Creating Synthetic EEG Signals Using a Denoising Diffusion Probabilistic Model

Keywords

Loading...
Thumbnail Image

Issue Date

2023-02-23

Language

en

Document type

Journal Title

Journal ISSN

Volume Title

Publisher

Title

ISSN

Volume

Issue

Startpage

Endpage

DOI

Abstract

Electroencephalography (EEG) is often used in neuroscientific research to measure brain signals non-invasively. However, due to the limitations of EEG a lot of potentially interesting data is lost. Some researchers have attempted to solve this problem with data augmentation. The state-of-the-art of synthetic EEG signal generation uses variational autoencoders (VAEs) or generative adversarial networks (GANs). In the past few years a new kind of generative model has been introduced, namely denoising diffusion probabilistic models (diffusion models). In this work we aim explore the possible capacity of diffusion models to generate synthetic EEG signals. We train a diffusion model adapted for one-dimensional data on the EEGdenoiseNet dataset. To compare it to the state-ofthe- art we also train TimeVAE on the same dataset. We evaluate both models using the minimal Euclidean distance, mean Wasserstein distance, and FID-score. Additionally, we visually examine the generated samples and compare them with a PSD-plot. For two evaluation metrics our diffusion model performs similarly as the state-of-the-art, and for one evaluation metric it significantly outperforms the state-of-the-art. Therefore, we conclude that our diffusion model is capable of generating synthetic EEG signals and more diverse samples than the state-of-the-art while maintaining similar quality.

Description

Citation

Faculty

Faculteit der Sociale Wetenschappen