Biosignal Oversampling Using Wasserstein Generative Adversarial Network
Munawara Saiyara Munia, Mehrdad Nourani, Sammy Houari
Abstract
Oversampling plays a vital role in improving the minority-class classification accuracy for imbalanced biomedical datasets. In this work, we propose a single-channel biosignal data generation method by exploiting the advancements in well-established image-based Generative Adversarial Networks (GANs). We have implemented a Wasserstein GAN (WGAN) to generate synthetic electrocardiogram (ECG) signal, due to their stability in training as well as correlation of the loss function with the generated image quality. We first trained the WGAN with fixed-dimensional images of the signal and generated synthetic data with similar characteristics. Two evaluation methods were then used for evaluating the efficiency of the proposed technique in generating synthetic ECG data. We used Frechet Inception Distance score for measuring synthetic image quality. We then performed a binary classification of normal and abnormal (Anterior Myocardial Infarction) ECG using Support Vector Machine to verify the performance of the proposed method as an oversampling technique.