takeaways

• The review, led by researchers at University Health Network in Canada, counted at least 1,357 FDA-cleared AI devices and checked how many had been tested on patient outcomes.
• The funnel is steep: 34 linked to registered prospective trials, 12 with posted results, 12 with peer-reviewed publications, 3 that assessed patient outcomes.
• Nearly three-quarters of the studies enrolled fewer than 500 participants and a quarter enrolled fewer than 100.
• Most trials ran at a single academic centre or a well-resourced health system, so performance in other settings is unknown.
• Healio notes that the FDA's 510(k) route requires "substantial equivalence" to an existing device, not prospective proof of clinical benefit.
• Clearance is still meaningful, but it is a floor, not proof of benefit. Your own evaluation has to fill the gap.

What the Review Looked At

The researchers searched the FDA's device database and a catalogue of cleared AI tools, then checked whether devices cleared before 5 December 2025 had been tested on patient health outcomes: death, illness, hospitalisation, readmission, quality of life and symptom burden.

As reported by Healio on 21 August 2026, the main findings were:

The authors also noted limits in the studies that did exist. Participants were often drawn from narrow groups, with pregnant patients, young people, older adults and non-English speakers often excluded. They also said there are no standard reporting or transparency requirements, and that logistics and weak financial incentives may discourage patient testing.

What This Does and Does Not Mean

It does not mean cleared devices do not work. Many AI tools, such as image-reading aids, are tested well on accuracy: how often they find what a specialist finds. That is a real and useful form of evidence.

It does mean accuracy and benefit are different questions. A tool can be accurate and still not improve outcomes, for instance if clinicians ignore it, if it creates extra work or if it flags findings that do not change treatment. Only studies that follow patients can show that.

It also means local performance is a gap. A model validated in one academic hospital may behave differently with your equipment, patient mix or workflow.

Five Questions to Ask Before You Buy

1. What was it tested on, and where? Ask for the study, the setting and the patient mix, and compare it with your own clinic.
2. Did it improve outcomes or only accuracy? Ask directly. An honest vendor will say if there is no outcome data.
3. How was it cleared? Ask which pathway applied and what it was compared with. Our guide to evaluating a healthcare AI tool walks through this.
4. How will we check it locally? Agree a pilot, a baseline and a measure before you roll it out.
5. What happens when it is wrong? Look for clear human review, an audit trail and a way to report errors.

Not every healthcare AI tool is a regulated medical device. Administrative tools, such as scheduling and patient messaging, usually are not, and they raise different questions about data, wording and escalation. Our AI patient communication guide for clinics covers those.

The Short Version

Regulatory clearance is a starting point. Most cleared AI tools have not been shown, in published trials, to improve what happens to patients, and the few that have were tested in settings unlike most clinics. Ask for the evidence, run a small local pilot and measure the outcome you care about.

For how we read the evidence in this area, see our clinical AI evidence digest and the post on AI medical scribes and the randomized trial.

Related Insights

• AI Medical Scribes: What the Randomized Trial Shows
• AI Sepsis Early Warning: What the Evidence Shows
• AI in UK Healthcare: What to Check Before You Buy

Sources

• Healio, Most AI tools cleared by FDA were not tested on clinical outcomes, August 21, 2026
• Abulibdeh R, et al., PLOS Digital Health, 2026, doi:10.1371/journal.pdig.0001597 (figures taken from the Healio report; we did not read the paper itself)