Identifying potential nonpulmonary vein triggers in persistent atrial fibrillation using digital twins and deep learning

Although pulmonary vein (PV) isolation (PVI) is an established atrial fibrillation (AF) treatment, its efficacy in persistent AF (PsAF) remains limited. Accordingly, post-PVI arrhythmia inducibility testing—mimicking non-PV ectopic activation—has often been employed during PsAF ablation, the value of non-inducibility as procedural endpoints or prognostic markers remains uncertain. Personalized atrial digital twins (DTs) can assess patient-specific arrhythmia inducibility; however, their clinical translation is hindered by the need for expert fibrosis characterization from MRI signals and substantial computational resources. To overcome these barriers, we leveraged deep learning (DL) to develop a practical framework for identifying potential non-PV trigger (NPVT) sites capable of inducing arrhythmia, without requiring patient-specific fibrosis thresholding or simulation. A comprehensive pacing-site set yielded significantly more patients’ DTs harboring potential NPVT sites than conventional catheter-based sites (85% vs. 54%). The DL model, InduceNet, trained on DT-derived results, accurately predicted potential NPVT sites (sensitivity 91%). This integrated DT–DL framework facilitates direct access, in clinical practice, to DT-derived insights into patient-specific potential NPVT sites, supporting more effective decision-making.