Calibration and Implementation Sensitivity of OASIS Hospital Mortality Predictions in Two Critical Care Databases Code and Aggregate Results
Authors/Creators
- 1. Affiliated Baotou Clinical Medical College of Inner Mongolia Medical University, Baotou, China
- 2. Department of Anesthesiology, Baotou Central Hospital, Baotou, China
Description
Version 2.0.0 contains PostgreSQL, R and Python code, the original SAP and dated post-unblinding deviations, anonymous hospital/group aggregate results, and figures for retrospective OASIS calibration evaluation and updating in eICU-CRD and MIMIC-IV. The primary eICU cohort contains 144,832 stays from 132 hospitals; MIMIC-IV contains 85,242 stays.
This revision adds source-provenance and day-1 support audits, three additional mapping analyses with 1,000 global-patient full-source-refit draws each, and complete MIMIC original-score and fixed-update results. All six scoring implementations are reported. The three added mappings yielded observed hospital-pass gains of 19.7, 20.5 and 15.2 percentage points. The restrictive worst-ABG-intubation/timed-airway proxy yielded only 4.5 points; unknown elective-status coding reversed baseline calibration error. Hospital-pass percentile ranges describe resampling stability, not validated confidence intervals or method optimality.
Read-only eICU raw-input SELECT replay matched 156,466 frozen rows. MIMIC cohort reconstruction and component-score recalculation from preloaded derived inputs matched 85,242 rows. The MIMIC upstream raw-event derivation and historical mimic-code commit remain unavailable. Separate R implementations checked 792 hospital-by-method evaluation fits; these computational checks are not independent human review.
Local updating uses 20 repeated patient cross-fitting runs with approximately 80% local training data. This is retrospective prediction-performance benchmarking, not prospective clinical deployment or hospital-quality ranking. MIMIC anchor-year strata are not admission calendar eras, and U3 was not evaluated in MIMIC. Predictor proxies, extensive unknown elective status, fallback measurements and older-group calibration worsening remain limitations.
Only code, protocols, reporting self-assessments, non-patient-level aggregates and figures are included. No patient records, patient identifiers, hospital crosswalks, credentials, certificates or signed agreements are included. Source data remain subject to PhysioNet credentialing and data use agreements.
Mixed file-level licenses: MIT for code; CC BY 4.0 for newly authored documentation and aggregate results. Underlying database and third-party terms remain unchanged. Earlier dated analysis history is retained inside this version package.