Abstract / Summary
Background: Accurate prediction of atrial fibrillation (AF) is essential for prevention. The CHARGE-AF score, based on routinely available clinical factors, provides a practical tool for estimating AF risk but has limited predictive accuracy. This study aimed to improve AF risk prediction by integrating artificial intelligence (AI)-based electrocardiogram (ECG) analysis of age and sex with genetic information in addition to established clinical risk factors.
Methods: We analyzed 39 478 UK Biobank participants without prior AF. A polygenic risk score for AF (AF-PRS), the AI-ECG age gap (AI-ECG-predicted age minus chronological age), and AI-ECG-predicted sex mismatch were evaluated on top of the CHARGE-AF model. Model performance was compared across sequential prediction models.
Results: Over a median follow-up of 2.7 years (interquartile range, 1.7-4.2; maximum, 6.7), 533 participants (1.3%) developed AF. AF-PRS (HR 1.61, 95% CI: 1.48-1.76) and the AI-ECG age gap (HR 1.37, 95% CI: 1.24-1.51) were independently associated with incident AF. Compared with CHARGE-AF alone (C-index 0.708, 95% CI: 0.686-0.730), adding AF-PRS improved discrimination (C-index 0.743; Δ0.035, 95% CI: 0.021-0.049; p < 0.001). Adding traditional ECG parameters did not enhance performance, whereas incorporating AI-ECG features (age gap and sex mismatch) into the combined clinical and genetic model provided additional gain in discrimination (C-index 0.754; Δ0.045, 95% CI: 0.028-0.062), reclassification (net reclassification improvement 0.408), and discrimination improvement (0.007; all p < 0.001).
Conclusions: Integrating genetic risk and AI-ECG-derived features enhances AF prediction beyond an established clinical model, supporting the use of combined digital and genetic biomarkers for risk stratification.