Machine Learning-Based Risk Stratification for Predicting Hospital Readmissions in Heart Failure Patients: A Multi-Center Retrospective Study

Abstract

Background: Heart failure remains a leading cause of hospitalization and rehospitalization worldwide. Traditional risk scores, though useful, often fail to capture the complex interplay of clinical, laboratory, and imaging data that influence patient outcomes. Machine learning (ML) techniques have emerged as promising tools to improve risk prediction and guide individualized care strategies.

Methods: We conducted a retrospective analysis of 3,200 heart failure patients admitted across three tertiary hospitals in the Middle East between 2019 and 2024. Clinical variables (demographics, comorbidities, medications), laboratory parameters, echocardiographic findings, and electronic health record (EHR) data were included. Several supervised ML models (random forest, XGBoost, logistic regression) were trained and validated to predict 30-day readmission. Model performance was compared using area under the receiver operating characteristic curve (AUC), sensitivity, specificity, and calibration metrics.

Results: The XGBoost model outperformed traditional logistic regression and random forest (AUC 0.84 vs. 0.71 and 0.78, respectively). Important predictors identified included serum sodium, NT-proBNP, prior hospitalizations, and medication adherence patterns extracted from EHR. The model demonstrated good calibration across risk strata, with a sensitivity of 82% and specificity of 76% in the validation cohort.

Conclusion: ML-based models provide more accurate prediction of 30-day readmissions compared with conventional risk scores in heart failure. By integrating routinely available clinical and EHR data, these tools may enable early identification of high-risk patients, inform discharge planning, and ultimately reduce healthcare burden. Future work will focus on prospective validation and embedding ML models into clinical decision support systems.