Deep learning enabled analysis of cardiac sphericity

Quantification of chamber size and systolic function is a fundamental component of cardiac imaging, as these measurements provide a basis for establishing both diagnosis and appropriate treatment for a spectrum of cardiomyopathies. However, the human heart is a complex structure with significant uncharacterized phenotypic variation beyond traditional metrics of size and function. Characterizing variation in cardiac shape and morphology can add to our ability to understand and classify cardiovascular risk and pathophysiology. We describe deep learning enabled measurement of left ventricle (LV) sphericity using cardiac magnetic resonance imaging data from the UK Biobank and show that among adults with normal LV volumes and systolic function, increased sphericity is associated with increased risk for incident atrial fibrillation (HR 1.31 per SD, 95% CI 1.23-1.38), cardiomyopathy (HR 1.62 per SD, 95% CI 1.29-2.02), and heart failure (HR 1.24, 95% CI 1.11-1.39), independent of traditional risk factors including age, sex, hypertension, and body mass index. Using genome-wide association studies, we identify four loci associated with sphericity at genome-wide significance. These loci harbor known and suspected cardiomyopathy genes. Through genetic correlation and Mendelian randomization, we provide evidence that sphericity may represent a subclinical manifestation of non-ischemic cardiomyopathy.