A Robust and Explainable Framework for Predicting Hospital Readmission in Patients with Diabetes

Abstract

Background: Hospital readmission among patients with diabetes remains a healthcare challenge, contributing to adverse clinical outcomes, increased healthcare expenditures, and reduced quality of care. This study presents a leakage-resistant and clinically interpretable machine learning framework for predicting hospital readmission risk using structured clinical and admission-related data. Methods: The framework incorporated preprocessing strategies to minimize information leakage, including duplicate-feature removal, categorical encoding, and feature normalization. Class imbalance was addressed using fold-specific strategies, including class-weighted learning and the Synthetic Minority Over-sampling Technique for Nominal and Continuous features (SMOTENC). Resampling was performed exclusively within each training fold, while validation and test data retained their original class distributions. Feature selection was conducted using correlation analysis and Recursive Feature Elimination (RFE). Principal Component Analysis (PCA) was used for exploratory variance analysis and visualization, preserving the interpretability of the predictive features Five models were evaluated: L2-regularized logistic regression, random forest, ANN, XGBoost, and Hybrid CNN–LSTM. SHAP and LIME provided global and local explanations. Results: The hybrid CNN-LSTM achieved the strongest performance, attaining an accuracy of 0.80 and an area under the receiver operating characteristic curve (ROC-AUC) of 0.86 on a hold-out dataset. Explainability analyses enhanced model transparency and supported clinical interpretation of predictions. Conclusions: Combining leakage-resistant preprocessing, balanced training, interpretable feature selection, and explainable deep learning provides a robust and clinically meaningful framework for diabetic readmission prediction.