Predicting 30-Day Hospital Readmission in Diabetes Patients

Abstract

This study focuses on predicting 30-day hospital readmission risk among diabetes patients using machine learning techniques. A leakage-free machine learning pipeline was developed using the UCI Diabetes dataset. Multiple classification models including Logistic Regression, Random Forest, XGBoost, and CatBoost were evaluated. CatBoost achieved the best performance with an F1-score of 0.276 and ROC-AUC of 0.670. SHAP-based model interpretation identified prior inpatient visits and discharge destination as major factors influencing readmission risk. The study demonstrates the potential of explainable machine learning approaches for identifying high-risk diabetic patients and supporting healthcare decision-making.