Trained on over 1.1 million echocardiographic videos from nearly 29,000 studies, the deep learning model PanAdipo detected prominent epicardial adipose tissue (EAT) with an AUROC of 0.91—outperforming conventional cardiac measures. Validated across four independent cohorts totaling roughly 22,800 patients, the model's risk scores correlated strongly with CT-derived epicardial fat (Spearman's rho=0.75) but only modestly with BMI (rho=0.19–0.40). In the community-based MESA cohort, higher PanAdipo scores independently predicted elevated HOMA-IR, triglycerides, and incident metabolic-associated steatotic liver disease, even after BMI adjustment (HR 1.25 per 1-SD increment).
Epicardial fat has emerged as a metabolically active depot distinct from generalized obesity—secreting pro-inflammatory adipokines that directly bathe coronary arteries and myocardium. Until now, quantifying EAT required cardiac MRI or CT, limiting clinical scalability. PanAdipo's ability to extract this signal from routine echocardiography—a test performed tens of millions of times annually—could transform risk stratification for cardiovascular-kidney-metabolic syndrome without additional imaging cost or radiation. The BMI independence is particularly clinically significant: it suggests the tool captures metabolic risk invisible to standard anthropometric measures, relevant for the growing population of metabolically unhealthy normal-weight individuals. Limitations include retrospective design, predominantly health-system-based derivation cohorts, and reliance on expert-annotated rather than CT-confirmed labels for model training. As a preprint not yet peer-reviewed, these findings require independent validation before clinical translation. Still, the multi-cohort external validation and mechanistic correspondence with CT imaging make this a genuinely promising, potentially paradigm-shifting cardiometabolic biomarker.
