Additional file 2 — Public code, aggregate results and reporting materials
Journal of Intensive Care collection revision, 21 October 2026

This supplement contains the analysis protocol, extraction/training/evaluation code, model specifications, aggregate results and reporting maps. It excludes fitted model objects and the complete coefficient/intercept/scaling files that reconstruct them. No patient-level source or derived records, individual predictions, or access credentials are included.

aggregate_results/: unchanged model performance, calibration, paired contrasts, coverage, time-stratum and hospital summaries. Calibration intercepts and slopes are descriptive results and remain included.
descriptives/: unchanged patient-summary tables, category distributions, cohort flow and aggregate code/label inventories; these are not individual source records or clinical adjudication. See README_code_inventory_finalisation.txt for the earlier removal of review-workflow columns.
reporting/: TRIPOD+AI (52 entries), abstract (13 entries) and STROBE (22 entries) maps. All 87 page/line locations were regenerated from the current 27-page main manuscript and 13-page supplement of the Journal of Intensive Care revision. The abstract contains 342 whitespace-delimited words excluding headings and keywords. Items 12g and 22 acknowledge that fitted models and complete parameters are not publicly supplied. The map is not clinical validation or an independently issued ethics determination.
reproducibility/: frozen protocol, model-lock hashes, feature metadata, selected regularisation settings, software versions and original analysis code. The age-imputation medians and other method settings are retained; metadata do not contain the complete fitted coefficients and preprocessing arrays needed for inference. model_lock.json documents the original model hash, not a model included here.
reproducibility/source_extraction/: original extraction sources. Read README_source_extraction.txt and README_reproduction.txt before independently reproducing the analysis with authorised source data.
figures/: captions, unchanged threshold-grid denominators and original figure-generation code. The five figures are supplied separately with the submission.

Access boundary
frozen_models.joblib and model_coefficients.csv are retained locally and are not part of this supplement. They have not been deposited as an accessible model resource. Any future sharing would require a PhysioNet release under the source-data agreement, consistent with the MIMIC-derived model guidance. No model DOI, repository availability or direct distribution on request is claimed.
Guidance: https://physionet.org/content/mimiciv/3.1/ (Guidelines for creating datasets and models from MIMIC, 24 April 2024).
Code-sharing obligation: https://physionet.org/content/mimiciv/view-dua/3.1/ (item 9).
Sources checked 8 October 2026. This package does not establish that every regression statistic is a restricted model artifact; ordinary aggregate results are retained.

Reproduction and version scope
Reproduction requires the investigator's own authorised MIMIC-IV 3.1/eICU 2.0 access, rebuilt local caches, compatible software and isolated path configuration. The published code permits independently rebuilding the analysis; this supplement cannot directly execute inference with the authors' original fitted models. See reproducibility/README_reproduction.txt for the distinction.
Original unit-test and numerical-validation outputs describe the original analysis, not new testing or refitting during this distribution revision. Scientific tables, figure exports, aggregate results, protocols, model locks and thresholds were not changed; only machine-specific local tool-path references in packaging/figure helper code were made portable. Sharing statements and access notes retain the prior public-distribution revision. The 21 October 2026 revision reframed the main-manuscript text for Journal of Intensive Care without changing any reported number, table, figure, model or threshold; the supplement changed only in its title. Reporting locations, the reference list and current file hashes were then regenerated. The original models and parameter files remain unchanged locally.
CONTENTS_SHA256.json inventories the current public archive, excluding the manifest itself.


Current-version provenance
The analytical-reference clarification appears on main page 5, printed lines 108–110. reporting/finalisation_checks.json and reporting/reporting_pagination_validation_final.json describe this current package. The authors' existing ethics and authorship declarations were retained; this document-editing pass did not independently verify institutional ethics arrangements or historical data-transfer conditions.

Post hoc respiratory-support extension
fio2_extension/ contains the pre-fitting lock packages (v1.0, v1.1), the FiO2 coverage audit, the implementation snapshot, the hash-chained execution log and aggregate results of the extension reported in the main text and Additional file 1 (S6, Table S8). Fitted extension models and complete coefficients are excluded, as for the original models. See fio2_extension/README_extension.txt.
