https://doi.org/10.1007/s00392-025-02625-4
1Universitäres Herz- und Gefäßzentrum Hamburg Klinik für Kardiologie Hamburg, Deutschland; 2Broad Institute of MIT and Harvard Cambridge, USA
Background: Atrial fibrillation (AF) risk estimation is feasible using clinical factors, inherited predisposition, and artificial intelligence (AI)-enabled electrocardiogram (ECG) analysis. However, whether integrating these distinct risk signals improves AF risk estimation is not known.
Methods: In the UK Biobank prospective cohort study, we estimated AF risk using three models derived from external populations: the well-validated Cohorts for Aging in Heart and Aging Research in Genomic Epidemiology AF (CHARGE-AF) clinical score, a 1,113,667-variant AF polygenic risk score (PRS), and a published AI-enabled ECG-based AF risk model (ECG-AI). We estimated discrimination of 5-year incident AF using time-dependent area under the receiver operating characteristic (AUROC) and average precision (AP). We additionally validated our results in hospital-based Mass General Biobank (~30,000 individuals with ECG) from the USA and the population-based FinnGen cohort study from Finland (~50,000 individuals with ECG).
Results: Among 49,293 individuals (mean age 65±8 years, 52% women), 825 (2.4%) developed AF within 5 years. Using single models, discrimination of 5-year incident AF was higher using ECG-AI (AUROC 0.705 [95%CI 0.686-0.724]; AP 0.085 [0.071-0.11]) and CHARGE-AF (AUROC 0.785 [0.769-0.801]; AP 0.053 [0.048-0.061]) versus the PRS (AUROC 0.618, [0.598-0.639]; AP 0.038 [0.028-0.045]). The inclusion of all components (“Predict-AF3”) was the best performing model (AUROC 0.817 [0.802-0.832]; AP 0.11 [0.091-0.15], p<0.01 vs CHARGE-AF+ECG-AI), followed by the two-component model of CHARGE-AF+ECG-AI (AUROC 0.802 [0.786-0.818]; AP 0.098 [0.081-0.13]). Using Predict-AF3, individuals at high AF risk (i.e., 5-year predicted AF risk >2.5%) had a 5-year cumulative incidence of AF of 5.83% (5.33-6.32). At the same threshold, the 5-year cumulative incidence of AF was progressively higher according to the number of models predicting high risk (zero: 0.67% [0.51-0.84], one: 1.48% [1.28-1.69], two: 4.48% [3.99-4.98]; three: 11.06% [9.48-12.61]), and Predict-AF3 achieved favorable net reclassification improvement compared to both CHARGE-AF+ECG-AI (0.039 [0.015-0.066]) and CHARGE-AF+PRS (0.033 [0.0082-0.059]). The results of the Mass General Biobank and FinnGen cohorts will be reported at the time of presentation.
Conclusions: Integration of clinical, genetic, and AI-derived risk signals improves discrimination of 5-year AF risk over individual models. Comprehensive risk prediction Models such as Predict-AF3 have substantial potential to improve prioritization of individuals for AF screening and preventive interventions.
Panel A depicts the model discrimination for a 5-year window of incident atrial fibrillation (AF) using a polygenic risk score (PRS) model (teal), an ECG-derived artificial intelligence (AI) prediction (turquoise) and a model based on the CHARGE-AF score (orange). Panel B depicts the model discrimination for a 5-year window of incident atrial fibrillation (AF) using a model combining PRS and CHARGE-AF (black), a model combining an ECG-derived artificial intelligence (ECG-AI) prediction and the PRS (grey), a model combining ECG-AI and the CHARGE-AF (orange), and a model combining ECG-AI, CHARGE-AF and the PRS (red). Panel C depicts the cumulative risk of AF across strata of high risk by each model where high AF risk is defined as 5-year AF risk >2.5% (approximating the top tertile of risk).