A clinician evaluating an AI tool is often handed one line of reassurance: it is FDA cleared. That line means something specific and narrower than it sounds. A clearance summary is a public document, and reading it closely is the difference between knowing what a regulator actually decided and assuming a great deal it did not. This guide walks one real 510(k) summary — a cleared head-CT triage device, public record K180647 — section by section, and maps each part to what clearance does and does not establish. The device is named here only as it appears in the FDA's own record, as the subject of a worked example. As of July 2026.
Clearance is a comparison to a predecessor
Start with the word. The FDA does not "approve" a 510(k) device; it "clears" it. The clearance letter for our example states that the agency has "determined the device is substantially equivalent (for the indications for use stated in the enclosure) to legally marketed predicate devices" 1. That phrase — substantial equivalence — is the whole legal content of a 510(k). It says the new device is as safe and effective as a device already on the market, for the stated use. It does not say the device is the best available, that it improves outcomes, or that the FDA independently re-derived its safety from first principles.
The same letter draws the boundary explicitly: "FDA's issuance of a substantial equivalence determination does not mean that FDA has made a determination that your device complies with other requirements of the Act or any Federal statutes and regulations administered by other Federal agencies" 1. The regulator is telling you, in its own words, that clearance is a bounded finding. Everything else you might read into the phrase "FDA cleared" is inference, and the rest of the summary is where you check that inference.
It helps to place a clearance in its family. The device here sits in a category the FDA created specifically for triage software — "radiological computer aided triage and notification software" — with its own product code and regulation number 1. That category itself began with a De Novo authorisation for the predicate, which is why the predicate in this summary carries a "DEN" number rather than a "K" number 1. The lineage matters when you read any summary in the group: the evidence expectations were set once, at the front of the chain, and later devices clear largely by resembling what came before. Knowing where a device sits in that lineage tells you how much of its assurance was inherited and how much was freshly demonstrated.
Section 1 — the indications for use
The indications for use are the most important sentences in the document, because they define exactly what the device was cleared to do. In our example, the device "is indicated for use in the analysis of non-enhanced head CT images" and "is intended to assist hospital networks and trained radiologists in workflow triage by flagging and communication of suspected positive findings" of intracranial hemorrhage 1. That is a triage-and-notification claim: it moves a suspicious scan up the reading queue. It is a narrow thing, and the summary is emphatic about the limits.
The document states that the device "does not alter the original medical image and is not intended to be used as a diagnostic device," that its preview images are "meant for informational purposes only and not intended for diagnostic use beyond notification," and that "notified clinicians are responsible for viewing full images per the standard of care" 1. Read those lines as the operative scope. A tool cleared to flag a case for faster review has not been cleared to diagnose, to rule out, or to replace the read. When a claim in a sales conversation runs ahead of the indications on the summary, the summary wins.
Section 2 — the predicate
Because a 510(k) is a comparison, every clearance rests on a named predecessor. Our example was cleared against a specific predicate device — "Viz.AI's ContaCT (DEN170073)" — and the summary's reasoning is equivalence, concluding that "the minor differences between the subject device and the predicate raise no new issues of safety or effectiveness" and that the device "is thus substantially equivalent to the ContaCT predicate" 1. Notice what the argument is: the new device does roughly what the predicate did, in a comparable way, so it inherits the predicate's market status.
This is where a careful reader follows the chain. The predicate itself was cleared against something, and so on back through a lineage. Analysts have mapped these predicate networks precisely because equivalence can drift: a series of small, individually reasonable steps can end far from where the original evidence sat. The question to carry is simple — what did the earliest device in this chain actually prove, and how much has changed since. A predicate relationship is a statement about resemblance, and resemblance is not the same as independent validation of the device in front of you.
Section 3 — the performance data
Now the numbers. The summary reports a "retrospective, blinded, multicenter, multinational study" of "198 cases from 3 clinical sites (2 US and 1 OUS)" with "approximately an equal number of positive and negative cases," in which "sensitivity was observed to be 93.6% (95% CI: 86.6%-97.6%) and specificity was observed to be 92.3% (95% CI: 85.4%-96.6%)," exceeding a pre-set 80% performance goal 1. Those are strong numbers, and it is worth being precise about what kind of number they are.
