EEG data augmentation using Wasserstein GAN

Electroencephalogram (EEG) presents a challenge during the classification task using machine learning and deep learning techniques due to the lack or to the low size of available datasets for each specific neurological disorder. Therefore, the use of data augmentation which consists of adding batches of data with patterns quite similar to the training data can offer an interesting solution. Inspired by the successes of the generative adversarial network (GAN) and specifically the Wasserstein GAN (WGAN) version, we propose a deep learning WGAN to generate artificial EEG with features related to each addressed pathogen to approximate the original training dataset. The experimental results demonstrate that using the artificial EEG data generated by our Wasserstein GAN significantly improves the accuracies of the classification models. The implementation was performed using a real dataset dealing with the Autism pathology which is provided by the King Abdulaziz University. Thus, we achieved great results using the presented data augmentation technique applied to the above-mentioned dataset.

I. Introduction

Data augmentation is one of the promising ways to solve the problem of data scarcity and especially during the training phase of the dataset classification based on deep learning techniques. This approach consists of generating additional samples to make the available datasets much bigger by applying significant processing on them in order to improve the classification accuracy and to increase its degree of stability.