clinicians.build interactive · July 29, 2026

Cleared Before Measured

Yesterday npj Digital Medicine published the number almost nobody publishes: what 13 physicians could do without the AI. The tool in that study — Sonio Suspect — was FDA-cleared 522 days earlier. Here is every imaging-AI clearance since 2016, one dot each, and the column that would tell you which ones ever ran that arm.

Primary source: Chalouhi et al., npj Digital Medicine — multireader fetal ultrasound study (Jul 28, 2026)
Device data: FDA 510(k) release, latest vintage via MIMI Labs
369
imaging-AI 510(k) clearances, 2016–Jul 2026
184
distinct companies behind them
20%
are detection/diagnosis codes — the rest triage or measure
522
days from Sonio's clearance to its unassisted-arm paper

369 dots. None of them carries an evidence field.

Each dot is one 510(k) clearance in an imaging-AI product code: the date FDA decided, against how many days the review took. Color is what the device is allowed to claim — detection/diagnosis (it calls a finding), triage/notification (it flags a case), quantification/processing (it measures or segments). The ringed dot is Sonio Suspect.

detection / diagnosis (74) triage / notification (90) quantification / processing (194) Sonio Suspect
clearances in view
median FDA review
from repeat filers
busiest filer in view
Hover any dot for the device, company, K-number and review time. Now do the stress test: drag min clearances per company to 5. The cloud thins to the serial filers and the median review drops from 140 days to 123 — that is not FDA getting faster, that is the same handful of companies filing incremental updates on a path they already know. Now slide it back to 1 and tick first clearance from each company only: the median climbs to 172 days. Same agency, same years, opposite story.
80/20 — read the ledger like a builder

There is no column for the baseline. A 510(k) record tells you the device, the predicate, the panel and the calendar. It does not tell you whether a human ever read the same cases without the tool. Sonio Suspect cleared on 2025‑02‑21 in 91 days; the multireader study reporting 54.2% unassisted sensitivity landed 522 days later, on 2026‑07‑28. For the other 368 dots you cannot tell from this ledger whether that arm exists, because the ledger was never designed to ask.

Four in five of these devices never had to face the question. Only 74 clearances sit in a detection/diagnosis code. 194 — more than half — are QIH, "automated radiological image processing": segment it, measure it, hand it back. The reader-study design that produces an unassisted arm is the natural fit for the 20% lane, and the boom you keep reading about is mostly the other 80%.

The market is 65 companies deep, not 184. 250 of 369 clearances (68%) come from firms with more than one. Aidoc alone has 27 — 7% of the entire landscape, at a median 81-day review against 180 days for the one-and-done filers. If you are a first-time filer budgeting off the headline median, you are budgeting off somebody else's tenth submission.

😤 "2026 is falling off a cliff — the boom is over." No: this vintage of the FDA release stops at July 17, 2026. Fifty-four clearances in 6.5 months annualizes to roughly 100, which would be the biggest year on record. Partial-year bars are the most common way to manufacture a trend, and they are why the last bar below is drawn hollow.
😤 "Product code is a lazy proxy for 'AI'." It is a proxy, and it has both errors. It sweeps in image-processing software with no learned component, and it misses AI that cleared under an older generic code or came through De Novo or PMA instead. Treat 369 as the shape of the imaging-AI 510(k) lane, not a census of clinical AI.
😤 "Review time isn't quality — long reviews might just be bad submissions." Agreed, and that cuts both ways: a 28-day Special 510(k) is not evidence of a well-studied device, it is evidence of a small change to one you already cleared. Neither axis here measures whether the thing works. That is the point.

What a truncated last bar does to a trend

Clearances per year in this same set. The 2026 bar is hollow because the data stops mid-July; the dashed cap is the straight-line annualization. Tick the box and watch the "decline" become a record year.

Same 369 clearances, grouped by FDA decision year. 2016–2019 are thin partly because the AI-specific product codes did not exist yet — QAS was created in 2018, QIH in 2019 — so early years undercount by construction as well.
Read the study that published the unassisted arm →