The evidence FDA actually accepted.
Every parsed AI/ML 510(k) summary as structured, source-quoted fields — validation design, endpoints, predicate chains — connected to what happened after clearance and how devices got paid.
Presence rates, never pooled.Every figure below is “reported in X of Y audited devices.” Performance values are never averaged across devices — each stays with its analysis unit, task, and a verbatim FDA quote. A device we haven’t parsed yet is queue, never “FDA accepted thin evidence.”
What evidence did FDA accept for devices like yours?
Keyword-based retrieval over the parsed corpus (names, algorithm descriptions, endpoints, source quotes) — not semantic search. Try: · ·
Corpus coverage — Radiology
The hatched segment is our parse queue, not device data. Queued devices never count in any rate below — every denominator is audited devices only.
Original findings from the corpus
Median FDA review time, received → decision
Across all 1,146 matched AI/ML 510(k)s; the slowest decile takes 8.5+ months (p90 ≈ 260 days). Computed from openFDA decision dates — not published anywhere else for this cohort.
PCCP adoption in 2025 AI clearances
Up from 0% before 2023 (5.0% in 2024). Nine in ten AI/ML clearances still file without a Predetermined Change Control Plan — a live gap as FDA pushes the mechanism.
What the audited corpus reports
Evidence reporting
Canonicalized — includes per-finding sensitivities, not just the summary's top-line slot.
“Not reported” is a real finding about what FDA accepted — the summary was audited and does not state it.
Metric presence (canonicalized)
- sensitivity284 of 1143
- specificity239 of 1143
- auroc163 of 1143
- ppv72 of 1143
- npv36 of 1143
- accuracy125 of 1143
- f125 of 1143
- dice236 of 1143
- iou10 of 1143
- detection_rate5 of 1143
Presence only — values are never pooled or averaged across devices. Each measurement lives on its device's page with its analysis unit and verbatim quote.
Median predicate age: 1.9y across 1427 of 2061 predicate edges (69% datable, dates backfilled from openFDA) — how old each cited predicate was when the child device cleared.
Study design & generalizability
Where FDA review is tightening — and what almost no premarket dataset reports. Each rate is presence over audited devices; a device that describes no clinical study is honestly “not reported” here.
Primary validation is prospective-clinical or a multi-reader design — not a retrospective bench run.
Validation data explicitly independent of the training/development set.
Tested across more than one scanner or device manufacturer.
Performance reported per subgroup, not just an overall number.
One corpus, whole lifecycle
Clearance evidence
1143 devices parsed with source-quoted fields
Postmarket signals
1,146 devices snapshotted · 777 of 1146 carry a drift signal
Reimbursement pathways
5 pathways across 4 devices (seed corpus, growing)
Every device page connects all four stages: clearance → evidence → postmarket → payment. Stages without data render as “not yet tracked”, never as empty or fabricated.
Machine access
The same retrieval over MCP — 12 tools across the lifecycle
curl -s https://constat.dev/api/mcp \
-H 'content-type: application/json' \
-d '{"jsonrpc":"2.0","id":1,"method":"tools/call",
"params":{"name":"evidence_search",
"arguments":{"panel":"Radiology","reports_any_sensitivity_metric":true}}}'device_evidence_lookup, evidence_search, predicate_chain, device_postmarket_lookup, reimbursement_lookup, and more. Every extracted field carries a source quote and page — descriptive only, never a compliance judgment.
Follow the corpus as it grows
New parses land weekly, most-recent-first. Get notified as coverage expands and presence rates shift — and be first on the client digest when it ships.