Unsupervised machine learning to identify subphenotypes among cardiac intensive care unit patients with heart failure

Aims Hospitalized patients with heart failure (HF) are a heterogeneous population, with multiple phenotypes proposed. Prior studies have not examined the biological phenotypes of critically ill patients with HF admitted to the contemporary cardiac intensive care unit (CICU). We aimed to leverage unsupervised machine learning to identify previously unknown HF phenotypes in a large and diverse cohort of patients with HF admitted to the CICU. Methods We screened 6008 Mayo Clinic CICU patients with an admission diagnosis of HF from 2007 to 2018 and included those without missing values for common laboratory tests. Consensus k-means clustering was performed based on 10 common admission laboratory values (potassium, chloride, anion gap, blood urea nitrogen, haemoglobin, red blood cell distribution width, mean corpuscular volume, platelet count, white blood cell count and neutrophil-to-lymphocyte ratio). In-hospital mortality was evaluated using logistic regression, and 1 year mortality was evaluated using Cox proportional hazard models after multivariable adjustment. Results Among 4877 CICU patients with HF who had complete admission laboratory data (mean age 69.4 years, 38.4% females), we identified five clusters with divergent demographics, comorbidities, laboratory values, admission diagnoses and use of critical care therapies. We labelled these clusters based on the characteristic laboratory profile of each group: uncomplicated (25.7%), iron-deficient (14.5%), cardiorenal (18.4%), inflamed (22.3%) and hypoperfused (19.2%). In-hospital mortality occurred in 10.7% and differed between the phenotypes: uncomplicated, 2.7% (reference); iron-deficient, 8.1% [adjusted odds ratio (OR) 2.18 (1.38-3.48), P < 0.001]; cardiorenal, 10.3% [adjusted OR 2.11 (1.37-3.32), P < 0.001]; inflamed, 12.5% [adjusted OR 1.79 (1.18-2.76), P = 0.007]; and hypoperfused, 21.9% [adjusted OR 4.32 (2.89-6.62), P < 0.001]. These differences in mortality between phenotypes were consistent when patients were stratified based on demographics, aetiology, admission diagnoses, mortality risk scores, shock severity and systolic function. One-year mortality occurred in 31.5% and differed between the phenotypes: uncomplicated, 11.9% (reference); inflamed, 26.8% [adjusted hazard ratio (HR) 1.56 (1.27-1.92), P < 0.001]; iron-deficient, 33.8% [adjusted HR 2.47 (2.00-3.04), P < 0.001]; cardiorenal, 41.2% [adjusted HR 2.41 (1.97-2.95), P < 0.001]; and hypoperfused, 52.3% [adjusted HR 3.43 (2.82-4.18), P < 0.001]. Similar findings were observed for post-discharge 1 year mortality. Conclusions Unsupervised machine learning clustering can identify multiple distinct clinical HF phenotypes within the CICU population that display differing mortality profiles both in-hospital and at 1 year. Mortality was lowest for the uncomplicated HF phenotype and highest for the hypoperfused phenotype. The inflamed phenotype had comparatively higher in-hospital mortality yet lower post-discharge mortality, suggesting divergent short-term and long-term prognosis.