Towards synthesizing physiologically accurate EEG signals with generative adversarial networks: an exploration of hyperparameter impact on sine samples generation

Electroencephalography (EEG) is important in medical diagnostics and cognitive neuroscience for brain activity recording and understanding, but limited data availability restricts the application of machine learning methods for classification and prediction purposes. Therefore, our ultimate aim is to synthesize physiologically accurate EEG data using Generative Adversarial Networks (GANs). In this preliminary study, we first investigate how the choice of GAN hyperparameters influences the quality of simpler EEG-like generated signals. Thereto, we first created sinusoidal samples as target datasets with single (12 Hz), double (5, 12 Hz), or triple frequencies (5, 12, 28 Hz) and added Gaussian noise. We trained 2,592 models to generate fake sinusoidal samples and examined the effects of varying the hyperparameters batch size, learning rate of generator (LrG) and discriminator (LrD), and the number of epochs for vanilla GAN, Wasserstein GAN (WGAN), and WGAN with Gradient Penalty (WGAN-GP). We compared the distributions of relevant target and fake signal features, such as amplitude, phase, and signal-to-noise ratio, using the Kolmogorov-Smirnov test. Our results showed that WGAN-GP models had better training stability for sinusoidal sample generation compared to WGAN, and significantly outperformed vanilla GAN. Smaller batch sizes accelerated convergence at a fixed number of epochs. Matching LrG and LrD balanced networks during training. Increasing the number of epochs (up to 100) enhanced performance. We conclude that the WGAN-GP model provides the best quality signals, when hyperparameters, particularly learning rate, are well tuned. We therefore propose that effective combinations of hyperparameters enhance training stability and performance of GANs for EEG signal generation.