Optimal lead selection for evaluation ventricular premature beats using machine learning approach

Abstract

Repolarization heterogeneity (RH) has been shown to increase with ventricular premature beats (VPBs). Moreover, several differences between left ventricle (Lv) and right ventricle (Rv), such as fibrillation threshold and anatomic properties have been presented. Nevertheless, few results exist regarding the influence of the origin site of VPBs on modulation of ventricular RH, as well as the optimal electrode location to assess the origin of VPBs. We studied electrocardiographic indices as a function of the coupling interval and the site of VPBs, in an isolated rabbit heart preparation (n=18) using ECG multi-lead (5 rows×8 columns) system. In both ventricles, results have shown significant increases in ventricular depolarization duration. Also, we have observed that when the VPBs were applied to the Lv, a significant decrease of the total repolarization duration was detected, while in the Rv premature stimulation we have not found significant changes of total repolarization duration. Also, we compared twenty machine learning classification techniques with the aim to find the optimal electrode placement (row4–column4 to Lv stimulation and row5–column3 to Rv stimulation) and interpret the site of origin of VPBs. It was observed that the Random Forest classifier has shown the best performance among all the techniques studied. Finally, we found differences in the overall duration of repolarization associated to transmural RH.