How the 510(k) Report is built

Sources and cadence

The data is reloaded from FDA’s published files every week, so each source is as current as FDA’s own files; the last refresh was Nov 3, 2026. Each page states how far each source runs.

• FDA 510(k), De Novo and PMA databases: decision, dates, applicant, product code, class, regulation and review panel for 233,963 decisions.
• FDA 510(k) summaries: the summary PDF filed with each clearance (92,896 read). Pages without a text layer were read with optical character recognition.
• FDA recall and enforcement reports: recall class, reason and root cause, matched to clearances through the 510(k) numbers FDA lists on each recall (41,646 recalls matched).
• MAUDE: 25,781,414 adverse event reports, counted per product code and year. Counts are reports naming a product code; they are not rates and are not matched to individual devices.

How a family tree is built

A 510(k) is cleared by showing the device is substantially equivalent to a legally marketed predicate. Each summary names its predicates; reading every summary gives each device its parents, and following parents back gives its ancestors.

• Predicate: a device the submitter claimed equivalence to. Reference device: one cited for a specific test or feature, without a claim of equivalence.
• Line of descent: the longest chain of predicates behind a device, following at each step the predicate with the deepest ancestry.
• Generation: one step along that chain. Ancestors: every device reachable through predicates. Descendants: every later device that reaches this one.
• De Novo grants and PMA approvals do not claim predicates, so their family trees start with them; later 510(k)s can still cite them.

Accuracy

Predicate citations were measured against an answer key of 200 FDA summaries, sampled across 1996–2026 (including 20 scanned 1990s documents and 10 De Novo decisions). The key was prepared separately from the extraction, by a second AI reading of each document’s text and page images; it has not yet been audited by a person. Against that key, the citations kept in the family trees are 100.0% precise (none was wrong) with 96.0% recall (4% of the key’s citations are missing).

Measured October 2026. Every number read from a summary is checked before it enters a family tree: it must be in FDA’s 510(k), De Novo or PMA lists and appear in the summary text, and a 510(k) predicate must have been decided no later than the citing device. A predicate FDA cleared while the citing submission was under review is kept, and marked on the page. Citations failing a check stay on the device page, marked –, with a note saying what FDA’s record is, and are left out of the tree.

Findings

Recalls run in families. Among 510(k)s cleared 2010–2019 that cite at least one predicate, devices whose direct predicate had a Class I recall were 1.6 times as likely to have a Class I recall of their own (in a separate recall event), compared with devices of the same product code whose predicates had none. The ratio is a Mantel-Haenszel risk ratio stratified by product code; for 2005–2019 it is 1.9. Without controlling for device type, the ratio is 16.7, which mostly reflects that some device types are recalled far more than others.

Coverage

• FDA does not post a summary or statement for every clearance: almost none before 1996, and for 2000–2001 only about a third. Those devices show no predicates and say why.

  510(k)s decided	Clearances	FDA posts a document	Read for predicates
  1976–1995	75,477	0%	0%
  1996–1999	16,042	82%	82%
  2000–2001	6,937	38%	38%
  2002–2009	25,956	98%	98%
  2010–2019	30,270	99%	99%
  2020–2026	21,369	99%	98%

• PMA summaries are rarely posted in a form we can read, so PMA pages show decisions, supplements and recalls but little text.
• About 10,000 older recalls carry no class in FDA’s enforcement file and are shown as not classified.

Accessibility

These pages aim to meet WCAG 2.2 AA in both colour themes. Each release is checked with axe-core at 1440, 1024, 390 and 320 pixels wide, by keyboard alone, for horizontal scrolling at 320 pixels, and through Chromium’s accessibility tree. They have not yet been tested with a screen reader such as NVDA or VoiceOver. Every chart has a table with the same numbers, and recalls are marked by shape and word as well as colour. If something does not work for you, use “Report an extraction error” on any device page to tell us.

Colours and what they mean