Abstract:

Objective To construct a machine learning-based risk prediction model that integrates multidimensional risk factors to predict rehospitalization in patients with acute decompensated heart failure (ADHF), and to provide clinical decision support for targeted prevention. Methods In this retrospective cohort study, 224 ADHF patients treated at Yellow River Sanmenxia Hospital between September 2022 and March 2025 were enrolled. Patients were classified into a rehospitalization group (n=72) and a non-rehospitalization control group (n=152) based on whether they were rehospitalized within 6 months after discharge. Clinical variables were collected and compared between groups. Feature selection was performed using least absolute shrinkage and selection operator (LASSO) regression combined with ten-fold cross-validation, and multivariate logistic regression was used to identify independent predictors. Prediction models were developed using three machine learning algorithms-support vector machine (SVM), random forest (RF), and gradient boosting decision tree (XGBoost)-and model performance was assessed with receiver operating characteristic (ROC) analysis. Results Among 224 ADHF patients, 72 experienced rehospitalization within 6 months, yielding a rehospitalization rate of 32.14%(72/224). No significant between-group differences were observed for age, sex, body mass index, education level, hypertension, diabetes, smoking history, or alcohol history (P>0.05). Significant differences were found for living situation, disease duration, New York Heart Association (NYHA) functional class at discharge, regular follow-up, adherence to a low-sodium diet, fluid-restriction adherence, medication adherence, self-care ability, anxiety, depression, appropriate exercise, and hyperlipidemia (P<0.05). LASSO, ten-fold cross-validation, and logistic regression consistently identified eight optimal predictive features: NYHA class at discharge, regular follow-up, adherence to a low-sodium diet, medication adherence, anxiety, depression, appropriate exercise, and hyperlipidemia. Using these eight predictors, SVM, RF, and XGBoost models were trained. ROC analysis showed model sensitivities of 88.33% (SVM), 69.44% (RF), and 88.89% (XGBoost); specificities of 65.79% (SVM), 78.95% (RF), and 82.24% (XGBoost); accuracies of 71.43% (SVM), 75.89% (RF), and 84.38% (XGBoost); and AUCs of 0.823 (SVM), 0.869 (RF), and 0.916 (XGBoost). Conclusion Rehospitalization within 6 months after discharge is common among ADHF patients. Rehospitalization risk is associated with discharge NYHA class, regular follow-up, adherence to a low-sodium diet, medication adherence, anxiety, depression, appropriate exercise, and hyperlipidemia. Machine learning models that integrate these multidimensional factors (SVM, RF, and XGBoost) show good predictive performance, with XGBoost achieving the best discrimination in this cohort.

Key words: heart failure, re-hospitalization, predictive model, machine learning, multidimensional risk factors