ML-ENHANCE: A Machine Learning Framework for Hospital Readmission Prediction with Cost-Optimized Decision Thresholds
Hospital readmissions within 30 days affect 15–20\% of patients, generating over $40 billion in potentially preventable costs annually. We developed two prediction models using 113,312 internal medicine admissions (2017–2023): ENHANCE, an interpretable clinical score, and ML-ENHANCE, a machine learning ensemble. Using decision-theoretic threshold optimization across cost ratios from 0.5:1 to 20:1, we demonstrate that ML-ENHANCE (AUC 0.752) substantially outperforms ENHANCE (AUC 0.696) and HOSPITAL (AUC 0.676). At the 5:1 cost ratio, ML-ENHANCE matches ENHANCE sensitivity (76.9\%) while achieving 8.3 percentage points higher specificity, translating to 7,400 fewer unnecessary interventions annually and projected savings of $69.3 million versus $14.7 million for ENHANCE. ML-ENHANCE achieves 4–5 times greater economic impact, supporting deployment in automated population health applications.