ABSTRACT NUMBER: ESOC2026A549 ELECTROCARDIOGRAM AND BRAIN MRI ALLOW INDIVIDUALIZED ESTIMATION OF STROKE RISK. A MULITMODAL DEEP LEARNING ANALYSIS

Abstract Background and aims Identification of cardiovascular risk factors (CVRF) is essential for targeted prevention of stroke. CVRF affect both the heart and the brain. We aimed to identify CVRF and their contribution to stroke risk by analysing electrocardiogram (ECG) and brain magnetic resonance imaging (MRI) using multimodal deep-learning. Methods We trained two deep-learning models on 23’671 participants in the population-based UK Biobank (UKBB) cohort to identify participants with hypertension, atrial fibrillation, diabetes and dyslipidemia. One model is based on T2 FLAIR Brain MRI, the other is based on ECG. We then combined the models. We evaluated our models on a separate test set from the UKBB (n=45’434). Finally, we investigated if the predicted CVRF were associated with incident stroke over a median follow-up of 4.8 years. Results The multimodal model showed best performance for identifying CVRF (Table 1). The predicted CVRF allowed risk stratification for stroke (Figure 1). Table 1. ROC-AUC for detecting CVRF from T2 FLAIR Brain MRI, ECG and combination of modalities. Figure 1. Cumulative incidence of stroke stratified by number of CVRF predicted by ECG and Brain MRI deep learning models Conclusions Multimodal deep-learning of Brain MRI and ECG allows improved detection of CVRF and risk stratification for stroke. Our models could inform prevention strategies, through identification of CVRF for targeted reduction of stroke risk. Conflict of interest Julian Deseoe: nothing to disclose, Ezequiel de la Rosa: nothing to disclose, Bjoern Menze: nothing to disclose, Susanne Wegener: nothing to disclose Table 1 - belongs to Results Figure 1 - belongs to Conclusions