Predicting incident atrial fibrillation from sinus rhythm electrocardiograms using deep learning: a meta-analysis with horizon stratification

Background/Introduction Atrial fibrillation (AF) frequently remains clinically silent until complications occur, limiting opportunities for early prevention. Deep learning (DL) analysis of sinus rhythm (SR) electrocardiograms has emerged as a novel approach for identifying individuals at risk of developing AF, but reported performance varies and has not been systematically quantified. Purpose To systematically evaluate and quantitatively synthesise the performance of DL models applied to SR electrocardiograms for the prediction of incident AF. Methods A systematic review and meta-analysis were conducted in accordance with PRISMA guidelines. PubMed and the Cochrane Library were searched for peer-reviewed studies published since 2010 evaluating DL models applied to sinus rhythm electrocardiograms for the prediction of incident AF. Eligible studies reported the area under the receiver operating characteristic curve (AUROC) with 95% CI. Risk of bias was assessed using a PROBAST-aligned framework. Random-effects meta-analysis was performed to pool AUROC estimates, with prespecified subgroup analyses by prediction horizon. Results A total of 796 records were identified through database searching. Thirteen studies met the criteria for qualitative synthesis and are summarised in Table 1. Eight studies provided sufficient data for quantitative synthesis and were included in the meta-analysis (Figure 1), representing more than 1.35 million patients and over 5.0 million electrocardiograms or ECG segments. The pooled random-effects AUROC was 0.789 (95% CI, 0.764–0.813), indicating moderate-to-good overall discrimination, with substantial heterogeneity (I² ≈ 96%). A fixed-effect analysis yielded a similar central estimate (AUROC ;= 0.793, 95% CI, 0.790–0.797). Short-term prediction horizons of 31 days or less demonstrated the highest pooled discrimination (AUROC 0.815, 95% CI 0.747–0.868), while intermediate-term horizons (31 days to 1 year) showed moderate performance (AUROC 0.777, 95% CI 0.733–0.815). Longer prediction horizons of one year or more were associated with lower discrimination (AUROC = 0.737; 95% CI, 0.622–0.827). Models using standard 12-lead electrocardiograms generally reported higher discrimination than those using single-lead or wearable recordings. Risk of bias was most frequently observed in the outcome definition and analysis domains. Conclusion DL models applied to SR electrocardiograms predict incident AF with good discrimination, particularly for short-term prediction. Substantial heterogeneity limited external validation, and the absence of outcome-driven evidence highlights the need for prospective validation and randomised trials before routine clinical implementation.