01Executed software pilotWhat has been done
Exploratory software experiments used 4,800 annotation-centred windows from six MIT-BIH records. A follow-up used whole-record offline WFDB XQRS and one-to-one matching of detected and reference events. The experiments reveal validation failure, class concentration and dependence on beat detection. They demonstrate a functioning experimental workflow, not patient-level diagnosis, independent clinical effectiveness, a Haar–Walsh advantage or measured ARM performance.
02Planned evaluationPatient grouping before model comparison
MIT-BIH v1.0.0 is the development resource. Records 201 and 202 belong to one subject and must remain together. The historical DS1/DS2 split is retained only as a documented reproduction mode; the corrected mode moves 202 into development with 201. Patient-grouped inner and outer folds will separate fitting, selection and evaluation of the selection procedure. The already-inspected legacy test is not a pristine confirmation set.
03Planned evaluationTask and representation controls
The planned primary binary target uses AAMI-aligned groups: N = {N,L,R,e,j} versus S/V/F, with excluded paced and unknown annotations counted separately. This differs from the historical annotation-N-versus-other task. RR-only, waveform-only and combined inputs will be compared using common data access and tuning budgets. Haar–Walsh descriptors, matched db4 DWT/SWT representations and a compact 1D CNN are candidates. Broad normal complexes, V-versus-N and S-versus-N analysis will test what morphology adds beyond past-interval RR.
04Planned evaluationQuality, rejection and calibration
The prespecified primary comparison tests a learned quality gate against classifier-confidence rejection at matched coverage. Reports will include classification risk, coverage, class-specific rejection, missed detections and false detections. Operational rules and ungated predictions provide further controls. The existing gate uses synthetic corruptions; it has not demonstrated quality generalization to real acquisition artefacts. Calibration and feature-group contributions will explain fitted model scores without claiming causal physiological mechanisms.
05Planned evaluationA frozen external evaluation
INCART is planned as the external evaluation after all models, thresholds and preprocessing decisions are frozen. Excerpts from the same patient or source Holter recording stay together. External labels will not tune the pipeline. Lead configuration, 257 Hz sampling, prevalence and annotation differences will be documented. New clinical recruitment and hospital integration require a separate future protocol.
06Planned evaluationFrom files to an embedded prototype
The planned Raspberry Pi 5 prototype replays recorded waveforms, detects QRS events, assesses quality, predicts and exposes feature contributions. Measurements include detection, filtering, normalization, feature extraction and inference, alongside p50/p95 latency, peak memory and complete artefact size. Warm-ups, timed iterations, threads, cooling and software versions will be recorded. Buffering and look-ahead are separate from compute time; offline XQRS does not establish causal streaming.
07Planned evaluationReproducibility and current status
Planned outputs are versioned patient manifests, configurations, dependency locks, model cards, trained artefacts, prediction and abstention logs, timing records and reproducibility instructions, subject to data and dependency licences. This is a research proposal with preparation evidence. Funding, external validation, device deployment and international partnerships are not presented as secured or completed.
Lower classification risk must be interpreted together with coverage and rejection by class.
