Po-04-208 clinical application of artificial intelligence in prediction of recurrence in atrial fibrillation patients with catheter ablation

The atrial fibrillation (AF) recurrence is still a challenging issue after catheter ablation. It is important to identify which patients are intended to have a post-ablation recurrence before ablation for decision-making. We aimed to build AI models in PVCT images and clinical variables. We also combined deep learning in PVCT images and machine learning of clinical parameters by nonparametric statistical method or I-score to build an ensemble model to predict the post-ablation one-year AF recurrence. We retrospectively analyzed 518 PAF patients who have undergone catheter ablation of PAF (Table 1). PVCT geometric slices (2-3mm for each slice, 20-200 slices for each patient, a total of 36943 images of slices in 518 AF patients) were used in the deep learning process for the prediction of AF trigger origin as model 1. In 1 year follow-up, there was no AF recurrence in 388 (74.9%) patients (Group 1) and AF recurrence in 130 (25.1%) patients (Group 2). The deep learning process (EfficientNet) of cardiac CT geometric slices was applied for the prediction of AF recurrence. A total of 43 clinical variables were obtained before catheter ablation, including baseline characteristics, personal habits, medical histories, a stroke risk score of AF, laboratory values of blood tests, transthoracic echocardiographic parameters, left atrial volume calculated by 3D geometric reconstruction, and medications. Using different machine learning methods, the predictive performance of clinical variables using all variables or selected features by Nonparametric Statistics ranged from an AUC of 0.691 to 0.760. the predictive performance of clinical variables by I-Score was between AUC of 0.702 to 0.729 (Table 2). The best predictive performance was an AUC of 0.729 using I-Score rank 3 parameters (Table 3) and the CatBoost machine learning method. We combined the features of PVCT images and clinical variables. The ensemble model was able to promote the predictive performance up to an AUC of 0.722 (accuracy of 75.9%, sensitivity of 50.0%, and specificity of 83.7%) in the test set. Deep-learning AI using pre-ablation cardiac CT and clinical variables can be applied in the prediction of recurrence in AF patients receiving catheter ablation. Application of this model may identify patients with post-ablation AF recurrence.