Automated Identification and Localization of Premature Ventricle Contractions in Standard 12-Lead ECGs

Abstract

It can take up to twelve hours to identify and precisely localize the origin of a premature ventricle contraction (PVC).This work is investigating a neural network (NN) as an automated alternative to a human expert for detecting and locating the arrhythmogenic zone—with the goal of accelerating the PVC detection process. The proposed shallow neural network contains one hidden layer with multiple hidden units. Three data sets consisting of a total of 328 samples of 12 lead resting ECGs were used to train as well as to evaluate the NN. After performing several iteration tests with different training sets, the most promising configuration was established. The first cohort consisted of a ratio of 1:1, the second cohort of a ratio of 25:4 (NO PVC:PVC).The study has resulted in high sensitivity and specificity values in NN’s performance given uniformly distributed training data. The proposed NN was shown to perform at a level comparable to that of a human expert.