Public analysis code and model-access boundary
8 October 2026

The original source files are supplied unchanged. No fitted model, complete reconstruction parameter file or individual-level data is supplied. To independently reproduce the analysis, obtain your own authorised access to the specified source databases, rebuild the required pair/cache files and configure an isolated workspace. Follow source_extraction/README_source_extraction.txt and the frozen protocol. Keep the work/oxygen_formal and work/oxygen_feasibility relative structure when unpacking reproducibility_code.zip. Dependencies are recorded in requirements.txt; paths in the original scripts require local configuration.

The original formal workflow is prepare.py -> freeze.py -> ehr.py -> test_formal.py -> train.py -> evaluate.py -> validate.py -> figures.py. Inspect the scripts and input dependencies before running. Training creates a new local frozen_models.joblib and lock; validation can export a new local model_coefficients.csv. Those generated materials remain subject to source-data sharing conditions. Do not overwrite the supplied historical protocol/model locks. They document the authors' original run; a new run must maintain its own provenance and is not guaranteed to reproduce the same serialized-file hash across environments.

The evaluate.py and validate.py references to frozen_models.joblib are expected: these files consume a locally fitted model. Such references do not indicate that the model is embedded in the source archive. The feature metadata retain predictor names, training counts, selected regularisation strengths and age-imputation medians already described in the methods; they do not provide complete fitted weights or scaling arrays.

report.py and package.py are historical reporting/packaging utilities, retained for provenance. Their generated local reports can refer to locally created model files, and their original packaging filter excludes individual records but does not exclude fitted models or complete parameters. Do not use that historical packaging function to prepare a public journal attachment. The current Additional file 2 was rebuilt with a separate distribution filter and is inventoried by CONTENTS_SHA256.json.

Exact replay of the authors' original fitted predictions is not available from this public supplement alone. frozen_models.joblib and model_coefficients.csv are retained locally; no hosted model resource is currently claimed. Any future sharing would require a PhysioNet release under the source-data agreement. No claim of a deployable clinical tool is made.
