Abstract Number: Esoc2026a1157 Prediction of Atrial Fibrillation Using Deep Learning From a Single-Lead Sinus Rythm Electrocardiogram Is Associated With Future Ischemic Stroke or Systemic Embolism

Abstract Background and aims Atrial fibrillation (AF) is a major cause of ischemic stroke. Previous studies have shown that Artificial Intelligence (AI) can predict incident AF from single-lead sinus rhythm ECGs by detecting electrical signatures of atrial disease. However, the clinical actionability of early AF detection remains debated. In this registry-based study, we tested the hypothesis that stroke incidence differs between patients predicted at high- versus low-risk of AF by a previously published deep learning (DL) model. Methods A French national database was used to construct a retrospective cohort of 11,024 patients who underwent ambulatory ECG monitoring without detected AF (Table 1). Patients were followed up to 5 years (mean follow-up 1.6 years) for ischemic stroke or systemic embolism. The high- versus low-risk threshold was determined on an independent cohort. The cumulative stroke incidence was compared between groups. Results Patients classified as high AF risk by the model (age: 70.7 ± 13 yo; sex: 46.8 % female) exhibited a 1.7-fold and 2.1-fold higher risk of ischemic stroke or systemic embolism at 1 and 4 years, respectively (Figure 1). Saliency maps highlight ECG regions contributing most strongly to AF risk prediction (Figure 2). Conclusions AI-based AF risk prediction from single-lead sinus rhythm ECGs identifies patients at significantly increased risk of ischemic stroke. These findings support prospective studies evaluating anticoagulant therapy for primary stroke prevention in patients in sinus rhythm with high predicted AF-related stroke risk. Conflict of interest Thomas Proudhon: Philips employee; Amine Keldhouni: Philips employee; Clémence Blanc: nothing to disclose; Baptiste Lefebvre: Philips employee; Laurent Fiorina: medical expert for Philips. Table 1 - belongs to Methods Figure 1 - belongs to Results Figure 2 - belongs to Conclusions