Predicting recurrence and outcomes after triggered atrial fibrillation using deep learning

Abstract Introduction Triggered atrial fibrillation (AF) is defined as new-onset AF occurring after an acute precipitating event, and may recur months or even years after the initial episode. Though long-term AF recurrence is increasingly recognized as an important marker for increased risk of AF-related adverse outcomes including stroke, risk stratification methods to guide longitudinal rhythm surveillance are limited. Artificial intelligence (AI) derived AF risk based on a 12-lead electrocardiogram has shown substantial predictive utility for incident AF and associated-outcomes and may do so for triggered AF as well. Methods We retrospectively analyzed 3,371 patients in an ambulatory cohort of primary care and cardiology patients with triggered AF occurring during a hospitalization. We examined associations between clinical covariates, common AF triggers, and AF recurrence using Fine-Gray models accounting for death as a competing risk. We investigated the association between AF recurrence (as a time-varying covariate) and a composite endpoint of AF-related adverse events (defined as stroke, heart failure, or all-cause mortality) using Cox proportional hazards models. We then developed and validated a penalized regression model to predict AF recurrence incorporating clinical factors, AF triggers, and predictions from a previously developed 12-lead ECG-based deep learning model that stratifies future AF risk. Results Over a median follow-up of 3.8 years, the 10-year subdistribution cumulative incidence of AF recurrence was 41% (95% CI 39-44). Recurrence rates varied by trigger, with respiratory illness-associated AF exhibiting the highest (46%, 95% CI 40-53) and sepsis-associated AF the lowest (34%, 95% CI 24-44). Time-varying AF recurrence was strongly associated with both increased adverse event risk (HR 2.29, 95% CI 1.87–2.80) as were other clinical factors including diabetes (HR 1.40, 95% CI 1.18-1.67) and current smoking (HR 1.39, 95% CI 1.04-1.84). A predictive model incorporating clinical factors, AF triggers, and deep learning-based AF predictions demonstrated moderate discrimination for AF recurrence (AUROC 0.768, 95% CI 0.707-0.830), surpassing models based on clinical features alone (AUROC 0.707, 95% CI 0.642-0.772). Conclusions AF recurrence is common after triggered AF and is associated with a substantially higher risk of adverse cardiovascular events. While there are no standardized approaches for recurrence risk stratification after triggered AF, models incorporating ECG-based AI predictions improve discrimination of recurrent AF compared to traditional clinical risk assessment and may inform rhythm surveillance strategies and preventive interventions.