Abstract 10515: Deep Phenotyping of Heart Failure with Machine Learning: An Echocardiographic Community Study

Abstract

Introduction: Heart failure (HF) is a clinical syndrome with high mortality. Left ventricular ejection fraction (EF) determines clinical phenotyping and HF is classified most often as preserved (HFpEF) or reduced (HFrEF) EF. The importance of expanding phenotyping beyond EF is recognized. We hypothesized that machine learning applied to echocardiography will provide information important for deep phenotyping. Methods: We collected clinical and echo data in a HF community cohort between 2013 and 2017. Unsupervised clustering was applied to 23 echo variables using partitioning around medoids (PAM), with Gower dissimilarity used to address the mixed variable types. Proportional hazards regression examined the association between cluster assignment and death. Results: We identified 4978 persons with HF [mean age 73.9 years (SD 13.9), 47% women] with echocardiography within 6 months of diagnosis. We identified 2 primary clusters using echo variables, as depicted using t-Distributed Stochastic Neighbor Embedding (t-SNE) applied to the dissimilarity matrix (Figure). Diastolic function variables exhibited the largest differences by clusters as shown by absolute standardized differences (Table). After 3.2 + 1.9 years of follow up, 2001 deaths occurred. Assignment to cluster 2 was associated with a near 20% excess risk of death independent of age, sex, comorbidity, and EF [HR 1.17 (1.07, 1.28) p<0.001]. Conclusions: Unsupervised clustering applied to echocardiography in a large community HF cohort identified 2 distinct clusters, which defined HF subgroups with different characteristics and were independently associated with risk of death.