Multi-phase ECG dynamic features for detecting myocardial ischemia and identifying its etiology using deterministic learning

Abstract

Timely detecting myocardial ischemia and identifying its etiology is of great value to the making of therapeutic regimen. Electrocardiogram (ECG) is a first-line screening tool for myocardial ischemia, but few based-ECG studies into the identification of its etiology have been reported. In this study, we propose a multi-phase ECG dynamic feature-based method for detecting myocardial ischemia and identifying its etiology using deterministic learning and the self-adaptive variational mode decomposition (VMD). Firstly, a self-adaptive approach is designed to produce the VMD components describing the ECG depolarization and repolarization phases, respectively. Then, dynamic modeling of ECG signal and multi-phase VMD components is carried out by deterministic learning, and multiple dynamic features under different phases of ECG activities are extracted from the learned dynamics. Subsequently, we design a hybrid feature reduction scheme to obtain low-dimensional but highly informative features while preserving the physical significance of features. Finally, classifiers are trained and evaluated on 297 ECG records from 297 patients, including 118 patients with non-ischemia chest pain, 41 patients with coronary slow flow (CSF), and 138 patients with coronary artery stenosis (COAS). The results show that logistic regression with a few dynamic features achieves promising performance in detecting myocardial ischemia (CSF and COAS), with the area under the receiver operating characteristic (AUROC) of 0.9126. Especially, it can effectively identify the etiology of myocardial ischemia, with an overall AUROC of 0.9238. The results show that our method provides a powerful tool for clinicians to analyze the etiology of ischemia.