Gap analysis and to-do list

Scope: 510(k) for software that classifies a 30-second single-lead ECG strip from the sensor as AF, not AF, or inconclusive. Predicate category 21 CFR 870.2345. Status as of 14 September 2026. Owners: Algorithm (this work), Corsano (regulatory, quality, clinical), Developer (original engine author).

Gap analysis

#	Submission element	Exists today	Gap	Owner	Status
1	Intended use and indications	wording proposed	formal statement agreed with regulatory	Corsano	open
2	Predicate and substantial-equivalence table	candidates identified (Apple, Fitbit ECG apps)	comparison table	Corsano	open
3	Frozen algorithm version	experiment scripts and model files on elon	repository, tag, model hashes, pinned build, clean strip-in / decision-out interface	Algorithm	in progress
4	Software requirements and design description	pipeline description, feature and threshold tables (RESULTS.md, methods page)	formal SRS / SDS for the frozen version	Algorithm	drafts exist
5	Model card and training procedure	hyperparameters, features, thresholds, tuning method documented	consolidated model card	Algorithm	drafts exist
6	Data documentation	dataset table, split rules, audit findings	provenance, licence, consent and annotator record per source; sensor set re-annotation record	Algorithm + Corsano (clinic, cardiologists)	partial
7	Unit and integration tests, traceability	none	tests with expected values, bit-for-bit reproduction on a fixed strip set, requirement-to-test matrix	Algorithm	open
8	Validation protocol, pre-specified	targets proposed	written protocol: sets, frozen thresholds, acceptance criteria, CI method, WFDB scoring, subgroups	Algorithm	open
9	Algorithm validation report (bench)	feasibility results on AFDB, MIT-BIH, held-out sensor patients	single pre-specified run on the frozen version, EC57 tables via WFDB tools	Algorithm	after 3 and 8
10	Cardiologist re-annotation of sensor strips	export prepared (2,660 strips, PNG + WFDB, no labels)	two blinded readers, adjudication, interval-level labels	Corsano	open, critical path
11	Clinical validation study on the device	none	protocol, ethics, simultaneous 12-lead reference, blinded reads, a few hundred subjects enriched for AF, statistical report	Corsano	open, longest item
12	Study analysis script	none	endpoint computation from reads and algorithm output, tested before unblinding	Algorithm	after 8
13	Risk management inputs	failure-mode list in RESULTS.md	hazards integrated into the ISO 14971 file	Algorithm (inputs) + Corsano (file)	partial
14	Labeling performance section	numbers exist	final numbers from 9 and 11 with CIs, inconclusive rate, exclusions	Corsano, after 9 and 11	open
15	Human factors (if consumer-facing)	Corsano process	study on interpretation of AF / not AF / inconclusive	Corsano	open
16	Cybersecurity, SBOM	Corsano process; dependency list exists	SBOM for the frozen build	Algorithm (SBOM) + Corsano	partial
17	Predetermined change control plan	decision pending	retraining procedure and re-test criteria, if retraining after clearance is planned	Algorithm + Corsano	decision needed
18	Beat detector provenance	engine code and models delivered; open questions to the developer	answers on unannotated code state, training environment, IP assignment	Developer + Corsano	open
19	Pre-submission meeting (Q-Sub)	none	confirm predicate, endpoints and sample size with FDA before the study	Corsano	recommended

To-do list, Algorithm owner

Freeze: create the repository, move the pipeline into a versioned module, pin dependencies, tag, record model hashes. Deliverable: tagged release with build instructions. Progress 15 Sep: freeze manifest generator written (model SHA-256, engine commit, feature list, hyperparameters, decision rule); repository and tag pending.
Tests: unit tests for features, plausibility gate, guard, thresholds, strip decision; integration test reproducing a fixed strip set bit for bit; traceability matrix. Progress 15 Sep: 12 unit tests written and passing (RR features incl. the new lag features, plausibility gate, window labels, hysteresis, gap merging, EC57 scorer, guard, strip rule, model determinism and monotonicity); integration test and traceability matrix pending.
Validation protocol: write and freeze before any further run; include the WFDB cross-check of the EC57 scorer. Progress 15 Sep: draft protocol page published for review; EC57 scorer cross-checked against the reference epicmp tool (agreement within one point, see results).
Pending experiment: engine-native features on AFDB (running); decide whether morphology features enter the frozen version. Also running: PhysioNet/CinC 2017 (8,528 public single-lead strips) as a second strip-level test set.
Documentation: SRS, design description, model card, data documentation with the SCI and sensor annotation records, SBOM, failure-mode list for the risk file.
Final validation run on the frozen version; algorithm validation report.
Study analysis script, tested on synthetic data.
Support the clinical study with the frozen algorithm interface.

To-do list, Corsano

Agree intended use and predicate; schedule a Q-Sub.
Commission cardiologist re-annotation of the exported strips (two readers, adjudication).
Secure data documentation from the clinic (SCI) and consent basis for the sensor set; obtain the developer’s answers and IP assignment. Added 15 Sep: obtain the annotation backend repository (ecg-base), which holds the expert-label audit trail and the training folders; the web annotation tool and the beat-classifier training service have been reviewed.
Plan and run the clinical validation study; statistician for sample size.
Risk file, human factors, cybersecurity, labeling, eSTAR assembly.

Order and rough timing

Weeks 0-3: freeze, tests, protocol (Algorithm); re-annotation commissioned, Q-Sub requested (Corsano). Weeks 3-6: documentation, SBOM, study script (Algorithm); annotation delivered, study protocol and ethics (Corsano). Weeks 6-8: final validation run and report (Algorithm). Months 3-8: clinical study, then labeling and submission (Corsano).