Machine Learning Approaches for Predicting 30-Day Hospital Readmissions: Evidence from Massachusetts Healthcare Data

Thirty-day hospital readmissions represent a critical challenge in healthcare, contributing to significant financial burdens, increased patient morbidity, and reflecting gaps in care continuity. This study aimed to develop and evaluate machine learning models for predicting 30-day hospital readmissions using a comprehensive, statewide healthcare dataset from Massachusetts. Employing a quantitative, predictive modeling design, this research compared the performance of Ridge regression with two advanced ensemble methods: Random Forest and Gradient Boosting. The models were trained and tested on a hospital-year panel dataset derived from the Massachusetts readmissions data book. Performance was evaluated using Root Mean Squared Error (RMSE) and the coefficient of determination (R²). The results demonstrated the superior predictive power of the ensemble methods over the traditional linear model. Gradient Boosting emerged as the top-performing model, achieving the lowest RMSE of 1.48 and the highest R² of 0.81, followed closely by Random Forest (RMSE = 1.52, R² = 0.80). In contrast, Ridge regression showed limited predictive capability (RMSE = 2.54, R² = 0.43). Feature importance analysis from the Gradient Boosting model identified the number of deaths/readmissions and the number of cases as the most influential predictors, with hospital quality ratings and geographic factors also contributing significantly. The findings indicate that machine learning, particularly Gradient Boosting, provides a robust and accurate tool for identifying patients at high risk of readmission. Implementing such models can enable healthcare systems to better allocate resources, tailor discharge planning, and ultimately improve patient outcomes by reducing costly and disruptive readmissions.