Abstract 4366137: Multimodal AI Signatures of Cardiac Remodeling Independently Predict Worse Aortic Stenosis Prognosis

Abstract

Introduction: Aortic stenosis (AS) follows a progressive course, with delays in diagnosis linked to worse outcomes. Traditional AS categorization relies on functional Doppler parameters, which may not adequately define structural and electrophysiological remodeling linked to a higher risk of progression and adverse outcomes. Research Questions: Multimodal integration of AI-enabled digital biomarkers of structural, functional, and electrophysiological remodeling may enable greater precision in the phenotyping of AS risk. Methods: We included 61,812 individuals from the UK Biobank who prospectively underwent 12-lead electrocardiography (ECG) and cardiac magnetic resonance (CMR) imaging. We quantified 3 AI-enabled digital biomarkers of AS risk: 1) Electrophysiological signature (AI-ECG score) : defined as the cosine similarity (0-1) between an individual’s ECG embedding and a typical ECG embedding for AS cases as derived from a foundation model trained across a large U.S. health system; 2) Structural signature : defined as the validated Digital AS Severity index (DASSi, 0-1), a biomarker of cardiac remodeling in AS measurable on long-axis cine-CMR clips; and 3) Flow signature : defined as the peak aortic ejection velocity on velocity-encoded CMR using DeepFlow. We examined the age/sex-adjusted, independent association of the 3 biomarkers with prevalent AS, as well as incident aortic valve replacement (AVR) and mortality ( Figure 1 ). Results: Among 61,812 individuals (mean age 66±8 years, n=32,167 [52.0%] female), 246 (0.40%) had prevalent AS at baseline. Each SD increase in the electrophysiological (AI-ECG score), structural (DASSi), and flow (peak velocity) biomarkers was associated with 1.4-, 1.8- and 2.2-fold higher odds of prevalent AS ( Figure 2 ). Over a median follow-up of 4.6 [IQR: 1.3-6.3] years, 73 (0.12%) individuals underwent AVR and 1037 (1.7%) died. All 3 biomarkers independently predicted AS progression based on incident AVR (per 1 SD: 2.56 [95%CI: 2.16-3.04], 2.56 [2.03-3.25] and 1.61 [1.24-2.08], respectively) ( Figure 3a ). AI-ECG score (HR 1.08 [95%CI: 1.01-1.15]) and DASSi (1.17 [1.10-1.25]), but not peak velocity (1.04 [0.98-1.10]) were associated with mortality ( Figure 3b ). Conclusion: We propose an AI-enabled multimodal approach using electrophysiological, structural and flow-related signatures to enable precision phenotyping of future AS risk. The scalability of AI-ECG and DASSi to point-of-care devices may define a novel paradigm for early AS monitoring.