Abstract
INTRODUCTION: Hospital readmissions within 30 days remain a critical issue in healthcare, signaling potential care discontinuities and contributing to escalating costs. Leveraging machine learning (ML) on electronic health records (EHRs) presents a promising strategy to identify patients at heightened risk for readmission and support early intervention. AIM: This study evaluates the performance of four ML models - logistic regression, random forest, XGBoost, and deep neural networks (DNN) - in predicting 30-day hospital readmissions. It also identifies the most influential predictors using SHapley Additive exPlanations (SHAP) values. MATERIALS AND METHODS: We conducted a retrospective analysis on 101,766 de-identified inpatient encounters from the University of California, Irvine (UCI) Diabetes 130-United States (US) Hospitals dataset. After preprocessing, including feature imputation, scaling, and one-hot encoding, we trained and validated models on an 80/20 train-test split. Evaluation metrics included accuracy, precision, recall, F1-score, and area under the receiver operating characteristic curve (AUC-ROC). RESULTS: XGBoost achieved the highest AUC-ROC (0.667), followed closely by logistic regression (0.642) and random forest (0.630). Despite DNNs demonstrating the highest recall for the positive class (0.143), their AUC-ROC (0.579) and precision (0.186) indicated lower reliability. SHAP analysis revealed that previous admissions, number of medications, and comorbidity indicators such as diabetes medication usage and admission type were key predictors influencing model decisions. CONCLUSION: XGBoost outperformed other models in predicting 30-day readmissions using EHR data, balancing performance and interpretability when coupled with SHAP values. These findings underscore the promise of ensemble models in improving discharge planning and reducing preventable readmissions. Future research should explore the inclusion of behavioral and social health features to further enhance predictive accuracy.