Abstract 13745: Deep Learning of Raw Echocardiography Videos Recovers HFpEF-Specific Prognostic Signal: The Atherosclerosis Risk in Communities Study

Abstract

Background: Echocardiography captures rich spatiotemporal information about cardiac structure and function that is incompletely captured by routine structured measurements. We evaluated whether deep learning applied directly to raw echo video can recover prognostic information for incident HF beyond conventional echo measurements, and whether this benefit differs by HF phenotype. Methods: Among 5,732 HF-free participants in the community-based ARIC cohort who underwent protocol echocardiography at study Visit 5 (2011-2013), we evaluated the prognostic performance of four models for adjudicated incident HFpEF (LVEF ≥50%) or HFrEF (LVEF <50%) using a cause-specific framework: clinical only (M1), structured echo only (M2), clinical + echo (M3), and a 3D convolutional neural network (CNN) trained on raw apical 4-chamber and parasternal long-axis clips using Cox partial-likelihood loss (M4). Clinical variables included age, sex, race, BMI, eGFR, NT-proBNP, diabetes, prior CHD/MI, atrial fibrillation, and smoking status. Structured echo variables included LVEF, LV mass, LAVi, E/e′, E/A ratio, A velocity, and longitudinal strain. Model performance was assessed using 10-fold cross-validated Harrell’s C-index. Grad-CAM was used to localize the spatiotemporal regions driving CNN predictions. Results: Mean age was 75±5 years, 59% were female, and 22% reported Black race. Over a median follow-up of 10.6 [IQR 6.8-11.6] years, HFpEF occurred in 451 and HFrEF in 359 participants. Based on the C-index, the structured echo model (M2) improved HFrEF prediction over clinical variables alone (M1) but did not improve HFpEF prediction ( Figure ). Deep learning (M4) significantly improved prediction over structured echo alone (M2) and over clinical + structured echo combined (M3) for HFpEF but not HFrEF. Grad-CAM localized attention to the basal-to-mid septal myocardium in HFpEF and to the mitral annulus, mitral valve, and basal LV cavity in HFrEF. Conclusions: A 3D CNN trained on raw echo video recovered prognostic signal for HFpEF beyond clinical and structured echocardiographic variables combined, while providing no incremental information for HFrEF risk, consistent with structured echo measures more fully capturing HFrEF-relevant cardiac dysfunction. These findings support explainability-driven analyses to interrogate HFpEF pathophysiology and identify novel imaging biomarkers of HFpEF risk.