Skip to content
← Back to blog
Industry·Updated October 22, 2026·6 min read

AI in healthcare by the numbers: more than 1,600 FDA-listed devices

Healthcare AI is easy to count: most clinical AI tools need FDA authorisation. The FDA list grew more than fourfold in five years, and it holds a design lesson.

Healthcare is where AI's progress is easiest to count, because clinical AI tools that qualify as medical devices generally need FDA authorisation before they reach patients (some decision-support software is excluded). In most industries adoption is fuzzy: a model ships inside a product and you infer its impact. In medicine there is a public list, a decision date, and a category. The trend line is steep, and it tells you something useful about how serious AI gets deployed everywhere else.

The numbers

  • The FDA's public list of AI-enabled devices now has more than 1,600 entries (our count of the list as published in September 2026). At its July 2025 update it held 1,247, and about 950 had been authorised as of August 2024. The FDA says the list is not comprehensive.
  • The list grew more than fourfold in five years: about 300 entries carry a decision date up to July 2020, about 1,300 up to July 2025 (our count by decision date).
  • Radiology dominates (76% of entries), then cardiovascular (9.5%) and neurology (4.5%). One 2025 analysis by the regulatory consultancy Innolitics put the median 510(k) review time for AI devices at 142 days.

The radiology concentration is not an accident. Imaging is data-rich, the inputs are standardised, and the "right answer" is well defined: a tumour is either present or it is not. That combination of clean data and a clear ground truth is exactly what makes a problem tractable for machine learning, in medicine and out of it. Where the data is messy and the correct answer is contested, clearances are far rarer.

A median review of around 142 days is worth sitting with. In software terms that is glacial; most teams ship features in days. But it is the price of a system that demands evidence before deployment, and it produces something the consumer-software world mostly lacks: a public, auditable record. Every entry on that list has met the FDA's premarket requirements, including a review of its safety and effectiveness. That bar has limits: about 96% of entries came through the 510(k) route, and most authorisations rely on pathways that do not require new clinical trials. One analysis of 691 devices found only 1.6% reported data from randomised clinical trials. So the growth is real, and it counts tools that passed a regulator's review, which is still a very different signal from "more AI products launched."

What's actually shipping

Many listed devices are assistive, not autonomous: flagging abnormalities for a radiologist to confirm, guiding a surgeon, supporting bedside evaluation. Autonomous ones exist (IDx-DR, authorised in 2018, was the first autonomous AI to detect diabetic retinopathy), but in the assistive ones a human stays in the decision by design. The model narrows attention and catches what a tired clinician might miss; the clinician carries the accountability. That division of labour is the single most important design pattern in the whole field, and it generalises far beyond hospitals.

It is worth being clear about why "assistive" is not a euphemism for "unambitious." An assistive tool that reliably catches the early-stage tumour a fatigued radiologist would have missed on the last scan of a long shift is enormously valuable, arguably more so than a fully autonomous system nobody trusts enough to deploy. The constraint of keeping a human accountable does not shrink the impact; it changes its shape. The model does the tireless, high-recall first pass; the human brings context, judgement and responsibility. That pairing is more powerful than either alone, and it is the template for AI in any domain where being wrong has real consequences.

The other reason this matters: liability and trust travel together. A clearance is not just a safety certificate, it is a statement about who is answerable when something goes wrong. By keeping a qualified human in the loop, the system keeps accountability somewhere a patient can actually direct it. Autonomous systems blur that line, which in our view is why they deserve the closest scrutiny.

Healthcare AI advances at the speed of evidence and oversight. That's not a brake on innovation; it's what makes it trustworthy.

The design lesson for everyone else

The FDA's Predetermined Change Control Plan, pre-authorising how a model may evolve after authorisation, is in our view a preview of how other regulated industries will handle software that learns. Instead of a new marketing submission for every planned change, you agree the boundaries of acceptable change up front and monitor against them. We expect the same logic to spread to finance, insurance and any domain where a model's output carries real consequences.

For teams building AI in any high-stakes setting, the healthcare playbook is worth copying directly:

  • Pick problems with clean inputs and a checkable answer first. That is where AI earns trust fastest.
  • Ship assistive, not autonomous. Keep a qualified human accountable for the decision, especially anything irreversible, the same human-in-the-loop principle most assistive devices are built around.
  • Plan for the model to change. Define in advance how it may be updated and how you will detect drift, rather than treating the launch version as final, the same discipline the EU AI Act will require of high-risk systems, from December 2027 for most standalone uses and August 2028 for AI inside regulated products such as medical devices.
  • Log everything. Evidence is what makes the system defensible when something goes wrong, a lesson AI in legal learned the hard way with hallucinated citations.

None of this slows good products down. It is what lets them survive contact with reality. If you're building AI for regulated healthcare products, see how we approach it or start a conversation.

Sources

Written by Zain Ali

Frequently asked questions.

The FDA's public list of AI-enabled medical devices had more than 1,600 entries in our count of the September 2026 version. At its July 2025 update it held 1,247, and about 950 had been authorised as of August 2024. The FDA says the list is not comprehensive. Every entry passed FDA premarket review, though about 96% came through the 510(k) route, and most authorisations rely on pathways that do not require new clinical trials.

Radiology dominates the FDA's list with about 76% of entries, followed by cardiovascular devices (9.5%) and neurology (4.5%). That concentration isn't an accident: imaging is data rich, the inputs are standardised, and the right answer is well defined, since a tumour is either present or it isn't. Clean data plus a clear ground truth is what makes a problem tractable for machine learning. Where data is messy and the correct answer is contested, clearances are far rarer.

One 2025 analysis by the regulatory consultancy Innolitics put the median 510(k) review time for AI devices at 142 days, and it notes that it used AI to identify the devices. In software terms that's glacial, since most teams ship features in days, but it's the price of a system that demands evidence before deployment. What it produces is something consumer software mostly lacks: a public, auditable record of what was authorised.

Many are assistive rather than autonomous: flagging abnormalities for a radiologist to confirm, guiding a surgeon, supporting bedside evaluation. Autonomous ones exist, such as IDx-DR, authorised in 2018 as the first autonomous AI to detect diabetic retinopathy. In the assistive ones the model narrows attention and catches what a tired clinician might miss, and the clinician carries the accountability.

It pre-authorises how a model may evolve after authorisation. Instead of a new marketing submission for every planned change, you agree the boundaries of acceptable change up front and monitor against them. In our view it previews how other regulated industries will handle software that learns, and we expect the same logic to spread to finance, insurance and any domain where a model's output carries real consequences.

Keep reading

All articles
Workplace

Is government where software agency revenue is growing? What the Census Bureau's 2026 services data says, and what it means before you staff a public-sector job

7 min read
Security

Does a CCPA opt-out have to cover every device and service on an account? What California's Disney settlement says to build

7 min read
Industry

AI in legal: review at machine speed, judgment at human pace

3 min read