Detection of Ventricular Late Potentials in Electrocardiograms Using Machine Learning

Abstract

The first signs of myocardial electrical instability reflect the depletion of regulatory systems at the cellular level. These changes at the first stage may not manifest themselves clinically as functional and anatomical changes. The development of methods for detecting early signs of cardiac abnormalities makes it possible to prevent life threatening pathological processes. Such a task includes eliciting a violation of the electrical homogeneity of the myocardium based on the registration of ventricular late potentials (VLP) by high resolution electrocardiography. The goal of this paper is to evaluate the performance of the Simson method, widely accepted as the standard method for detecting VLP and to compare it to other methods based on time-frequency analysis. The simulation of VLP with different signal-to-noise ratio conducted in this study allows us to generate a variety of VLP of different shapes, which correspond to the states of norm and pathology. Comparison of Simson method, acknowledged as the standard method for VLP detection, time-frequency and wavelet analysis as well as combinations of their features is performed in order to determine whether VLP presence in ECG. Machine learning approach is used to find the efficiency of each set of features. A new method for VLP detection based on wavelet analysis is proposed as a suitable replacement for the Simson method. Using the proposed features made it possible to separate healthy and sick patients with an overall accuracy of 99%, wherein 98% of the cases with VLP were correctly identified outperforming the Simson method.