UGECGAD: EVALUATING THE EFFICACY OF UNSUPERVISED GAN ARCHITECTURES FOR ECG ANOMALY DETECTION

Anomaly detection in Electrocardiogram (ECG) signals is critical for reliable diagnosis and continuous patient monitoring. Although traditional approaches are incapable of describing the rich nonlinear structure in ECG data, they are ineffective in detecting these abnormalities. Although Generative Adversarial Networks (GANs) has recently attracted increasing interest for ECG anomaly detection, research available currently does not provide a systematic evaluation of their advantages and disadvantages for practical situations. This paper addresses this gap by critically reviewing unsupervised GAN-based approaches, evaluating their ability to reconstruct normal ECG signals and accurately detect deviations. We introduce a rigorous empirical comparison of different GAN architectures and their adversarial variants, highlighting key challenges in their implementation. Our results show that the ECG-Adversarial Autoencoder (ECG-AAE) is superior to the other GAN-based approaches in terms of training anomaly, and provides the best performance in anomaly detection. This study contributes new insights into the robustness of ECG-AAE, establishing its potential for precise and reliable ECG anomaly detection in practical healthcare applications