Reproduction code and cross-site aggregate data for "Separate selection from correction: a paired chart–monitor audit of rare low-SpO₂ events and held-out-hospital prediction"

This record contains the reproduction code and cross-site aggregate results supporting the manuscript “Separate selection from correction: a paired chart–monitor audit of rare low-SpO₂ events and held-out-hospital prediction.” The archive includes 12 aggregate CSV/JSON files, Python/SQL source snapshots, a SHA-256 manifest, and a standalone reproduction script. From the extracted archive, run python reproduce.py with Python 3.10 or newer. The script verifies file integrity and regenerates numerical evidence for Tables 1–4, Supplementary Tables S1 and S4–S8, and the numerical inputs to Figures 3–4. The archive excludes the manuscript, patient-level records, hospital identifiers and individual hospital result rows, prediction vectors, and trained model weights. The minimum individual hospital size in Supplementary Table S7 is withheld. Hospital-equal losses are calculated from cross-site sums; the underlying site-derived sums and medians cannot be independently reconstructed from this archive. Pooled performance metrics and patient-bootstrap intervals are checked against retained source results. Raw-level recomputation requires authorized access to the eICU Collaborative Research Database v2.0 and the protected prediction pipeline.