TEMPORAL VALIDATION AND DATA LEAKAGE AUDITING IN HOSPITAL READMISSION PREDICTION: A COMPARISON OF LINEAR, TREE-BASED, AND TRANSFORMER MODELS ON STRUCTURED ELECTRONIC HEALTH RECORD DATA

Hospital readmission prediction models often fail deployment due to temporal validation errors and feature leakage.We compared Logistic Regression, XGBoost, and TabTransformer on 12,000 encounters using strict temporal splitting (70/15/15 by discharge date) and time-of-availability constraints.With 25 discharge-available features, XGBoost achieved AUROC = 0.63, AUPRC = 0.36, and exceptional calibration (ECE = 0.015), outperforming Logistic Regression (AUROC = 0.61, ECE = 0.220) and TabTransformer (AUROC = 0.55).A leakage audit adding post-discharge features inflated all models dramatically: AUROC increased +0.31 to +0.36 (exceeding 0.91), demonstrating that temporally inadmissible features create non-deployable optimism.Key predictors included patient age, comorbidity burden, renal dysfunction, and admission acuity.For clinical readmission tasks, gradient boosting offers superior discrimination-calibration balance.Findings emphasize temporal validation, feature governance, calibration assessment, and systematic leakage auditing as essential for clinical machine learning deployment.