Predicting Hospital Readmission Risk Using Explainable Machine Learning Models: A Multi-Hospital Study

Abstract

Unplanned 30-day hospital readmissions represent a critical challenge in healthcare delivery, contributing to over $26 billion in annual expenditures in the United States, with more than $17 billion attributed to preventable events. While conventional risk stratification tools such as the LACE Index and HOSPITAL Score offer ease of calculation, they demonstrate limited discriminative performance (AUC-ROC: 0.60-0.68) due to their linear assumptions and omission of social determinants of health (SDOH). This comprehensive review synthesizes evidence from multi-hospital studies employing machine learning (ML) and explainable artificial intelligence (XAI) frameworks to enhance readmission prediction. We examine diverse patient cohorts across multiple institutions, including elderly multimorbid populations, renal transplant recipients, and traumatic brain injury patients, analyzing data standardization approaches utilizing HL7 FHIR and OMOP Common Data Model architectures. Gradient-boosted ensemble methods, particularly XGBoost and LightGBM, consistently outperform traditional scores, achieving AUC-ROC values up to 0.867 in specialized populations when augmented with SDOH variables. The integration of SHAP and LIME explanation frameworks enables both global model auditing and local clinical decision support, addressing the interpretability dilemma that has historically hindered ML adoption. Federated learning strategies, particularly personalized approaches like FedPer, demonstrate promise in maintaining performance across heterogeneous hospital settings while preserving data privacy (AUC-ROC: 0.780-0.795). Implementation through SMART on FHIR frameworks facilitates seamless EHR integration, enabling real-time risk stratification with actionable feature attributions.