From patterns to prognosis: machine learning–derived clusters in advanced heart failure

Introduction Advanced heart failure (HF) is a clinically heterogeneous condition with poor prognosis, and traditional classification systems often fail to capture the complexity needed for personalized care. This study aimed to identify clinically meaningful phenotypic subgroups among patients with advanced HF using unsupervised machine learning and to evaluate their association with long-term outcomes. Methods A retrospective analysis was conducted on 524 patients with advanced HF who underwent comprehensive clinical, echocardiographic, hemodynamic, and cardiopulmonary exercise assessments. Using k-means clustering on standardized, multidimensional data, two distinct phenotypes were identified. The primary composite outcome was defined as all-cause mortality, left ventricular assist device implantation, or heart transplantation. Associations between cluster assignment and outcomes were evaluated using Kaplan–Meier analysis and Cox proportional hazards regression. Results The first cluster, representing patients with relatively preserved hemodynamics and functional status, was associated with a more favorable prognosis, while the second cluster included older individuals with significant biventricular dysfunction, higher pulmonary pressures, and poorer exercise capacity. These patients experienced a markedly higher rate of the composite outcome over a median follow-up of 2.4 years, with Cluster 2 showing a significantly increased risk (hazard ratio HR: 3.84; 95% CI: 2.72–5.43; p 0.001). Conclusion Machine learning–based clustering revealed two distinct phenotypes in advanced HF with differing clinical features and prognoses. This approach may enhance risk stratification and inform individualized therapeutic strategies in this high-risk population.