Contents I. Introduction
Physicians tend to rely on the electrocardiogram, also known as the ECG or EKG, to diagnose heartbeat disorders [1]. This involves analyzing the shape and heart rate variability (HRV) obtained from the recorded electrocardiogram. An ECG is a non-invasive medical tool that records the heart's bioelectrical contractions. However, the ECG signal is known for its sensitive nature to various types of noise from diverse sources, which can cause damage to its features and affect its clinical value. These noises might complicate the task of doctors making precise clinical interpretations and potentially lead to insufficient diagnoses, which can drive them to use other forms of complicated diagnosis. Furthermore, it poses challenges for advanced artificial intelligence technologies specifically developed for ECG-based heart disease detection or classification. Various types of noise can cause distortion in the ECG signal, such as; baseline wander (BW), electrode motion artifacts (EM), power-line interference, and electromyography (EMG) noise. As each one of these noises damages significantly the ECG signal, in different forms, a filtering process is a fundamental step before any analysis performed on the signal. This process tends of removing unwanted noise or disturbances from a signal. Noise can be random or have a specific pattern, and it often distorts the quality of the signal. Thus, in the specific context of electrocardiograms (ECGs), noise elimination is a vital step in removing disturbances that interfere with these signals. To address this challenge, a variety of commonly employed techniques have been extensively studied and presented in academic literature to mitigate or eliminate these disturbances. However, because of the progress in computational computing, a widely discussed framework known as “Generative Adversarial Networks” (GANs) has emerged in the literature [2]. GANs comprise two specific models: a generator, which is trained to produce new data based on the patterns in the training data, and a discriminator, which is trained to differentiate between original samples and those generated by the generator. As the discriminator effectively continues to distinguish the generated data samples, it drives the generator to effectively provide additional new data samples that closely resemble the real ones, effectively deceiving the discriminator. Scientists have primarily implemented Generative Adversarial Networks (GANs) in image processing to produce lifelike examples, achieving uplifting results across various applications [3]. These include image-to-image translation, such as converting summer backgrounds to winter or day to night, and generating realistic photos of objects and scenes. Specifically, in image denoising applications [4], [5], they have delivered outstanding results. Furthermore, several types of adversarial models have been developed from the foundational GAN model, such as the Wasserstein Generative Adversarial Network (WGAN) [6], the Conditional Generative Adversarial Network (CGAN) [7], and the Cycle Generative Adversarial Network (cycleGAN) [8]. Several studies used the GANs for the denoising the ECG signals. The advantage of these techniques is that they can generalize to a wide range of noises, by learning the properties of the different noises, with just a single model.
