Plan to certification
Certify the 30-second strip claim first; what is met, what is missing, and the order of work.
Scope
A 30-second single-lead ECG strip from the sensor is classified as AF, not AF, or inconclusive. No continuous-monitoring or episode claim in the first submission.
Targets and status
| Metric | Proposed acceptance | Current | Status |
|---|---|---|---|
| Strip sensitivity, held-out patients | >= 95 %, lower 95 % CI >= 90 % | 98.7 % | met |
| Strip specificity | >= 97 %, lower 95 % CI >= 95 % | 98.6 % | met |
| Inconclusive rate | <= 15 % on a clinical-grade set | 31.8 % on an artefact-heavy set | needs a clean study set |
| AFDB duration Se / +P | >= 95 % / >= 90 % | 96.5 % / 92-94 % | met |
| AFDB episode Se, episodes >= 30 s | >= 95 % | 100 % | met |
| Beat detection Se / +P | >= 98 % / 98 % | 98.1 % / 98.7 % | met |
What remains, in order
- Cardiologist re-annotation of the existing sensor strips (two readers, adjudication, interval level).
- Clinical validation study on the device: simultaneous 12-lead reference, blinded reads, a few hundred subjects enriched for AF, pre-specified endpoints.
- Freeze and version the algorithm: repository, tag, model hash, unit tests, design description.
- Written validation protocol, then a single pre-specified run on AFDB and on the held-out sensor patients, scored with the WFDB tools.
- Software documentation, ML data and model description, labeling; human factors if consumer-facing.
Pending experiment: engine-native features on AFDB (P-wave, QRS width, quality), expected to raise long-recording episode precision; relevant only to a later continuous-monitoring claim.