Abstract 11938: Performance of ECG-Based Artificial Intelligence for Atrial Fibrillation Risk Prediction Across the Demographic Spectrum
Abstract
Background: Atrial fibrillation (AF) is common and highly morbid. By estimating future AF risk, ECG-based artificial intelligence (ECG-AI) models may prioritize individuals for screening and related preventive efforts. However, understanding performance variation across the broad demographic spectrum for which AF risk estimation may be relevant is needed prior to implementation. Methods: We evaluated ECG-AI, a previously validated deep learning model for AF risk prediction originally trained in over 100,000 patients at Massachusetts General Hospital (MGH), on a separate cohort of patients with longitudinal primary or cardiology care from MGH and Brigham and Women’s Hospital (BWH) between 2000 and 2019. In subgroups of age (by decade), sex, and race, we computed discrimination of incident AF using the areas under the receiver operating characteristic (AUROC) and precision-recall curve (AUPRC), and model calibration using integrated calibration index (ICI). As a reference, we performed an analogous assessment of the validated 11-component CHARGE-AF clinical risk score, among individuals with available data. Results: Among 76,967 participants (mean age 55.2 years, 56.7% women), ECG-AI demonstrated good discrimination of 10-year incident AF overall (AUROC 0.810 [95% CI: 0.805-0.816]). Discrimination was consistent across sex (women: AUROC 0.806 [0.797-0.814]; men: 0.802 [0.793-0.810]). Discrimination persisted across age strata, with modestly higher AUROC in younger groups (e.g., AUROC 0.756 [0.719, 0.790] in ages 30-39 vs. 0.714 (0.698, 0.730) in ages 70-79) (Figure 1). Discrimination was similar across race (White: AUROC 0.806 [0.800, 0.813]; non-White: 0.819 [0.806, 0.831]). When compared to CHARGE-AF in 58,413 participants, ECG-AI had more consistent performance across sex, age, and race, with the largest relative improvements observed among younger age groups (Figure 2). Conclusions: Estimation of future AF risk using ECG-AI is highly consistent across sex and race. Although discrimination was higher in younger patients, performance was better preserved than a validated clinical risk factor score. Findings support the application of ECG-AI across broad demographic groups to prioritize individuals for AF screening and prevention.