Predictive Modeling for Hospital Readmission Risk Using XGBoost and LightGBM Techniques with Class Imbalance

This paper reports an end-to-end machine-learning framework to predict the likelihood of patient readmission. Scaling the process and proper data management required scalability for SQLAlchemy and PostgreSQL. Additionally, an ETL pipeline was created to integrate and transform structured data from EHRs. This, then, gathered all the essential knowledge related to a patient, while model development included training followed by the comparison of many classifiers: specifically, Random Forest, XGBoost, and LightGBM. The LightGBM algorithm outperformed the others, with an AUC-ROC of 0.88 and an accuracy of 83%, both of which were rather good. As a result, SHAP (SHapley Additive exPlanations) values made the model interpretable, making the predictions accessible to clinicians as well. To further promote data-driven clinical decision-making, a Streamlit web application was made available that allowed real-time interaction with the model itself.