Supporting archive for: Artifact-specific stress testing of RR feature fusion in single-lead ECG rhythm classification

This Zenodo record is the supporting archive for the manuscript "Artifact-specific stress testing of RR feature fusion in single-lead ECG rhythm classification", submitted as a Research paper to Physiological Measurement. It contains five archives:S1 - analysis code (research_notes), independent audit scripts, frozen protocols and protocol chronology, eligibility manifests, source checksum and label audits, the training-derived spectral threshold, environment and hardware version reports, and the separately dated post hoc Challenge 2017 transfer rule;S2 - aggregate CSV/JSON result summaries, including three-seed run settings, per-class metrics and per-seed confusion matrices, spectral-fallback rates, all-15-condition trigger diagnostics, post hoc alternative-detector and true-single-lead checks, RR-only noise metrics, the descriptive sex-reporting audit, and paired patient-bootstrap intervals;S3 - three-seed trained PyTorch checkpoints for the final neural comparisons;S4 - paired PTB-XL per-record probabilities (waveform CNN, generic RR gate, safeguarded gate) on clean and all 15 recorded-artifact conditions, aligned by ECG identifier, truth label and patient identifier;S5 - independent Chapman-Shaoxing-Ningbo per-record probabilities for the three primary models, plus exploratory PhysioNet/CinC Challenge 2017 N/A probabilities. No raw ECG signals are redistributed here. To reproduce the analysis, obtain the exact public dataset versions identified in the manuscript (PTB-XL 1.0.3; Chapman-Shaoxing-Ningbo 1.0.0; MIT-BIH NSTDB 1.0.0; PhysioNet/CinC Challenge 2017), verify the publisher hashes, then follow the frozen protocol order documented in S1 and S2. Methodological notes: only lead II was extracted from clinical 12-lead recordings; NSTDB noise scaling uses per-record AC-RMS SNR rather than the PhysioNet nst QRS-power convention; the spectral cutoff is the 99th percentile of clean PTB-XL training waveforms; the safeguard was introduced after validation failures but frozen before test and external evaluation. OpenAI Codex assistance is disclosed in the manuscript.