Methodology
How we appraise the evidence
Before trusting any single verdict on this site, read how it was produced. This page explains the scorecard framework, why FDA clearance status is never treated as clinical validation, and how the tracker sources and limits its dated entries.
The risk-of-bias scorecard
Every appraisal closes with a structured scorecard rather than a prose-only conclusion, following a PROBAST/TRIPOD+AI-style structured-judgment pattern: study design, sample size and setting, external or prospective validation, and an overall risk-of-bias rating. Rendering that rating as an instrument dial — rather than a colored badge — is deliberate: it is a measured judgment against a fixed set of criteria, not a marketing claim.
Well-conducted study design, adequate sample, external validation present.
Some concerns in design, sample, or validation that limit confidence.
Serious methodological concerns — retrospective-only, no external validation, or undersized samples.
Vocabulary used in the scorecard is defined in the Glossary, including external validation and risk of bias. See a scorecard applied in practice on any appraisal.
Clearance and evidence are two separate facts
FDA clearance confirms that a device may legally be marketed under a given pathway — 510(k), De Novo, or PMA — for a stated intended use. It does not confirm that the device's clinical claims have been independently validated in peer-reviewed evidence. Every appraisal on this site reports clearance status and the risk-of-bias verdict as two structurally distinct fields, front-loaded above the narrative, so the two claims are never read as one.
See this separation applied on any appraisal's readout panel, or read the related FAQ answer on whether FDA clearance equals clinical validation.
Tracker sourcing and limitations
The FDA Clearance & Incident Tracker is sourced from FDA's AI-enabled device list, MAUDE adverse-event reports, and secondary hazard authorities such as ECRI-style hazard reporting. Every entry is dated and cites its primary source.
- Under-reporting. MAUDE is a self-reported system and is known to under-count adverse events; absence of an incident entry does not mean absence of a problem.
- Unconfirmed causation. Reported-incident entries describe what was reported, not a confirmed finding of AI-model fault. The tracker does not editorialize on causation.
- Update cadence.FDA's device list updates periodically rather than in real time, so tracker entries reflect the most recent available source data, not a live feed of regulatory action.
Open the Tracker to see this methodology applied, or subscribe to monitor new entries as they publish.