Premature Ventricular Conduction Detection and Localization From the ECG Using a Neural Network

The identification and localization of premature ventricle contractions (PVC) can be a lengthy procedure. In each treatment, a great deal of time is spent in the precise localization of the origin of the cardiac arrhythmia. This work investigates the acceleration of the PVC detection process, using standard 12-lead ECG data as input and localizes PVC on the right ventricular (RV) or left ventricular (LV). The proposed neural network (NN) is a shallow NN which consists of only one hidden layer with multiple hidden units. Three data sets consisting of a total of 328 resting ECG samples are used to train and evaluate the NN. Multiple iteration tests with different training sets have been done to identify the most promising configuration. The training cohorts differ in the distribution of data with PVC (cohort 1 ratio 1:1, cohort 2 ratio 25:4; NO PVC: PVC). High sensitivity and specificity values have been reached in NNs with uniformly distributed training data providing a sufficient performance, which might be comparable to an expert.