Machine learning–based identification of mortality risk factors in hospitalized heart failure patients: a multicenter real-world study

Background Despite advances in guideline-directed medical therapy, heart failure (HF) continues to carry a high burden of mortality and hospitalization. Existing risk models are largely derived from historical cohorts or highly selected clinical trial populations, which may not reflect the complexity and therapeutic standards of contemporary "real-world" clinical practice. There remains a critical evidence gap in mortality risk assessment for the modern HF era, requiring the identification of contemporary determinants that align with current clinical practice. Purpose This study aims to evaluate the predictors of mortality in hospitalized HF patients in real-world settings using a machine learning approach. Methods This study is part of a prospective, observational, national HEROES registry. Patients with HF were enrolled in 41 Polish centers, and the only exclusion criterion was age < 18 years. All data (clinical, laboratory, etc.) were reported by site investigators via an electronic case report form. Follow-up data on all-cause mortality were obtained from the Polish Ministry of Digital Affairs. Five machine learning architectures—Elastic Net, Ridge, Lasso, Random Survival Forest, and XGBoost—were evaluated using a 70/30 training-to-testing split. The optimal model was selected based on the highest concordance index (C-index) on the independent testing dataset and was subsequently selected for downstream risk stratification and hazard ratio analysis. Results The study cohort comprised 1143 hospitalized patients with HF, followed for a median duration of 435 days (IQR: 324–544), during which 211 (18.5%) death events occurred. The Elastic Net model was selected for final analysis as it yielded the highest performance, with a C-index of 0.73 on the independent testing dataset (Figure 1A). Time-dependent AUC analysis revealed consistent discriminative accuracy across the entire follow-up duration (Figure 1B). The model identified a prognostic signature consisting of several key variables (Figure 1C). Among the factors included in the model, those most strongly associated with a higher risk of mortality were anemia, hepatic dysfunction, prior valvular intervention, chronic HF diagnosis, age, peripheral artery disease, and atrial fibrillation, whereas a reduced risk was associated with the increased physical limitation score, symptom frequency score, and KCCQ-12. Based on the model-derived risk score, the total population was stratified into three distinct risk categories (log-rank p < 0.001) (Figure 1D). The high-risk group exhibited a hazard ratio (HR) for death of 6.93 (95% CI: 4.20–11.42, p < 0.001) compared to the low-risk reference group, whereas the median-risk group had an HR for death of 2.65 (95% CI: 1.54–4.58, p < 0.001). Conclusions Precision risk profiling enables the early identification of vulnerable heart failure patients who may benefit from intensified monitoring and targeted therapeutic interventions.