Postdoctoral Fellow University of Pittsburgh Pittsburgh, Pennsylvania, United States
Background: Atrial septal defects (ASDs) are frequently treated using transcatheter closure. In the current era, procedural success has become the norm, yet, some patients still develop late atrial arrhythmias. Moreover, the risk factors for atrial arrhythmias after closure remain unclear. We aimed to develop a predictive machine learning (ML) model that could accurately predict the occurrence of atrial arrhythmias following transcatheter closure of ASDs.
METHODS AND RESULTS: We retrospectively analyzed data from patients with secundum-type ASDs who underwent transcatheter closure between 2008 and 2024 at University Hospitals Leuven. Patients with a history of previous atrial arrhythmias were excluded in initial model training. We used pre-trained deep neural network (DNN), which was adapted through transfer learning from a large, external electrocardiogram (ECG) dataset. In addition to the ECG-derived features, clinical, demographic, biochemical, and hemodynamic variables were integrated into ensemble survival models (CoxNet, Random Survival Forests and Gradient Boosting Models). Feature selection used a permutation-based method with shadow features and binomial modeling to identify and rank predictors by importance. Model performance was evaluated using the integrated Brier score and the area under the receiver operating characteristic curve (AUC) in a time-dependent manner and at final follow-up. A total of 148 adult patients (median age 44.4 years, range 30.6-57.8) were eligible for inclusion in the study, with 105 (70.9%) of them being female. The patients were followed for a total of 1055 person-years (median follow-up 7.3 years, range 3.1-11.3), during which 28 patients (18.9%) developed atrial arrhythmias. The final ensemble ML model, which incorporated both ECG-derived features and biochemical data, demonstrated robust predictive performance, with an integrated Brier score of 0.044 and a mean AUC of 0.823. Subgroup and sensitivity analyses further confirmed that the model's performance was consistent across various patient profiles, including different age groups, sex, and comorbidity statuses.
Conclusion: A novel machine learning-based risk model for predicting atrial arrhythmias after transcatheter ASD closure is both feasible and effective. By utilizing a transfer learning approach to extract predictive features from 12-lead ECGs, we developed a model that demonstrated strong performance metrics, effectively addressing potential sample size limitations often encountered in machine learning applications for congenital heart disease. However, further research, including external validation across multiple centers and larger patient cohorts, is necessary to refine this risk stratification model before clinical implementation.
