Abstract
Background and objective Heart failure (HF) is a leading cause of hospitalizations and early readmissions, contributing significantly to morbidity, mortality, and healthcare costs. Identifying factors that predict 30-day readmission can help design targeted interventions to improve patient outcomes. Therefore, this study aimed to identify these predictors.. Methods A retrospective cohort study involving 300 patients with HF was conducted. Demographic, clinical, laboratory, comorbidity, and socioeconomic factors were analyzed. The primary objective was to identify clinical, laboratory, and comorbidity-related factors independently associated with 30-day hospital readmission in patients with HF. A secondary objective was to evaluate and interpret the performance of logistic regression and machine learning models (random forest and XGBoost) in predicting readmission risk. Results The cohort had a mean age of 68.4 ±10.2 years, with 186 (62%) males and 114 (38%) females. Readmission was observed in 93 (31%) patients. Readmitted patients more frequently had reduced left ventricular ejection fraction (LVEF <40%; 122, 41%), elevated B-type natriuretic peptide (BNP, 168; 56%), creatinine >1.5 mg/dL (87, 29%), and a Charlson Comorbidity Index (CCI) score ≥4 (196, 65.3%). Multivariate regression confirmed reduced LVEF (adjusted odds ratio (OR): 1.74, 95% confidence interval (CI): 1.09-2.96, p = 0.021), elevated creatinine (adjusted OR: 1.89, 95% CI: 1.11-3.11, p = 0.015), and higher CCI score (adjusted OR: 2.31, 95% CI: 1.41-3.77, p = 0.001) as independent predictors. Random forest achieved the best performance (accuracy 0.72, precision 0.61, recall 0.58, F1-score 0.59, area under the receiver operating characteristic (ROC-AUC) curve 0.44) but still showed poor discrimination (ROC-AUCs for logistic regression and XGBoost were 0.43 and 0.42, respectively). Conclusions Comorbidity burden, impaired renal function, and reduced cardiac function are key predictors of 30-day readmission in HF patients. Machine learning models provided useful interpretability but showed poor discrimination, highlighting their role as exploratory tools for hypothesis generation rather than significant improvements in predictive performance.