Abstract
BACKGROUND: Multiple risk scores and biomarkers have been proposed for the prediction of atrial fibrillation (AF), but it is unknown how these compare with each other and whether they could be combined.
OBJECTIVE: This study aimed to evaluate and compare approaches for incident AF prediction.
METHODS: The artificial intelligence-enhanced electrocardiogram risk estimator-AF (AIRE-AF), a convolutional neural network with a discrete-time survival loss function, was developed to predict incident AF. It was trained using a dataset of 1,163,401 electrocardiograms from 189,539 patients from the Beth Israel Deaconess Medical Center and externally validated in the UK Biobank (n = 38,892). AIRE-AF was compared with other risk prediction approaches including the Cohorts for Heart and Aging Research in Genomic Epidemiology (CHARGE)-AF, a clinical risk score.
RESULTS: In the Beth Israel Deaconess Medical Center cohort, AIRE-AF predicted incident AF with a C-index of 0.750 (0.743-0.758). AIRE-AF was superior to CHARGE-AF, left atrial size, and N-terminal pro-B-type natriuretic peptide. The addition of CHARGE-AF and left atrial size provided a minor improvement in performance (C-index improvement 0.017). There was no additive value of N-terminal pro-B-type natriuretic peptide in combination with AIRE-AF. The single best-performing single predictor in the volunteer population (UK Biobank) was CHARGE-AF (C-index 0.750 [0.734-0.769]). The best-performing combination of 2 predictors was AIRE-AF and CHARGE-AF (C-index 0.768 [0.743-0.792]). The addition of a polygenic risk score to AIRE-AF and CHARGE-AF provided a further significant improvement in performance (C-index 0.791 [0.766-0.816]).
CONCLUSION: We present the first comprehensive evaluation of methodologies for predicting incident AF. Risk prediction with a model including AIRE-AF and CHARGE-AF resulted in similar performance to more complex models.