iTechGuides is reader-supported. When you buy through links on our site, we may earn an affiliate commission. As an Amazon Associate I earn from qualifying purchases. Learn more
By a December 2025 count, 1,357 AI- or machine-learning-enabled medical devices had been authorized for marketing by the U.S. Food and Drug Administration (FDA). Yet a 2026 evidence census found publicly identifiable registered prospective trials evaluating patient-centered outcomes for just three of them. That gap is important—but it does not mean the other devices had no testing, that FDA review was absent, or that all medical AI is unsafe.
The key is to separate three questions: Did a device meet the regulatory requirements for its intended use? Was its technical performance tested? And has using it been shown to improve outcomes patients care about, such as survival, recovery, or quality of life?
What the 2026 device count does—and does not—show
In a study published August 19, 2026, Rawan Abulibdeh and co-authors counted 1,357 FDA-authorized AI/ML medical devices listed through December 5, 2025. Their structured evidence census looked for registered prospective trials and publicly available results; it was not a clinical trial, a formal risk-of-bias assessment, or a pooled analysis of whether the devices work.
The Tool Desk
Outbyte Driver Updater FREEFix the driver behind crashes, sound loss and screen glitchesFind Drivers →Outbyte PC Repair FREERepair Windows errors before they cause bigger problemsFix Now →Among the 1,357 devices, the authors identified the following publicly traceable evidence:
| 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 the census authors’ counts of identifiable public evidence, not a finding that every other device lacked all testing. The study searched for trial registrations and publications using FDA submission summaries, ClinicalTrials.gov, PubMed, and manual review. Company testing that was not publicly registered, confidential evidence submitted to regulators, bench tests, and some other evaluations may not appear in that count. The authors also noted that some FDA summaries were difficult to access.
The study counted outcomes such as mortality, major morbidity, hospitalization or readmission, validated quality of life, function, and symptom burden as patient-centered. Measures such as sensitivity, specificity, area under the curve (AUC), workflow efficiency, and other technical metrics did not qualify on their own. A device can therefore have evidence that it performs a technical task without public evidence that deploying it improves patients’ lives.
FDA authorization, technical accuracy, and patient benefit are different tests
FDA authorization concerns marketing for an intended use
The FDA says devices on its AI-enabled-device list met applicable premarket requirements, including a review of safety and effectiveness appropriate to the intended use and technology. The agency also cautions that the list is periodically updated and not comprehensive, and that publicly releasable summaries do not include most information that may have been submitted.
What’s actually slowing this PC down?
Pick the symptom - the matching free tool is one click away.
Regulatory pathways differ. For a 510(k), a manufacturer generally seeks clearance by showing that a device is substantially equivalent to a legally marketed predicate. That pathway can allow clearance without a new prospective trial demonstrating better patient outcomes, 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 Digital Health authors report that 510(k) is the predominant pathway among the devices they examined and note that predicate chains can carry forward limitations in the available evidence.
Technical performance is not the same as clinical benefit
A diagnostic model may be assessed on whether its output agrees with a reference standard or correctly identifies cases in a particular dataset. That can establish something useful about performance under the study conditions. It does not by itself show that clinicians using the device make better decisions, or that those decisions improve health outcomes.
For example, a high sensitivity or AUC does not tell a patient whether the tool changes the chance of receiving timely treatment, recovering, avoiding complications, or living longer. Those questions require evaluation of how the device is used in care and what happens to patients, not only whether its predictions match labels.
What earlier FDA-device studies found about testing designs
A 2025 JAMA Network Open analysis by Windecker and colleagues examined public FDA materials for 903 AI-enabled devices listed as of August 31, 2024. It reported performance studies 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 reported studies as their denominator, not all 903 devices.
Windows Errors? Fix Them Before They Spread
Repair common Windows errors and clear accumulated junk for a smoother, more stable PC - no reinstall needed.Free scan · no reinstallOutdated Drivers Are Slowing You Down
One free scan finds every outdated or missing driver and matches the right update for your exact hardware.Free scan · exact hardware matchThis earlier analysis and the 2026 census ask related but different questions, use different device cutoffs, and should not be merged into one denominator. Windecker and colleagues relied on public FDA materials and did not search other sources for additional clinical evaluation reports, so their figures are not a complete inventory of all evaluations.
