Forecasting 1-year atrial fibrillation risk from sinus rhythm ECGs in Arrhythmia Naïve Patients

Background Prior studies used binary classifiers to identify patients at risk of developing atrial fibrillation (AF) from normal electrocardiograms (ECGs), overlooking clinically relevant time-to-event information. Objective To develop a survival-based model to estimate AF risk from normal ECGs in patients without prior AF and stratify them into risk groups up to one year before AF onset. Methods This retrospective cohort included patients with ECGs evaluated at the University of Pittsburgh Medical Center between 2022 and 2023 (index visit). We identified patients with prior normal sinus rhythm ECGs within one year before the index visit, where primary outcome was new-onset AF detected on the index ECG. From prior ECGs, we extracted 417 features and trained multiple survival models to predict incident AF. Data were split into training (60%), validation (15%), and test (25%) sets. The best-performing model, by validation C-index, stratified patients into low-, intermediate-, and high-risk groups. Kaplan–Meier curves evaluated time-to-AF across risk groups. Interpretability analyses identified ECG features associated with AF risk. Results Among 26,015 patients (age 62±19; females 52%; average follow-up 80±100 days; AF incidence 23.1%), the gradient boosting survival model achieved highest performance (test C-index 0.781). It stratified 72.8%, 23.2%, and 4.0% of patients into low, intermediate, and high-risk groups, with AF incidence rates of 5.1%, 21.5%, and 49.8% (log-rank χ 2 =838.9, p<0.0001). Predictors included older age, lower P wave amplitude, and increased heart rate variability. Conclusion Survival modeling of normal sinus ECGs enables time-dependent AF risk estimation and clinically actionable risk-group stratification, extending beyond binary classification.