How to read an FDA 510(k) summary for an AI or digital health device
Every FDA 510(k) clearance leaves a public record: a K number, a decision letter and usually a summary of what the company submitted. Here is what each part tells you about an AI or digital health device, and what it leaves out.
Docket, covering health regulation and AI in care
, 3 min read
90FDA review days the agency aims for on a 510(k)
What to know
- A 510(k) clearance means the FDA found a device substantially equivalent to a legally marketed predicate, not that it proved the device works better.
- The public summary names the predicate, the indications for use and the testing the company relied on.
- The indications for use statement, not the press release, sets what the device is cleared to do.
When a company announces that the FDA has cleared its software or wearable, the press release is the company's version. The FDA keeps its own record, and anyone can read it in the agency's public 510(k) Premarket Notification database. For AI tools and digital health devices, that record is the best way to check what was actually cleared. This guide walks through it piece by piece.
The K number
Every 510(k) submission gets a number that starts with K, followed by six digits. The first two digits show the year the FDA received the submission, so K253055 arrived in 2025. Search the database by that number, the company name or the device name. The entry shows the decision date, the decision (for a clearance, "substantially equivalent"), the review panel, the product code and whether a third party did the review.
The product code is worth a click. It links to the device classification, which tells you the regulation the device falls under and its risk class. Most software cleared through a 510(k) is Class II, a moderate-risk class that comes with special controls such as specific testing requirements. Our explainer on clearance, approval and De Novo covers how the classes and pathways fit together.
The decision letter
The letter is the FDA's formal notice to the company. Much of it is standard language: the company may market the device, it must still follow general controls such as registration, labeling and complaint handling, and clearance is not approval. Companies are not supposed to describe a cleared device as "FDA approved."
The 510(k) summary
The summary is the most useful document. Federal rules let a company file either a summary or a statement that it will share the safety and effectiveness information on request, but most now file a summary. It usually appears in the database some weeks after the decision. Look for these sections:
- Predicate device. A 510(k) argues that the new device is as safe and effective as one already on the market. The summary names that predicate, often with its own K number, so you can follow the chain back.
- Device description. What the product is made of: the hardware, the software, the inputs it takes and the outputs it gives.
- Indications for use. The exact patients, settings and purpose the clearance covers. This line is the boundary. A tool cleared to flag possible findings for a radiologist to review is not cleared to make the diagnosis.
- Comparison of technological characteristics. Where the new device matches the predicate and where it differs, with the company's case for why the differences do not raise new safety questions.
- Performance data. The bench, software and clinical testing the company submitted.
Reading the performance data on an AI device
For AI software, the performance section often reports standalone results, such as sensitivity and specificity on a test set, and sometimes a reader study comparing clinicians with and without the tool. Check how many cases were tested, where they came from, and whether the test data was kept separate from the data used to build the model. Many summaries say little about patient demographics or how many sites contributed data. A short section is common and is itself information.
The summary may also mention a predetermined change control plan, which lets the company update the model in ways the FDA agreed to in advance. Our PCCP explainer covers what those plans allow.
What the record does not tell you
A clearance rests on equivalence to a predicate, not on proof of better outcomes for patients. Researchers who traced AI devices through the public database found that most algorithm updates came with no reported clinical benefit, as our story on a JAMA Network Open study of AI device updates describes. The record also does not show price, how many hospitals use the product or how it performs once deployed.
Why it matters
Company announcements tend to describe a clearance in its broadest terms. The indications for use and the performance section show what the FDA actually reviewed. When we cover clearances such as Bracco and Avicenna.AI's aneurysm software or PulseAI's smartwatch ECG model, the FDA record is where we check the claims.
This article is general information about how technology works. It is not medical advice. Talk to a clinician about your own health.
FDA and health regulationDigital health and AI in careWearables
Written by an AI correspondent under standards set by editor Jacob Knight. Spotted an error? Tell us and we will fix it.