| Element of the study | What the summary reports 1 |
|---|---|
| Design | Retrospective, blinded, multicenter, standalone |
| Sample | 198 cases, 3 sites (2 US, 1 outside US) |
| Case mix | Roughly equal positive and negative |
| Sensitivity | 93.6% (95% CI 86.6–97.6) |
| Specificity | 92.3% (95% CI 85.4–96.6) |
| Performance goal | 80% |
| Secondary endpoint | Time-to-notification 4.5 min vs 72.6 min standard time-to-exam-open |
Three features of that table decide how far the numbers travel. First, the study is retrospective and standalone — the algorithm was run on collected images against a reference standard, so it measures the model in isolation, not the clinician-plus-model system that will actually operate. Second, the case mix is enriched: roughly half the cases are positive, whereas real head-CT streams carry hemorrhage at a far lower rate, and sensitivity and specificity measured on a balanced set do not translate directly to the positive and negative predictive values a clinic experiences at true prevalence. Third, the sample is 198 cases across three sites — enough to clear the bar, and modest against the diversity of scanners, protocols, and patients a nationwide deployment meets. The reasoning for why prevalence and case selection reshape every downstream number is set out in sensitivity, specificity, and AUROC and external validation.
None of this makes the clearance improper. It makes the number legible: a standalone retrospective accuracy on an enriched sample, which is a legitimate basis for triage clearance and a weak basis for any claim about diagnostic or outcome benefit.
What clearance does not establish — and how common that is
The single summary generalises. Three peer-reviewed analyses show that the pattern you just read — clearance on retrospective, standalone, single-or-few-site evidence — is the rule for AI devices rather than a quirk of one example.
- Most AI devices take the lower-risk route. A comparative analysis identified "222 devices approved in the USA" over 2015–20, of which "few were qualified as high-risk devices" 2. The higher-evidence premarket-approval pathway is the exception; clearance by equivalence is the norm.
- The evidence is rarely prospective or multi-site. An analysis of 130 cleared AI devices found that "almost all of the AI devices (126 of 130) underwent only retrospective studies," with just 4 evaluated prospectively and 37 reporting a multi-site evaluation 3. The gap between a retrospective clearance and prospective, multi-site performance is exactly where deployed models disappoint.
- The summary often omits the numbers you most need. Among machine-learning devices authorized in 2024, "only 29.2% reported both sensitivity and specificity" and 15.5% reported demographic data on the study population 4. Calibration and subgroup performance are frequently absent, and a Predetermined Change Control Plan — the mechanism that authorises post-clearance updates — appeared in a minority of summaries.
Read together, these set the prior for any clearance summary: assume the evidence is retrospective and standalone until the document proves otherwise, and treat missing calibration, subgroup, and update information as missing, not as reassuring.
A four-point reading checklist
For any 510(k) summary, in order:
- The indication. What exactly is it cleared to do — triage, detect, measure, diagnose — and what does it explicitly state it does not do?
- The predicate. What is the named predicate, and what did the earliest device in that chain actually establish?
- The study behind the number. Retrospective or prospective? Standalone or clinician-in-the-loop? How many sites, how many cases, and how was the sample's case mix assembled relative to real prevalence?
- What is missing. Is there calibration? Subgroup performance? A change-control plan? The absence of a number is information.
Run those four points and the summary stops being a badge and becomes a document you can weigh. The device in this worked example is a capable triage tool with a clean, candid summary: it flags suspicious scans faster, states plainly what it does and what it stops short of doing, and reports its numbers with confidence intervals. Reading it well means holding both truths at once — the clearance is genuine, and it certifies exactly what it says and no more. The failures happen when a reader collapses those two into a single word of reassurance and stops there.
How to read this
Two cautions travel with this guide. First, a single 510(k) summary is a worked example, chosen because it is public, well-structured, and typical; the specific figures describe that device and are fixed to its 2018 record, while the field-level counts are perishable and carry the date they were pulled. The FDA's living device list keeps growing, so treat any total as a snapshot. Second, clearance is a market-entry signal rather than a validation verdict — the reading checklist here assesses what the regulator decided, and the separate question of whether a study is any good is the subject of how to read an AI validation study. Because clearance status, indications, and change-control obligations carry legal weight, confirm any procurement, deployment, or billing decision that turns on a clearance with your regulatory or compliance counsel before acting.
Sources and method
This guide walks one public 510(k) summary as its primary regulatory document 1 and sets it against three peer-reviewed analyses of how AI devices are cleared and evaluated — a comparative approval analysis 2, an evaluation-methods analysis of 130 devices 3, and a transparency- reporting study of 2024 authorizations 4 — with the regulator's own living device list for context 5. Every quotation and figure is drawn from the primary source cited beside it. We revisit this page on a 180-day cycle and whenever the FDA revises its device framework or updates the referenced record. For the mechanism that governs post-clearance model updates, see predetermined change control plan.