Machine Learning-Driven Risk Clusters for Adverse Outcomes in Patients Hospitalized with Acute Heart Failure: Using a Retrospective Cohort Data

Aim Heart failure (HF) poses a growing public health burden, yet conventional risk stratification models fail to capture the multidimensional complexity of acute HF by overlooking nutritional and lifestyle-related factors critical to prognosis. This study aimed to classify patients hospitalized with acute HF into clinically distinct risk subgroups using machine learning and to evaluate their prognostic significance for individualized nursing care. Methods and Results We retrospectively analyzed records of 1,104 patients admitted to the cardiac intensive care unit of a tertiary hospital in Seoul, Korea (2013–2023). Adverse outcome was defined as all-cause mortality or readmission. Random forest identified key predictors, and clustering using Gower distance and Partitioning Around Medoids defined patient groups. Survival was assessed using Kaplan–Meier and Cox proportional-hazards models. Patients had a mean age of 66.4±14.6 years, and 31.3% experienced adverse outcomes (25.6% died, 8.4% were readmitted). Key predictors included nutritional risk index, hemoglobin, creatinine, left ventricular ejection fraction, and age. Three clinically distinct clusters were identified: Cluster 1, Middle-aged unhealthy lifestyle; Cluster 2, Older multimorbidity; and Cluster 3, Malnutrition-renal dysfunction (log-rank p < .001). Compared with Cluster 1, Cluster 3 showed a significantly higher risk of adverse outcomes (HR=1.82, 95% CI 1.33–2.50, p < .001). Conclusion Machine learning–driven clustering identified three HF phenotypes with divergent prognoses, with the malnutrition–renal dysfunction phenotype conferring nearly twofold higher mortality risk. These findings support cluster-specific nursing strategies, including early nutritional risk screening and renal monitoring, to guide individualized care in acute HF.