Predictive Analysis of Readmission Risk of Heart Failure Patients within 28 days

Heart failure (HF) is one of the emerging global health issues and a leading cause of readmission as well as a source of healthcare system strain due to escalating expenses. Proper identification of patients with high risks of readmission early, especially within 28 days, presents an opportunity to intervene before the patient is readmitted. In our paper, we suggest a machine learning-based model to forecast the risk of early readmission with the help of structured electronic health record (EHR) data. We put different preprocessing methods, SHAP-based feature pruning, and cost-sensitive learning into our pipeline, as well as tuning thresholds. We tested and trained several models like Support Vector Machines (SVM), XGBoost, LightGBM, and a Stacked Ensemble model. The stacked model, with threshold tuning and SHAP interpretability, appeared superior in terms of the cases of high-risk detection. Our findings indicate that the use of interpretability and recall-optimized thresholding would provide a fair and clinically feasible readmission prediction approach.