EEGWave: a Denoising Diffusion Probabilistic Approach for EEG Signal Generation

EasyChair Preprint 10275

Abstract

The importance of brain-computer interface (BCI) systems in modern-day healthcare and robotics is indisputable. BCIs are commonly based on signals recorded by electroencephalography (EEG) due to the ease of use and relatively low cost of the measurement technology. Evoked potentials (EP) are well-measurable with EEG and can be utilized to control BCIs. The processing of these signals is a very complex task, because of the low signal-to-noise ratio, the inter-subject, and inter-measurement variability. In recent years, deep learning (DL) methods have proven to be powerful algorithms for signal processing and decoding. However, a large amount of data with good quality and variety is needed in order to train DL models properly and use them as generally as possible. Finding such a publicly available data set or even creating one is a resource-intensive task. Denoising Diffusion Probabilistic Models (DDPM) got more attention over generative adversarial networks (GAN) only recently and in image processing tasks they achieved state-of-the-art results. In this work, We propose a DDPM, called EEGWave, for brain signal generation and show that DDPMs are promising options for augmenting EEG data sets. We present our results on a data set which contains EPs to present the performance of the model in the signal synthesis task.

Keyphrases: Brain Computer Interface (BCI), Diffusion probabilistic model, EEGWave, Electroenchephalography (EEG), Generative Modelling, deep learning, signal generation, signal synthesis

@booklet{EasyChair:10275,
  author    = {Szabolcs Torma and Luca Szegletes},
  title     = {EEGWave: a Denoising Diffusion Probabilistic Approach for EEG Signal Generation},
  howpublished = {EasyChair Preprint 10275},
  year      = {EasyChair, 2023}}