Transportability of a Released Deep-Learning Electrocardiogram Model for Atrial Fibrillation Across German and Chinese Cohorts: Reproducibility Package
Description
This record contains the reproducibility package for a frozen, zero-training external evaluation of the released Ribeiro et al. six-output 12-lead deep neural network's atrial-fibrillation output in PTB-XL version 1.0.3 and the Chapman-Shaoxing/Ningbo ECG database version 1.0.0. The package contains analysis and inference scripts, tests, a frozen protocol and design lock, software requirements, derived cohort and label audits, complete prediction tables, bootstrap and statistical outputs, figure source data, publication figures, tables, and manifests. No PhysioNet ECG waveforms, waveform arrays, WFDB headers, released model binaries, inference batch caches, credentials, direct identifiers, or local-machine metadata are redistributed. Record-level tables retain only public de-identified source identifiers and fields needed for provenance, patient-cluster resampling, subgroup analyses, and the prespecified first-ECG sensitivity analysis. Users must preserve de-identification and comply with the cited upstream terms.
Files
Steps to reproduce
1. Download all three deposit files into one directory and verify archive integrity with: shasum -a 256 -c SHA256SUMS.txt. The expected result is frozen_ribeiro_af_external_validation_v1.0.0.zip: OK. 2. Unzip the archive and read README.md plus manuscript/reproducibility_README.md. Create a Python environment from environment/requirements_frozen.txt. 3. The archive already contains the frozen prediction tables and all derived outputs needed for the statistical rebuild. From the extracted project root run: .venv-ribeiro/bin/python -m scripts.analysis_pipeline --ptbxl-predictions results/inference/ptbxl_predictions.csv --chapman-predictions results/inference/chapman_predictions.csv --output results/analysis --threshold 0.390 --replicates 2000 --seed 20260821 --n-jobs -1. 4. Run the test suite with: .venv-ribeiro/bin/python -m unittest discover -s tests -p 'test_*.py'. The verified baseline contains 44 tests and must end with OK. 5. Rebuild figures and tables using the commands in manuscript/reproducibility_README.md and compare row counts, source hashes, and manifests. 6. Full waveform-level reinference requires separately obtaining PTB-XL v1.0.3, Chapman-Shaoxing/Ningbo v1.0.0, and the Ribeiro pretrained model from the cited upstream DOI records. Raw waveforms and model binaries are intentionally not included. Verify all frozen hashes before inference.