Abstract 063: Short-term Repeatability of Artificial Intelligence Estimated Electrocardiographic Age

Abstract

Background: Artificial intelligence (AI) models produce precise interpretations of electrocardiogram (ECG) data and can estimate cardiac age from raw ECG waveforms (ECG-age). Although research suggests potential for cardiovascular disease (CVD) risk assessment, research on the short-term repeatability of ECG-age over time is scarce. Assessing the short-term repeatability of ECG-age is essential for establishing its precision and stability. Methods: We trained a convolutional neural network machine learning model to predict ECG-age using 18,869 patients from the publicly available German PTB-XL dataset with 10-second digital 12-lead ECG recordings. The model was adapted for one-dimensional signals to capture how aging impacts ECG waveforms, resembling a residual network used in image classification. The model was applied to two 10-second 12-lead ECGs taken at each of two separate visits, collected 1-2 weeks apart, from participants at the University of North Carolina at Chapel Hill’s General Clinical Research Center. The intraclass correlation coefficient (ICC), standard error of measurement (SEM), and minimally detectable change (MDC) estimated repeatability. Estimated variance was decomposed into between-participant, between-visit, and within-visit components. Results: Data to estimate ECG-age were available for 58 participants free of cardiovascular or metabolic conditions (mean age = 52±5 years; 55% female; 66% White). The mean (SD) ECG-age at visit 1 and visit 2 were 43.5 (4.1) and 44.0 (4.5) years, respectively. ECG-age demonstrated moderate repeatability between visit 1 and visit 2 (ICC=0.64; 95% confidence interval: 0.52, 0.76). The SEM was 2.6 years and MDC was 7.2 years. Between-participant variance accounted for the largest source of variation, followed by within-visit variation (Table 1). Discussion: ECG-age had moderate short-term repeatability, with consistency between visits. Notable within-visit variation suggests that measurement error or environmental and technical factors may influence repeated ECG-age assessments. These findings provide an important foundation for validating this novel AI metric in CVD research. Future studies should focus on strategies to minimize within-visit variation, improving the reliability of ECG-age as a potential clinical tool.