Recommended Free Tools
By December 5, 2025, 1,357 AI- and machine-learning-enabled medical devices had FDA clearance or approval. But a 2026 census found publicly identifiable registered prospective trials evaluating patient-centered outcomes for just 3 of them. That gap does not mean the other devices were untested or unsafe. It means authorization, evidence that a device performs a technical task, and evidence that using it improves patients’ lives are different things.
What the device count does—and does not—tell you
Rawan Abulibdeh and co-authors counted 1,357 AI/ML-enabled devices with FDA clearance or approval through December 5, 2025, in a study published in PLOS Digital Health on August 19, 2026. The count is a snapshot, not a live total: the FDA’s device list is periodically updated, and the agency says the list is not comprehensive.
As an Amazon Associate I earn from qualifying purchases.
Most devices in the study were for imaging. The authors classified 1,059 of 1,357 devices (78%) as radiology. Stanford’s AI Index Report 2026 counted 1,039 of the same total of 1,357 (76.6%) in radiology, using its dataset and classification. Those figures are not interchangeable; they reflect different source methods.
The headline count answers how many devices appear in a defined FDA-related census. By itself, it does not say how many have been shown to improve diagnosis, treatment decisions, survival, recovery, or quality of life.
#1 Best Overall
How few had public prospective evidence on patient outcomes?
The PLOS study linked the FDA device database with the American College of Radiology Data Science Institute’s AI Central catalogue, searched PubMed for published outcome studies, and looked for prospective trial registrations in ClinicalTrials.gov. It identified the following evidence among the 1,357 devices:
| Evidence identified | Devices | Share of 1,357 |
|---|---|---|
| Linked to a registered prospective trial | 34 | 2.5% |
| Had trial results posted | 12 | 0.9% |
| Had a peer-reviewed publication | 12 | 0.9% |
| Evaluated patient-centered outcomes | 3 | 0.2% |
These are separate evidence categories, not a sequence that should be added together. The study counted patient-centered primary endpoints such as mortality, major morbidity, hospitalization or readmission, validated quality of life, function, and symptom burden. Measures such as sensitivity, specificity, area under the curve (AUC), workflow efficiency, and other technical metrics did not qualify on their own.
Of the 34 registered trials, 32 (94%) were industry-led. Nine included any subgroup analysis, and none analyzed language. These details matter because evidence from one population or setting may not tell clinicians how a system performs for every group of patients.
What’s actually slowing this PC down?
Pick the symptom - the matching free tool is one click away.
Why FDA authorization is not the same as proof of better outcomes
The FDA says devices on its list met applicable premarket requirements, including a review of safety and effectiveness appropriate to the device’s intended use and technology. The agency also cautions that public summaries do not include most information that may have been submitted. A limited public record is therefore not a complete record of everything reviewed by regulators or held by a manufacturer.
Rank #3
The kind and extent of evidence required depend on the device and its regulatory pathway. Under the 510(k) pathway, a device is cleared based on substantial equivalence to a legally marketed predicate. That process can allow clearance without a new prospective patient-outcomes trial, but it is still a premarket review. De Novo and premarket approval (PMA) are distinct pathways; their requirements should not be treated as identical to 510(k). The PLOS authors report that 510(k) predominates among AI/ML devices and note that predicate chains can carry forward evidence limitations.
As a co-author of the PLOS study, Sebastián Andrés Cajas Ordóñez put the distinction this way in an interview reported by Earth.com: “FDA clearance means the device resembles one already on the market. It does not mean it helps anyone.” That is an interview quote, not FDA language or a substitute for the qualifications above: FDA authorization does involve applicable review, but authorization alone does not establish improved patient outcomes.
Rank #4
What “tested” can mean in practice
Testing is not one threshold. A device may be assessed on technical performance, evaluated retrospectively using existing records, studied prospectively in actual care, or tested in a randomized trial. The endpoint also matters: showing that an algorithm identifies a pattern is different from showing that deploying it changes treatment or helps patients.
An earlier analysis by Windecker and colleagues examined public FDA materials for 903 devices listed as of August 31, 2024. Performance studies were reported for 505 devices (55.9%). Among those reported studies, 193 were retrospective (38.2%), 41 were prospective (8.1%), and 12 were randomized (2.4%). The design percentages use the 505 devices with reported performance studies as their denominator, not all 903 devices. The authors did not search other sources for additional clinical evaluation reports, so this is not a complete inventory of evidence.
Study design is only one part of the question. A prospective study may still use a technical endpoint rather than a patient-centered one; a result from one hospital may not generalize to another. In the 2026 census, the authors did not formally score risk of bias. Their method also could not capture unregistered or confidential company testing, and some FDA summaries were difficult to access. An unlisted trial is not proof that no other evidence exists.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.How performance can fall short in a hospital
A 2021 external validation by Wong and colleagues illustrates why real-world performance needs to be checked rather than inferred from a model’s intended purpose. The researchers retrospectively assessed the Epic Sepsis Model among 27,697 adult patients and 38,455 hospitalizations at Michigan Medicine, from December 6, 2018, through October 20, 2019. At a score threshold of 6 or higher, the model failed to identify 1,709 of 2,552 sepsis cases (67%) and generated alerts for 6,971 of 38,455 hospitalizations (18%). Its hospitalization-level AUROC was 0.63 (95% confidence interval, 0.62–0.64); the authors described poor discrimination and calibration.
This is evidence about one model, one threshold, one retrospective evaluation, and one health system—not a current audit of every hospital or later version of the model, and not a verdict on medical AI as a whole. A hospital assessing any predictive system needs to know how it was validated locally, what threshold it uses, who receives alerts, and what happens after an alert.
What to ask about a medical AI system
Patients and clinicians can ask a hospital or care team for specifics rather than treating “FDA-cleared” or “AI-powered” as a complete explanation:
- What is the intended task? Is the system analyzing an image, flagging risk, recommending an action, or doing something else? Is it an aid for a clinician or does it affect care directly?
- What kind of evidence supports it? Was it tested on technical performance, retrospective records, prospective patients, or in a randomized study? What was the endpoint?
- Where and in whom was it evaluated? How many sites and patients were included, and how closely do those patients and care settings match the people and hospital using it?
- Were relevant groups examined? Ask whether results were reported by factors such as age, sex, race or ethnicity, language, pregnancy, or pediatric status where relevant. A headline average can conceal variation.
- How is it monitored in routine care? Who reviews its output, what action follows, and how does the hospital detect missed cases, unnecessary alerts, or changes in performance?
The 2026 PLOS census is a map of publicly identifiable evidence, not a pooled test of whether medical AI works or harms patients. It supports a specific conclusion: public prospective evidence of patient-centered benefit was rare among the devices counted, while the available public data do not establish that every other device lacked testing.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
