In 2018, IDx-DR earned the first autonomous AI diagnosis authorization by enrolling 900 patients in a prospective trial across 10 primary care sites. In 2022 and 2023, two competitors updated their diabetic retinopathy algorithms and cleared 510(k)s with zero new patients. They reran old data and went to market.
That is not a loophole. It is how the predicate system is supposed to work, and if you are building AI SaMD, it should change how you think about the study you are about to fund.
We pulled the AI/ML universe from our FDA Device Explorer corpus: 1,483 records, 43 of them De Novos that founded new device categories. We assigned each follow-on 510(k) a generation number based on its predicate-chain distance from the founding De Novo, then verified five categories line by line against FDA decision summaries. We may have missed devices whose summaries do not report enrollment cleanly, so treat the counts as a floor on precision, not a census.
Diabetic retinopathy screening is the cleanest slope. IDx-DR's De Novo took 900 prospectively enrolled patients. The second generation paid less: EyeArt cleared with 655, AEYE-DS with 531. By the third generation, IDx-DR v2.3 and EyeArt v2.2.0 both cleared on retrospective reanalysis of their own pivotal data.
Autism diagnosis aids ratcheted even harder. Cognoa's De Novo enrolled 425 children in a double-blind study; EarliPoint entered with 500; then EarliPoint cleared twice more, in 2023 and 2025, by reprocessing the same pivotal dataset, and Cognoa's Canvas Dx update cleared on its original De Novo validation plus a PCCP. Four consecutive clearances in the category, zero new enrollment.
Cardiac ultrasound guidance shows the anchoring mechanism. Caption Guidance's pivotal had nurses scan 240 patients. UltraSight cleared with 240. HeartFocus cleared with 240. The predicate's design became the specification, until UltraSight's second-generation product cleared on bench testing alone against a retrospective 134-patient dataset.
And the largest families never paid a prospective toll at all. ContaCT founded CADt triage on retrospective CT reads and now has 123 descendant clearances; OsteoDetect's radiograph CADe lineage has 43, QuantX's CADx lineage 22. Categories born retrospective start at the floor.
Colonoscopy polyp detection went the other way. GI Genius founded the category with a 263-patient randomized trial in Italy; SKOUT followed with a 1,359-patient US RCT, CADDIE with 841, MAGENTIQ-COLO with 950. A CADe device promises a clinical benefit, more adenomas found per colonoscopy, and a benefit claim drags a randomized clinical endpoint behind it. There is no retrospective shortcut to inherit.
Sleep apnea wearables broke the pattern for a different reason. Samsung's De Novo enrolled 620 subjects; Apple then validated its notification feature on 1,499. When the sponsor is a consumer giant validating a flagship feature, the marketing value of a definitive study outweighs the regulatory minimum. The ratchet describes what FDA will accept, not what every sponsor will choose.
If you are the category founder, price the toll into the business plan and collect the rent. Your pivotal becomes the special controls and the anchor number every rival is measured against. That is a reason to run the leanest defensible design: every extra hundred patients you enroll is a standard you are gold-plating for competitors who will pay a fraction of it.
If you are a fast follower, the predicate's enrollment is your ceiling. Anchor your pre-sub proposal at or below that number and make FDA argue you upward. The retinopathy curve, 900 to roughly 600 to zero, is the empirical case that the agency accepts less as a category's safety record accumulates.
If you are iterating on your own cleared device, the cheapest patients are the ones you already enrolled. Reused pivotal data carried IDx-DR v2.3, EyeArt v2.2.0, and two EarliPoint clearances through without a single new consent form. Design your first pivotal with that reuse in mind, and add a PCCP so the next change may not need a 510(k) at all.
One caution: claim expansion resets the meter. When AEYE-DS added a handheld camera, it ran two new prospective studies totaling 679 patients, more than its original clearance.
Before you commit to a validation budget, find your category's generation curve and decide where you sit on it. We map predicate lineages like these for clients weekly inside FDA Device Explorer, and the difference between a generation 1 and a generation 3 evidence package is routinely seven figures.
Data from FDA Device Explorer by Innolitics (fda.innolitics.com), verified against FDA 510(k) and De Novo decision summaries, August 2026.