Augmenting Electrocardiogram Datasets Using Generative Adversarial Networks

Abstract

In this work, we explored and proposed various techniques for the synthesis of realistic Electrocardiograms (ECG). In the last two decades, several methods were developed to automatically classify the ECG signals. Deep learning has evolved rapidly and now demonstrates the state-of-art performance in many fields, including the ECG domain. In this thesis, we first summarize the existing deep learning models in the context of ECG classification, providing the details about their architecture, hyper-parameters, performance results, training and testing datasets used, the number of classification classes, etc. Many of these models are trained and tested on very small datasets. However, deep learning models have to be trained on large datasets for accurate classification. In this work, we propose a few generative models to augment the existing ECG datasets and demonstrate its effectiveness under various network configurations.