Disentangling inter-subject variations: Automatic localization of ventricular tachycardia origin from 12-lead electrocardiograms

An automatic, real-time localization of ventricular tachycardia (VT) can improve the efficiency and efficacy of interventional therapies. Because the exit site of VT gives rise to its QRS morphology on electrocardiograms (ECG), it has been shown feasible to predict VT exits from 12-lead ECGs. However, existing work have reported limited resolution and accuracy due to a critical challenge: the significant inter-subject heterogeneity in ECG data. In this paper, we present a method to explicitly separate and represent the factors of variation in data throughout a deep network using denoising autoencoder with contrastive regularization. We demonstrate the performance of this method on an ECG dataset collected from 39 patients and 1012 distinct sites of ventricular origins. An improvement in the accuracy of localizing the origin of activation is obtained in comparison to a traditional approach that uses prescribed QRS features for prediction, as well as the use of a standard autoencoder network without separating the factors of variations in ECG data.