Adversarial Denoising of EEG Signals: A Comparative Analysis of Standard GAN and WGAN-GP Approaches

Imad Eddine TIBERMACINE^1*Samuele Russo²Francesco Citeroni¹Giuseppe Mancini¹Abdelaziz Rabehi³Amal H. Alharbi^4*El-Sayed M. El-kenawy^5,6Christian Napoli^7,8,9

• ¹Department of Computer, Automatic and Management Engineering, Faculty of Information Engineering, Computer Science and Statistics, Sapienza University of Rome, Rome, Lazio, Italy
• ²Department of Psychology, Sapienza University of Rome, Rome, Sicily, Italy
• ³Ziane Achour University of Djelfa, Djelfa, Algeria
• ⁴Department of Computer Sciences, College of Computer and Information Sciences, Princess Nourah bint Abdulrahman University, Riyadh, Saudi Arabia
• ⁵School of ICT, Faculty of Engineering, Design and Information & Communications Technology (EDICT), Bahrain Polytechnic, Isa Town, Bahrain
• ⁶Applied Science Private University, Amman, Amman, Jordan
• ⁷Sapienza University of Rome, Rome, Italy
• ⁸National Research Council (CNR), Roma, Lazio, Italy
• ⁹Department of Computational Intelligence, Czestochowa University of Technology, Czestochowa, Poland

Introduction: Electroencephalography signals frequently contain substantial noise and interference, which can obscure clinically and scientifically relevant features. Traditional denoising approaches, such as linear filtering or wavelet thresholding, often struggle with nonlinear or time-varying artifacts. In response, the present study explores a Generative Adversarial Network framework to enhance EEG signal quality, focusing on two variants: a conventional GAN model and a Wasserstein GAN with Gradient Penalty. Methods: Data were obtained from two distinct EEG datasets: a "healthy" set of 64-channel recordings collected during various motor/imagery tasks, and an "unhealthy" set of 18-channel recordings from individuals with orthopedic impairments. Both datasets underwent comprehensive preprocessing, including band-pass filtering (8-30 Hz), channel standardization, and artifact 1 Tibermacine et al. trimming. The training stage involved adversarial learning, in which a generator sought to reconstruct clean EEG signals while a discriminator attempted to distinguish between real and generated signals. The model evaluation was conducted using quantitative metrics such as signal-to-noise ratio, peak signal-to-noise ratio, correlation coefficient, mutual information, and dynamic time warping distance. Results: Experimental findings indicate that adversarial learning substantially improves EEG signal fidelity across multiple quantitative metrics. Specifically, WGAN-GP achieved an SNR of up to 14.47 dB (compared to 12.37 dB for the standard GAN) and exhibited greater training stability, as evidenced by consistently lower RRMSE values. In contrast, the conventional GAN model excelled in preserving finer signal details, reflected in a PSNR of 19.28 dB and a correlation coefficient exceeding 0.90 in several recordings. Both adversarial frameworks outperformed classical wavelet-based thresholding and linear filtering methods, demonstrating superior adaptability to nonlinear distortions and dynamic interference patterns in EEG time-series data. Discussion: By systematically comparing standard GAN and WGAN-GP architectures, this study highlights a practical trade-off between aggressive noise suppression and high-fidelity signal reconstruction. The demonstrated improvements in signal quality underscore the promise of adversarially trained models for applications ranging from basic neuroscience research to real-time BCI in clinical or consumer-grade settings. The results further suggest that GAN-based frameworks can be easily scaled to next-generation wireless networks and complex electrophysiological datasets, offering robust and dynamic solutions to long-standing challenges in EEG denoising.