A DenseNet-based abnormal ventricular potentials onset delineation: a feasibility study

Abnormal ventricular potentials (AVPs) are fractionated and complex electrograms (EGMs), typically associated with slow conduction areas in the myocardium. As such, in ventricular tachycardia (VT), their identification supports the localization of the myocardial areas sustaining the arrhythmia in the electroanatomic mapping. Their latency can be related to their spatial localization, suggesting the entrance or the exit of a reentry pathway. In this work, we explored a novel approach to delineate the onset of the pathological deflections in AVPs. For this preliminary study, an expert cardiologist was asked to mark the onset of such pathological deflections in the identified AVPs in EGMs obtained from five patients affected by post-ischemic VT. On this basis, we trained a DenseNet-based convolutional neural network to automatically detect the AVP onsets, by using a time-frequency representation of the signal. The results of the model predictions have been compared to the ground truth in a regression analysis, yielding a root mean square error of 26 ms and a correlation coefficient of 0.6 on average. Notably, the identification of both onset and duration of abnormalities in EGMs may support the process of estimation of local conduction velocity and direction in arrhythmogenic regions.