Enhancing Predictive Accuracy in Healthcare Readmission through Ensemble Learning with Feature Selection and Imbalanced Data Handling

Abstract

Healthcare readmission rates represent a significant burden on healthcare systems globally, contributing to increased costs and potentially indicating suboptimal patient care. This research proposes an enhanced predictive model for healthcare readmission using ensemble learning techniques, specifically focusing on Gradient Boosting Machines (GBM) and Random Forests, augmented with a rigorous feature selection process and strategies to mitigate the challenges posed by imbalanced datasets. We employ a hybrid feature selection approach combining filter and wrapper methods to identify the most relevant predictors. Furthermore, we address the class imbalance problem inherent in readmission data using Synthetic Minority Oversampling Technique (SMOTE) and cost-sensitive learning. The performance of the proposed model is evaluated using various metrics, including AUC-ROC, precision, recall, F1-score, and Brier score. The results demonstrate a significant improvement in predictive accuracy compared to baseline models and existing approaches, offering a promising avenue for proactive intervention and improved patient outcomes. The interpretability of the model is further enhanced through SHAP (SHapley Additive exPlanations) values, providing insights into the factors driving readmission predictions.