Deep-Learning Premature Contraction Localization in 12-lead ECG from Whole Signal Annotations

Abstract

Since common electrocardiography (ECG) diagnostics approaches are time-consuming and arrhythmia-type sensitive, deep-learning methods are state-of-the-art for their detection accuracy. However, premature ventricular contractions' (PVC) localization via common deep-learning approaches requires large training set, therefore Multiple Instance Learning (MIL) framework was applied, where model is trained from whole-signal annotations. Proposed MIL framework is based on 1D Convolutional Neural Network (CNN), with global max-pooling in the last layer. The detection of PVCs' positions was done by the peak detector with specified parameters - threshold, minimal distance and peak prominence. Our method was tested on database containing 1590 ECGs, including 672 signals with PVCs. Dice coefficient reaches 0.947. This simple deep-learning method for the localization of PVC achieves a promising performance while being trainable from the whole-signal annotations instead of positional labels.