Predicting Atrial Fibrillation Relapse Using Bayesian Networks: Explainable AI Approach.

Area of Science:

• Cardiology
• Artificial Intelligence
• Medical Informatics

Background:

• Atrial fibrillation (AF) is a common arrhythmia with significant health impacts.
• Predicting AF recurrence after ablation remains challenging, limiting effective patient management.
• Existing risk models may not fully capture crucial clinical factors for AF relapse.

Purpose of the Study:

• To develop an explainable AI model using Bayesian networks for predicting AF relapse post-ablation.
• To evaluate the model's predictive performance using various clinical variables.
• To assess the model's adaptability and potential to enhance clinical decision-making.

Main Methods:

• Developed an explainable AI model based on Bayesian networks for AF relapse prediction.
• Utilized clinical data from 480 patients undergoing percutaneous pulmonary vein isolation (PVI).
• Evaluated model performance using AUC-ROC with 5, 6, and 7 predictors, including age, BMI, and epicardial fat.

Main Results:

• The Bayesian network model showed promising predictive performance, with AUC-ROC reaching 0.752 with 7 predictors.
• Model accuracy remained acceptable even with missing predictor data, demonstrating adaptability.
• The model effectively estimates AF relapse risk using readily available clinical variables.

Conclusions:

• An interpretable Bayesian network model can reliably predict AF relapse after PVI.
• The model's use of accessible clinical variables and adaptability makes it suitable for real-world application.
• This AI tool can aid clinicians in managing AF patients post-ablation.