Multi-national assessment of clinical factors, genetic scores, and electrocardiogram-based artificial intelligence for atrial fibrillation prediction
S Kany, E Aaltonen, S F Friedman, J German, J T Raemoe, L C Weng, S Koyama, M S Kim, A Shahid, A Fahed, A Palotie, N Mars, M Maddah, P T Ellinor, S KhurshidAbstract
Background
Atrial fibrillation (AF) is a primary driver of global cardiovascular morbidity, yet screening strategies based solely on age often result in modest yields. AF risk estimation is feasible using clinical factors, common genetic variation, and artificial intelligence (AI)-enabled electrocardiogram (ECG) analysis. It remains unknown whether integrating these distinct risk signals representing systemic comorbidity, inherited predisposition, and subclinical electrical remodeling improves AF risk estimation across diverse population-based and healthcare-related cohorts.
Methods
We leveraged data from 64,799 individuals across three independent prospective cohorts representing distinct countries and health setting: the UK Biobank (UKB; N=49,293), Mass General Brigham Biobank (MGBB; N=12,812), and FinnGen (N=2,694). We estimated 5-year incident AF risk using three previously validated approaches: the CHARGE-AF clinical score, a contemporary AF polygenic risk score (PRS), and a deep learning-based AF risk model (ECG-AI). We assessed discrimination using time-dependent area under the receiver operating characteristic curve (AUROC) and average precision (AP). In UK Biobank, we integrated models into two-component combinations as well as all three (“Predict-AF3”), with each composite model evaluated in all three samples.
Results
At 5 years, cumulative AF incidence was 2.4% (UKB), 8.7% (MGBB), and 6.9% (FinnGen). Notably, we observed minimal correlation between the PRS and either clinical factors or ECG-AI (r≤0.04 across all cohorts), suggesting that genetic risk represents a unique dimension of AF vulnerability. The integrated Predict-AF3 model achieved the highest numerical discrimination across all cohorts (UKB AUROC 0.817 [0.802-0.832], AP 0.110; MGBB AUROC 0.750 [0.733-0.767], AP 0.232 [0.211-0.258]; FinnGen AUROC 0.857 [0.816-0.894], AP 0.268), although the CHARGE-AF+ECG-AI model performed well (e.g., FinnGen AUROC 0.847). There was a graded increase in 5-year AF incidence with the number of individual models predicting high-risk status (e.g., MGBB 26.12% [22.79-29.32%] with three components estimating ≥10% risk vs. 2.39% [1.91-2.87%] with zero). Predict-AF3 also significantly improved net reclassification, particularly in FinnGen where it outperformed the clinical+genetic model (NRI 0.121 [95% CI 0.051-0.212]).
Conclusions
Clinical, genetic, and AI-derived risk signals provide complementary information for AF prediction. Although consideration of all factors is ideal, and identifies a subset of individuals at exceedingly high risk for AF development, certain two-component models can achieve high performance. Optimal methods of AF risk estimation likely vary based on intended setting and available data.Figure 2figure 1