Post hoc respiratory-support extension (FiO2, PEEP, ventilation status)
21 October 2026

locks/FiO2_Extension_Lock_v1.zip and locks/FiO2_Extension_Lock_v1_1.zip are the unchanged pre-fitting lock packages (analysis_lock.json SHA-256 v1.0 aceea519836407271d91578574b4bf0697fc4cb575ed2dffd1c7bdc999e28959; v1.1 d1307d434c3ae68fcbc75091ad6aa010163213fa6cf9f0ba86532f424677b76a). v1.1 changed only the ventilation-state rule before fitting. coverage_audit/ holds the descriptive FiO2 coverage audit that preceded the lock.

results/ holds the executed analysis: step0/ pre-fitting eICU ventilation-text audit; code/ the implementation snapshot hashed before the first fit; fit_start.json, model_lock.json, evaluation_started.json, evaluation_complete.json, execution_events.jsonl (hash-chained) and final_verification.json document the order lock < fit start <= model lock < evaluation start. Aggregate results: key_results_for_manuscript.csv, paired_differences.csv, model_metrics.csv, class_metrics.csv, calibration_bins.csv, error_coverage_curves.csv, reassurance.csv, hospital_differences.csv, feature_coverage.csv, feature_denominators.csv, S2_selection_descriptives.csv, training_cv_scores.csv, calibration_threshold_search.csv, coefficient_replay_checks.csv.

Excluded from this public supplement, as for the original models: frozen_extension_models.joblib, model_metadata.json. model_metadata_public.json retains each model's selected regularisation, feature order, training counts, fold sizes and imputation medians but omits coefficients, intercepts and scaling arrays. SHA256SUMS_original_results_package.txt is the authors' original checksum list; 46 of its 48 entries are present here and match byte-for-byte (the protocol/ entry is stored as locks/FiO2_Extension_Lock_v1_1.zip); the remaining 2 entries are the excluded files. Patient-level feature and prediction tables are retained locally and not distributed.

Interpretation boundaries: ventilation status was unknown for 71.8% of external measurements; S2 is a ventilation-enriched complete-case subset whose external interval includes zero; S3 is a retrospective adjustment; no model passed the 5% error gate.
