A High-Precision Machine Learning Algorithm to Classify Left and Right Outflow Tract Ventricular Tachycardia

• DOI: 10.3389/fphys.2021.641066
• Publication Date: 2021-02-25T06:25:36Z
• Authors: Jianwei Zheng; Guohua Fu; Islam Abudayyeh; Magdi Yacoub; Anthony Chang; William W. Feaster; Louis Ehwerhemuepha; Hesham El-Askary; Xianfeng Du; Bin He; Mingjun Feng; Yibo Yu; Binhao Wang; Jing Liu; Hai Yao; Huimin Chu; Cyril Rakovski
• Keywords: Artificial intelligence; Artificial Intelligence and Robotics; Physiology; electrocardiography; Cardiology; Catheter Ablation of Cardiac Arrhythmias; Receiver operating characteristic; Ablation; Ventricular outflow tract; 03 medical and health sciences; 0302 clinical medicine; Tachycardia; Diagnosis; catheter ablation; Health Sciences; Atrial Fibrillation; Machine learning; QP1-981; Internal medicine; outflow tract ventricular tachycardia; Supraventricular Tachycardia; Health Information Technology; Ventricular tachycardia; Molecular Mechanisms of Cardiac Arrhythmias; Computer science; Algorithm; classification; Ventricular Tachycardia; Medicine; Catheter ablation; Cardiology and Cardiovascular Medicine; Numerical Analysis and Scientific Computing; Other Computer Sciences; artificial intelligence algorithm

IntroductionMultiple algorithms based on 12-lead ECG measurements have been proposed to identify the right ventricular outflow tract (RVOT) and left ventricular outflow tract (LVOT) locations from which ventricular tachycardia (VT) and frequent premature ventricular complex (PVC) originate. However, a clinical-grade machine learning algorithm that automatically analyzes characteristics of 12-lead ECGs and predicts RVOT or LVOT origins of VT and PVC is not currently available. The effective ablation sites of RVOT and LVOT, confirmed by a successful ablation procedure, provide evidence to create RVOT and LVOT labels for the machine learning model.MethodsWe randomly sampled training, validation, and testing data sets from 420 patients who underwent successful catheter ablation (CA) to treat VT or PVC, containing 340 (81%), 38 (9%), and 42 (10%) patients, respectively. We iteratively trained a machine learning algorithm supplied with 1,600,800 features extracted via our proprietary algorithm from 12-lead ECGs of the patients in the training cohort. The area under the curve (AUC) of the receiver operating characteristic curve was calculated from the internal validation data set to choose an optimal discretization cutoff threshold.ResultsThe proposed approach attained the following performance: accuracy (ACC) of 97.62 (87.44–99.99), weighted F1-score of 98.46 (90–100), AUC of 98.99 (96.89–100), sensitivity (SE) of 96.97 (82.54–99.89), and specificity (SP) of 100 (62.97–100).ConclusionsThe proposed multistage diagnostic scheme attained clinical-grade precision of prediction for LVOT and RVOT locations of VT origin with fewer applicability restrictions than prior studies.