Test-Retest Reliability of Artificial Intelligence-Enhanced Electrocardiography: A Multi-Center Study

作者：Lovedeep Singh Dhingra, Philip M. Croon, Bruno Batinica, Evangelos K. Oikonomou, Rohan Khera · 发表于：medRxiv · 年份：2025 · DOI：10.1101/2025.11.04.25339526 · 被引用次数：1 · 研究领域：ECG Monitoring and Analysis、Cardiac electrophysiology and arrhythmias、Atrial Fibrillation Management and Outcomes

ABSTRACT Background Artificial intelligence-enhanced electrocardiography (AI-ECG) enables structural heart disease (SHD) detection. However, its utility as a clinical assay requires consistency of outputs when repeated under similar conditions. Objectives To characterize test-retest reliability of contemporary AI-ECG models across diverse health systems, identify factors associated with discordance, and determine the predictive significance of screen status changes. Methods We identified ECG pairs recorded 1-30 days apart in the same individual at the Yale-New Haven Health System (YNHHS), Massachusetts General Hospital (MGH), and Emory University Hospital (EUH). We evaluated internally developed ECG signal- and image-based SHD models, and the EchoNext-Mini model, including disease-specific components and ensemble composites. Reliability was quantified with concordance correlation coefficients (CCCs) and categorical concordance percentages. In patients without prevalent heart failure (HF) with serial ECGs 30-90 days apart, we evaluated association of screen status discordance with new-onset HF risk. Results We included 731,466 ECG pairs (median interval 5-6 days). At YNHHS, disease-specific model CCCs ranged from 0.77-0.86 across the signal- and image-based model families and 0.50-0.97 for EchoNext-Mini output nodes. Composite SHD models had CCCs of 0.90 (signal-based), 0.90 (image-based), and 0.81 (EchoNext-Mini). The image-based ensemble model achieved categorical screen-sta...