Data Lineage and Temporal Validation in Hospital Readmission Prediction: A Reproducibility Study of Integrated Health Information Systems

Abstract

Hospital readmission models are vulnerable to data leakage when predictor timestamps, data refresh schedules, and clinical events are not aligned with the intended prediction time. We audited an integrated health-information pipeline containing 64,732 adult admissions and reproduced a 30-day readmission model under three feature-extraction strategies. A conventional retrospective extract produced an AUROC of 0.79, but performance fell to 0.72 when all variables unavailable at discharge were removed and to 0.70 under prospective temporal validation. Several high-ranking predictors were generated or finalised after the nominal prediction time, including coding and reconciliation variables. Reconstructing data lineage allowed 11 leakage-prone features to be identified and removed. Although discrimination decreased, calibration improved in the later validation period after model updating. The study demonstrates that apparent readmission performance can be materially inflated by hidden timing assumptions in integrated hospital systems. Reproducible clinical prediction requires explicit lineage, feature-availability timestamps, and temporal validation that mirrors the intended deployment workflow.