An Innovative Approach for Long ECG Synthesis with Wasserstein GAN Model

Deep neural networks (DNNs) have set new standards in identifying and classifying irregular patterns in ECG (electrocardiogram) signals, surpassing previous methods. Despite the easy access and affordability of ECG sensors, a critical bottleneck remains the limited availability of reliable data for complex heart rhythms like second and third-degree atrioventricular block, ventricular tachycardia, and supraventricular tachycardia. This shortage has been a significant obstacle to improving DNN algorithms. Recent studies have turned to Generative Adversarial Networks (GANs) to create synthetic ECG data, enhancing the diversity of training datasets. However, much of this research has only managed to produce basic ECG components, missing the intricate details found in real patient data that includes multiple heartbeats. Our research has taken a groundbreaking approach by converting ECG signals into a two-dimensional format, allowing us to utilize advanced GAN models originally developed for image processing. This method has enabled us to generate extended, realistic ECG sequences closely mimicking those from actual patients. We have tested and refined our model using two databases, Physionet and Chapman, and have successfully produced 10-second ECG sequences showcasing a variety of heart rhythms previously unachieved in other studies. Our innovative technique not only surpasses existing methods in generating high-quality, realistic ECG data but also sets a new benchmark in ECG synthesis.