Hospital Readmission Predictive Modeling

Abstract

Unplanned hospital readmissions within 30 days of discharge affect 12-20% of hospitalised patients across European health systems, costing an estimated EUR 15-25 billion annually and serving as a widely used marker of healthcare quality. Despite two decades of modelling efforts, the discriminative performance of readmission prediction has plateaued at C-statistics of 0.60-0.70 -- barely better than chance -- raising questions about whether readmissions are fundamentally unpredictable or whether existing models simply use the wrong data. We analysed 124,682 hospitalisations across three Italian academic hospitals (Rome, Milan, and Naples) from January 2016 to December 2021, comparing five predictive modelling approaches: administrative claims models, clinical regression models using structured EHR data, NLP-enhanced models incorporating discharge summary text, temporal deep learning models using time-series EHR data, and multi-modal ensemble models integrating all data sources. We evaluated prediction for all-cause 30-day readmission across four high-volume conditions: heart failure (HF), pneumonia, COPD exacerbation, and hip/knee arthroplasty. Our Readmission Prediction Utility Score (RPUS) integrates discrimination, calibration, net reclassification improvement, actionability of identified risk factors, and implementation feasibility. The multi-modal ensemble achieved the highest RPUS (0.916), with C-statistic 0.78 for all-cause readmission -- a meaningful advance over administrative models (C-stat 0.62). NLP-enhanced models (RPUS 0.892) contributed the largest single-source improvement by capturing social determinants, functional status, and care plan adherence barriers from unstructured clinical notes that coded data systematically misses.