Predictive modeling for hospital readmission risk among cardiovascular patients: a data-driven approach

Abstract

Hospital readmission is an important clinical issue in cardiovascular disease, contributing significantly to morbidity, mortality, and health care costs. This problem is compounded in low-resource settings by the lack of comprehensive, localizeddatasets. To address this gap, we performed a six-month manual data collectionon 3,867 cardiovascular patients in Bangladesh, capturing a wide range of demographic, clinical, and behavioral variables, including age, type of residence, ejectionfraction, and multicomorbidity profiles. Notably, 64% of patients were readmittedduring the follow-up period, highlighting the urgent need for predictive tools tosupport early intervention. Using this dataset, we developed a machine learningframework to predict the risk of readmission at six months. The models includedgradient boosting (XGBoost), TabNet (a deep neural network for tabular data),and stacked ensemble classifiers. The pipeline incorporated robust preprocessingtechniques, such as MICE-inspired KNN imputation for missing values, SMOTE forclass imbalance, and feature engineering through principal component analysis andinteraction feature generation. To improve transparency and clinical interpretability, we used SHAP and LIME explainability tools, which revealed key predictorsof readmission, such as medication adherence, comorbidity burden, and follow-upstatus. The final Super XGBoost Ensemble achieved 92% accuracy, an AUC of 0.96,and an F1-score of 0.94, outperforming baseline models and demonstrating the advantage of integrating demographic context with clinical depth. These results notonly contribute to the field of cardiovascular informatics, but also offer a replicable,interpretable framework for policymakers and healthcare providers aiming to reducereadmissions in resource-limited environments.