AI-enabled ECG risk predicts incident atrial fibrillation with incremental value of AI-estimated heart age: comparison with the CHARGE-AF model from the UK Biobank
Y Baek, P S Yang, S C Lee, W I Choi, B Y Joung, D H KimAbstract
Background/Introduction
Atrial fibrillation (AF) is the most common sustained arrhythmia and is associated with substantial morbidity and mortality. The CHARGE-AF score is a validated clinical model based on demographic and clinical factors but does not incorporate electrocardiographic features reflecting atrial remodeling. Artificial intelligence (AI) applied to ECGs may capture subtle electrical signatures of atrial vulnerability, improving early prediction of AF.
Purpose
This study aimed to evaluate whether an AI-enabled ECG risk score (SmartECG-AF) improves AF prediction compared with the CHARGE-AF model and to assess the incremental value of AI-estimated Heart Age for enhancing predictive performance.
Methods
We analyzed 45,239 UK Biobank participants who had baseline 12-lead ECGs recorded between 2006 and 2010, without prior AF. Participants were followed through December 2023 for incident AF events. AF risk during sinus rhythm was quantified using an AI-driven ECG analysis model applied to baseline ECG waveforms. Model discrimination was assessed using receiver-operating characteristic (ROC) analysis and Harrell’s C-index, with CHARGE-AF as the reference. Fine–Gray competing-risk models estimated 5-year AF incidence across 1-year predicted risk strata (low 0–1%, intermediate 1–2%, high ≥2%).
Results
During a median follow-up of 7.6 years (IQR 6.2–8.9), 1,383 participants (3.06%) developed new-onset AF. AF risk estimated by AI provided improved discrimination compared with CHARGE-AF (AUC 0.719, 95% CI 0.695–0.743 vs. 0.708, 95% CI 0.687–0.729). Adding age further improved performance (AUC 0.740, C-index 0.751), and inclusion of AI-estimated Heart Age achieved the highest accuracy (AUC 0.757, C-index 0.757) (Figure 1). Fine–Gray model–based prediction curves demonstrated an exponential increase in 1-year AF risk with higher AI-ECG scores and older age, consistent across age strata (Figure 2A). When stratified by predicted 1-year AF risk, the 5-year cumulative AF incidence was 2.1%, 7.1%, and 12.8% for low-, intermediate-, and high-risk groups (Gray’s test p < 0.001), indicating significant long-term stratification of AF risk (Figure 2B).
Conclusion
AF risk quantified by an AI-driven ECG analysis model effectively predicts incident AF and outperforms the CHARGE-AF model, especially when combined with AI-estimated Heart Age. This approach transforms a standard ECG into a dynamic, individualized biomarker for early detection and prevention of AF.