JMIR Publications

ABSTRACT

Background:

The high prevalence of atrial fibrillation requires the development of reliable methods to determine the probability of success in patients who undergo ablation procedures. Currently, common practise is to rely on medical scores, such as ATLAS, which often require information that is not always available to all patients, thus limiting their practical use. Such limitations could be addressed with adaptable and dynamic artificial intelligence models, provided that they are interpretable for use in clinical practise.

Objective:

This study aims to investigate the potential use of Bayesian networks as clinical decision tools to predict the relapse of atrial fibrillation, prior to a percutaneous PVI procedure.

Methods:

An explainable artificial intelligence model based on Bayesian networks was developed to calculate conditional probabilities of risk based on the clinical condition of each patient. The model was tested for 5, 6 and 7 predictors of atrial fibrillation relapse and validated using four different sampling techniques.

Results:

Model performance in predicting atrial fibrillation relapse was evaluated using as predictors age, sex, smoking, pre-ablation AF type, and obstructive sleep apnea (5 predictors), to which were added body mass index (6) or left atrial volume and epicardial fat (7 predictors). Predictive performances were measured by AUC-ROC at 0.661 (CI 0.603-0.718), 0.703 (CI 0.652-0.753), and 0.752 (CI 0.701-0.800), respectively.

Conclusions:

The model shows acceptable diagnostic accuracy even in the absence of some predictive features, thus providing a valid clinical tool to estimate the risk of relapse of atrial fibrillation with easily available clinical variables. Furthermore, this model can accommodate new risk factors and dynamically learn from new data, hence improving performance as new patient data are added to the model.