PREDICTION OF ATRIAL FIBRILLATION RECURRENCE AFTER REPEAT ABLATION USING MACHINE LEARNING

Abstract

BackgroundPulmonary vein isolation is a widely used method to treat Atrial fibrillation (AF). However, many patients experience AF recurrence, leading to repeated ablation requirements. In this study, we sought to identify whether left atrial (LA) geometry can identify patients at risk of AF recurrence after second ablation.ObjectiveTo predict AF recurrence after repeat ablation using machine learning on clinical covariates and LA geometric features calculated from CT scans.MethodsThis retrospective single-center study included 102 patients who experienced AF recurrence after a single radiofrequency ablation procedure and underwent a second ablation. 38 patients had AF recurrence within 12 months of second ablation. Cardiac CT images taken prior to second ablation and clinical covariates were collected for each patient. From the segmented 3D LA geometry, 11 geometric features were calculated which along with 35 clinical covariates were analyzed for model development. Performance was analyzed using k-fold cross-validation on 76 patients and relevant features and optimized model were selected to maximize the cross-validation performance (Fig. 1).ResultsThe selected geometric features include sphericity and compactness (how much the shape of LA resembles that of a sphere) while the selected clinical features include the CHA2DS2-VASc score, history of any prior cardioversion and ARB medication. On the held-out test set of 26 patients, the radial-basis support vector machine (RBF-SVM) model achieved the best performance and obtained an accuracy of 0.67±0.08, specificity of 0.67±0.07, sensitivity of 0.67±0.15 and NPV of 0.77±0.09 (Fig. 2).ConclusionView Large Image Figure ViewerDownload Hi-res image Download (PPT)Receiver operating characteristic (ROC) curve for the SVM-RBF classifier model on the test patients. BackgroundPulmonary vein isolation is a widely used method to treat Atrial fibrillation (AF). However, many patients experience AF recurrence, leading to repeated ablation requirements. In this study, we sought to identify whether left atrial (LA) geometry can identify patients at risk of AF recurrence after second ablation. Pulmonary vein isolation is a widely used method to treat Atrial fibrillation (AF). However, many patients experience AF recurrence, leading to repeated ablation requirements. In this study, we sought to identify whether left atrial (LA) geometry can identify patients at risk of AF recurrence after second ablation. ObjectiveTo predict AF recurrence after repeat ablation using machine learning on clinical covariates and LA geometric features calculated from CT scans. To predict AF recurrence after repeat ablation using machine learning on clinical covariates and LA geometric features calculated from CT scans. MethodsThis retrospective single-center study included 102 patients who experienced AF recurrence after a single radiofrequency ablation procedure and underwent a second ablation. 38 patients had AF recurrence within 12 months of second ablation. Cardiac CT images taken prior to second ablation and clinical covariates were collected for each patient. From the segmented 3D LA geometry, 11 geometric features were calculated which along with 35 clinical covariates were analyzed for model development. Performance was analyzed using k-fold cross-validation on 76 patients and relevant features and optimized model were selected to maximize the cross-validation performance (Fig. 1). This retrospective single-center study included 102 patients who experienced AF recurrence after a single radiofrequency ablation procedure and underwent a second ablation. 38 patients had AF recurrence within 12 months of second ablation. Cardiac CT images taken prior to second ablation and clinical covariates were collected for each patient. From the segmented 3D LA geometry, 11 geometric features were calculated which along with 35 clinical covariates were analyzed for model development. Performance was analyzed using k-fold cross-validation on 76 patients and relevant features and optimized model were selected to maximize the cross-validation performance (Fig. 1). ResultsThe selected geometric features include sphericity and compactness (how much the shape of LA resembles that of a sphere) while the selected clinical features include the CHA2DS2-VASc score, history of any prior cardioversion and ARB medication. On the held-out test set of 26 patients, the radial-basis support vector machine (RBF-SVM) model achieved the best performance and obtained an accuracy of 0.67±0.08, specificity of 0.67±0.07, sensitivity of 0.67±0.15 and NPV of 0.77±0.09 (Fig. 2). The selected geometric features include sphericity and compactness (how much the shape of LA resembles that of a sphere) while the selected clinical features include the CHA2DS2-VASc score, history of any prior cardioversion and ARB medication. On the held-out test set of 26 patients, the radial-basis support vector machine (RBF-SVM) model achieved the best performance and obtained an accuracy of 0.67±0.08, specificity of 0.67±0.07, sensitivity of 0.67±0.15 and NPV of 0.77±0.09 (Fig. 2). ConclusionReceiver operating characteristic (ROC) curve for the SVM-RBF classifier model on the test patients. Receiver operating characteristic (ROC) curve for the SVM-RBF classifier model on the test patients.