Abstract Body (Do not enter title and authors here): Genome-wide association studies (GWAS) of cardiac structure and function have historically relied on a limited set of manually derived echocardiographic measurements, and more recently deep-learning derived features from CT/MRI. Despite representing the most common cardiovascular imaging modality, clinical echo archives remain largely untapped for genomic analysis at scale. We applied PanEcho, a deep learning model, to automate phenotyping of routine Echo across a diverse cohort from the Penn Medicine Biobank (PMBB).
PanEcho processed 68,637 images/videos across 8,036 echocardiographic studies from 4,987 PMBB participants (European (EUR)= 3,342, African (AFR)= 1,413), using 16-frame cine clip per study or up to 16 randomly selected stills if video was unavailable. Frame- or image-level predictions were averaged to produce a single study-level value for each trait. Model accuracy was evaluated by mean absolute error (MAE) against cardiologist-reported values. Predictions were averaged at the trait-level when individuals had multiple studies. We performed GWAS using SAIGE, adjusting for sex, age, and the first 6 genetic PCs.
Across the 21 echo traits, PanEcho had MAE ranging from 0.16 units (e.g. interventricular septal thickness) to 33.3 mL (left ventricular end-diastolic volume); fine-tuning could further reduce error for several traits. In EUR GWAS, we identified known cardiomyopathy loci with a trait-consistent effect: loci on BAG3 (rs2234962) was nominally associated with increased ejection fraction (β=0.063, P=3.1×10^-2). rs80076162 in CASP7 was associated with increased peak aortic valve velocity (AVPkVel; p=2.7×10^-6), suggesting a possible link to valvular flow dynamics. In AFR group, rs11153734 upstream of PLN was associated with aortic root diameter (p=1.05×10^-6). Two novel loci reached genome-wide significance: rs13079713 near EPHB1 for global longitudinal strain (EUR; β=-0.14, p=3.7×10^-8) and rs2467493 near CA10 with higher AVPkVel (AFR; β=0.31, p=1.9×10^-8). These findings suggest novel biological contributors to myocardial function and valvular flow across diverse populations.
Here we present the first large-scale GWAS using automated deep-learning phenotyping of routine echocardiograms. Our results highlight known and novel cardiac loci, show a scalable route to integrate imaging and genomics, refine myocardial biology, and improve polygenic risk prediction for precision cardiovascular medicine.