Onset of ventricular tachycardia (VT) is clinically significant, including as a trigger to defibrillator implants. In this paper, we propose a reliable technique to detect such onset using convolutional neural networks (CNNs). The proposed CNN adds convolution and pooling layers below the input layer and above the hidden and output layers of usual neural network (NN). Such layers would learn suitable linear features from training data, while eliminating the need to extract the traditionally used adhoc features. Employing such subject-specific features, we reported the performance of the proposed classifier using Creighton University ventricular tachyarrhythmia database (CUVT). In particular, we achieved mean (± standard deviation) performance of 95.6 (± 00.6) using subject-specific evaluation scheme over 100 random independent iterations. © 2016 CCAL.