An Explainable Clinical Decision Support System for Hospital Readmission Prediction in Dermatology Patients

Abstract

Hospital readmissions continue to be a major concern for healthcare providers across the globe, affecting patient outcomes while placing significant financial and operational pressure on healthcare institutions. In the United States, approximately 20% of Medicare beneficiaries are readmitted to hospitals within 30 days of discharge, contributing to healthcare expenditures of more than $26 billion each year. Despite extensive research on readmission prediction, patients undergoing dermatological and skin-related surgical procedures remain a relatively unexplored group. These patients often face distinct challenges, including post-operative wound care complications, infection risks, and the influence of underlying comorbid conditions.
This research introduces a comprehensive machine learning framework designed to estimate the likelihood of 30-day hospital readmission among patients who have undergone skin-related procedures. The proposed framework encompasses the complete predictive analytics workflow, including data preprocessing, feature extraction, model development, performance evaluation, and interpretability analysis. Four supervised learning algorithms—Logistic Regression, Random Forest, Extreme Gradient Boosting (XGBoost), and Support Vector Machine (SVM)—were trained and evaluated using a de-identified electronic health record (EHR) dataset containing both clinical and procedure-specific information.
Experimental findings indicate that the XGBoost model delivers superior predictive performance compared with the other classifiers. The model achieved an AUC-ROC score of 0.921, an accuracy of 91.2%, a precision of 91.5%, and a recall of 88.2%. To improve transparency and support clinical decision-making, SHAP (SHapley Additive exPlanations) was employed to identify the most influential factors contributing to readmission risk. The analysis revealed that previous hospitalization history, the number of existing comorbidities, hospital length of stay, patient age, and the severity of the skin procedure were the most significant predictors.
To facilitate practical clinical adoption, the predictive system has been integrated into an interactive web-based dashboard that allows healthcare professionals to assess patient readmission risk in real time. The platform also supports automated generation of clinical reports, enabling more informed discharge planning and targeted interventions. The proposed framework demonstrates the potential of machine learning and explainable artificial intelligence to enhance readmission prediction, improve patient management strategies, and reduce avoidable healthcare costs.