Unsupervised Clustering Applied to Electronic Health Record-derived Phenotypes in Patients with Heart Failure

BackgroundHigh-dimensional electronic health records (EHR) data can be used to phenotype complex diseases. The aim of this study is to apply unsupervised clustering to EHR-based traits derived in a cohort of patients with heart failure (HF) from a large integrated health system. MethodsUsing the institutional EHR, we identified 8569 patients with HF and extracted 1263 EHR-based input features, including clinical, echocardiographic, and comorbidity data, prior to the time of HF diagnosis. Principal component analysis, Uniform Manifold Approximation and Projection, and spectral clustering were applied to the input features after sex stratification of the cohort. The optimal number of clusters for each sex-stratified group was selected by highest Silhouette score and by within-cluster and between-cluster sums of squares. Determinants of cluster assignment were evaluated. ResultsWe identified four clusters in each of the female-only (44%) and male-only (56%) cohorts. Sex-specific cohorts differed significantly by age of HF diagnosis, left ventricular chamber size, markers of renal and hepatic function, and comorbidity burden (all p<0.001). Left ventricular ejection fraction was not a strong driver of cluster assignment. ConclusionReadily available EHR data collected in the course of routine care can be leveraged to accurately classify patients into major phenotypic HF subtypes using data driven approaches.