EEG Signal Denoising Using Beta-Variational Autoencoder

Electroencephalography (EEG) signals are a valuable source of information for investigating brain activity and different types of brain-related disease diagnoses. However, EEG signals are often contaminated by various kinds of noises/artifacts. Several methods have been proposed for EEG reconstruction/denoising to facilitate signal analysis, but such algorithms often fail when the EEG contains extreme artifacts. This paper presents a novel method for reconstructing EEG signals using a variant of the variational autoencoder (VAE) called beta-VAE. Through extensive evaluation of our model on the DEAP dataset, we show that the β-VAE architecture learns a compressed representation of the EEG signal in an unsupervised manner, and the reconstructed signal contains less artifact. We compare our proposed method with different baselines and state-of-the-art techniques for EEG signal denoising, demonstrating significantly reduced reconstruction error under artificially induced noise. The results suggest that our approach has great potential for improving the analysis and understanding of EEG signals in clinical and research settings.