Machine learning-based predictive model for atrial arrhythmia following transcatheter atrial septal defect closure.

Background: Atrial septal defects (ASDs) are frequently closed percutaneously. Despite successful closure, many patients still develop atrial arrhythmias. There is inconsistent data on the risk factors associated with these atrial arrhythmias. As such, we aimed to develop a machine learning (ML) model predicting atrial arrhythmias following transcatheter ASD closure in adolescents and adults. Methods: Patients with secundum-type ASDs undergoing transcatheter closure between 2008 and 2024 at a single center were retrospectively analyzed. Patients with prior atrial arrhythmias were excluded. A deep neural network, adapted via transfer learning from a large external ECG dataset, was used to extract predictive features from preprocedural 12-lead ECGs. These features were combined with clinical, demographic, biochemical and hemodynamic variables in ensemble survival models. Model performance was assessed using the integrated Brier scores and the area under the receiver operating curve (AUC). Results: A total of 148 adult patients (median 44.4 years [30.6-57.8], 105 females [70.9 %]) were eligible for included. There were a total of 1055 person-years of follow-up (median follow-up 7.3 [3.1-11.3]), during which 28 patients (18.9 %) developed atrial arrhythmias. The final ensemble ML model incorporating ECG-derived features demonstrated strong predictive performance (integrated Brier score 0.044, mean AUC 0.823). Subgroup and sensitivity analyses confirmed the robustness of the model across various patient profiles. Conclusions: We developed a novel ML-based risk model using a transfer learning approach to predict atrial arrhythmias after transcatheter ASD closure. Further research and external validation are needed to refine the proposed risk stratification prior to clinical implementation.