OBJECTIVES: Patients with heart failure (HF) are heterogeneous in terms of etiology, clinical course, and responsiveness to intervention.  This study uses cluster analysis to identify HF phenotypes and to explore whether clustering minimizes heterogeneity. METHODS: A retrospective study was performed using Optum Labs Data Warehouse, a large warehouse including administrative claims data on over 100 million privately insured and Medicare Advantage enrollees.  We identified patients with prevalent HF (ICD-9 codes 428.xx, 402.x1, 404.x1 or 404.x3) using established claims-based algorithms.  The cohort index date was either the discharge date of a single inpatient admission with HF or the second of two outpatient visits for HF within 24 months.  We then selected those at highest risk for future hospitalization using four models predicting all-cause and HF-specific hospitalization at 90 and 360 days. Patient’s similarity to other patients was measured by Euclidean distance across demographic features and comorbidities.  Patients were clustered into similar groups using hierarchical clustering and Ward’s method.  The optimal number of clusters was selected by analyzing inter-cluster distances and expert opinion.  Inter- and intra-cluster variation was measured in the clustering features as well as in utilization measures not used in the clustering. RESULTS: Twelve patient clusters were identified in which patients varied considerably in terms of their age, sex, race and comorbidity burden.  Diabetes was a key differentiating feature in 5 of the clusters.  Clusters ranged in size from 780 to 4,214patients.  Future annual inpatient admission rates varied from 124.65 per 100 patients in the healthiest group to 254.49 per 100 patients in a cluster dominated by patients with renal disease and diabetes. CONCLUSIONS: Even within a cohort of high-risk patients with HF, clustering can stratify patients into heterogeneous groups. Future work will attempt to distill differentiable treatment response in these sets.