Operationalizing Machine Learning in Hospital Diabetes Care: A Multi-Center Study of Readmission Prediction and Clinical Utility

Abstract

The increasing burden of diabetes-related hospitalizations and high rates of 30-day readmissions pose significant challenges to healthcare systems worldwide. This study aims to develop and operationalize machine learning (ML) models for predicting early readmissions among patients with diabetes using a large, multi-center clinical dataset. A retrospective observational design was employed using the Diabetes 130 U.S. Hospitals dataset, comprising 101,766 inpatient encounters. Following data preprocessing and feature engineering, multiple ML algorithms, including logistic regression, random forest, and gradient boosting, were developed and evaluated. Class imbalance was addressed using the Synthetic Minority Oversampling Technique (SMOTE), and model performance was assessed using standard metrics such as AUC-ROC, precision, recall, and F1-score. The results demonstrate that ensemble models, particularly gradient boosting, outperform traditional approaches in predicting 30-day readmissions. Feature importance analysis identified prior healthcare utilization, medication patterns, and comorbidity burden as key predictors. The models enabled effective risk stratification, supporting clinical decision-making and targeted interventions. This study highlights the importance of integrating predictive accuracy with interpretability and generalizability to enhance real-world applicability. The findings provide a practical framework for deploying ML-based decision support systems in hospital settings to improve patient outcomes and reduce avoidable readmissions.