Identifying high-risk Fontan phenotypes using K-means clustering of cardiac magnetic resonance-based dyssynchrony metrics

Abstract

Background: Individuals with a Fontan circulation encompass a heterogeneous group with adverse outcomes linked to ventricular dilation, dysfunction, and dyssynchrony. The purpose of this study was to assess if unsupervised machine learning cluster analysis of cardiovascular magnetic resonance (CMR)-derived dyssynchrony metrics can separate ventricles in the Fontan circulation from normal control left ventricles and identify prognostically distinct subgroups within the Fontan cohort. Methods: This single-center, retrospective study used 503 CMR studies from Fontan patients (median age 15 y) and 42 from age-matched controls from January 2005 to May 2011. Feature tracking on short-axis cine stacks assessed radial and circumferential strain, strain rate, and displacement. Unsupervised K-means clustering was applied to 24 mechanical dyssynchrony metrics derived from these deformation measurements. Clusters were compared for demographic, anatomical, and composite outcomes of death, or heart transplantation. Results: Four distinct phenotypic clusters were identified. Over a median follow-up of 4.2 y (interquartile ranges 1.7–8.8 y), 58 (11.5%) patients met the composite outcome. The highest-risk cluster (largely comprised of right or mixed ventricular morphology and dilated, dyssynchronous ventricles) exhibited a higher hazard for the composite outcome compared to the lowest-risk cluster while controlling for ventricular morphology (hazard ratio [HR] 6.4; 95% confidence interval [CI] 2.1–19.3; P value 0.001) and higher indexed end-diastolic volume (HR 3.2; 95% CI 1.04–10.0; P value 0.043) per 10 mL/m2. Conclusion: Unsupervised machine learning using CMR-derived dyssynchrony metrics identified four distinct clusters of patients with Fontan circulation and healthy controls with varying clinical characteristics and risk profiles. This technique can be used to guide future studies and identify more homogeneous subsets of patients from an overall heterogeneous population.

Keywords

• Unsupervised machine learning
• Pediatrics
• Fontan
• Cardiovascular magnetic resonance imaging
• Dyssynchrony