Most listed devices are used in imaging
Radiology made up 1,059 of the 1,357 devices (78%) in the Abulibdeh et al. census. Stanford’s AI Index Report 2026 counted 1,039 of 1,357 devices (76.6%) in radiology across 17 clinical specialties. The counts differ because the sources use their own datasets and classifications; they should be attributed to their respective sources rather than combined.
The concentration in imaging helps explain why a large device count is not the same as broad proof of patient benefit. A list spanning many medical specialties includes tools with different intended tasks, workflows, regulatory pathways, and evidence needs.
Why performance can change in a real hospital
Real-world results depend on more than an algorithm’s reported accuracy. The patients, data, clinical workflow, alert threshold, and product version can differ between the original evaluation and a hospital’s daily practice. An external validation of one sepsis prediction system illustrates why local evaluation matters.
Do these 3 things before closing this tab:
1Scan for outdated or missing drivers - takes under a minute2Repair Windows errors before they cause bigger problems3Fix the driver behind crashes, sound loss and screen glitchesOne external validation of the Epic Sepsis Model
Wong and colleagues retrospectively evaluated the Epic Sepsis Model at Michigan Medicine in 27,697 adult patients with 38,455 hospitalizations 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, threshold, period, and health system—not a current audit of all hospitals or later versions, and not a verdict on medical AI as a whole. It shows why a hospital should ask how a tool performs in its own setting, what alert burden it creates, and whether its use changes patient outcomes. An erratum corrected errors in Table 2 and the Results; the figures above reflect the corrected study details.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.What to ask when a hospital uses an AI device
Patients and clinicians can make a more informed assessment by asking about the specific device and the role it plays in care. Useful questions include:
- What is the intended task? Is the tool flagging an image, predicting risk, prioritizing a queue, or recommending an action?
- What regulatory pathway applies? Was it cleared through 510(k), authorized through De Novo, approved through PMA, or evaluated by another applicable route? What exact intended use was reviewed?
- What kind of evaluation supports it? Was the evidence bench-based, retrospective, prospective, or randomized? Was the evaluation conducted independently of the manufacturer?
- What outcome was measured? Was it technical accuracy, a workflow measure, a clinician decision, or a patient-centered outcome such as complications, function, or quality of life?
- Where and in whom was it evaluated? How many sites and patients were included, and how closely did the population resemble the people treated here?
- Were subgroup results examined? Consider age, sex, race and ethnicity, language, pregnancy, and pediatric coverage where relevant. In the 2026 census, 9 of 34 registered trials included any subgroup analysis; the authors found no language analysis.
- How is it used in the workflow? Does it inform a clinician, generate an alert, or automate a step? Who reviews its output, and what happens when the tool and clinician disagree?
- Is performance monitored after deployment? Ask how the hospital detects changes in accuracy, alert burden, or outcomes as patient populations, workflows, and software versions change.
What remains uncertain
The 2026 census describes evidence the authors could identify for devices on a list through December 5, 2025. The FDA list is periodically updated, so it may include additional devices by October 2026; the census authors quoted in Earth.com’s coverage said the analysis had not yet been repeated. The census does not inventory all evidence held by manufacturers or submitted confidentially to regulators, compare AI devices with non-AI devices, or determine whether particular products help or harm patients.
Recommended Free Tools
Accordingly, the defensible conclusion is narrower than “medical AI is untested.” Publicly identifiable registered prospective trials evaluating patient-centered outcomes were rare in this device census. That is a meaningful gap in transparent evidence about patient benefit, but it does not settle the performance or value of any individual device.
Quick Recap
Sources
- Rawan Abulibdeh and co-authors, “1,357 AI medical devices cleared, 3 actually tested on patient outcomes,” PLOS Digital Health, published August 19, 2026; device cutoff December 5, 2025.
- Windecker et al., “Generalizability of FDA-Approved AI-Enabled Medical Devices for Clinical Use,” JAMA Network Open, 2025.
- U.S. Food and Drug Administration, “List of Artificial Intelligence-Enabled Medical Devices,” periodically updated resource, accessed October 8, 2026.
- Wong et al., “External Validation of a Widely Implemented Proprietary Sepsis Prediction Model in Hospitalized Patients,” JAMA Internal Medicine, 2021.
- Stanford Institute for Human-Centered AI, AI Index Report 2026, chapter 6, medicine.
- Eric Ralls, Earth.com, October 4, 2026, coverage of the PLOS study and interview with co-author Sebastián Andrés Cajas Ordóñez.
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.
