Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Synthesizing EEG Signals from Event-Related Potential Paradigms with Conditional Diffusion Models

Guido Klein, Pierre Guetschel, Gianluigi Silvestri, Michael Tangermann

TL;DR

A novel approach to conditional diffusion models that utilizes classifier-free guidance to directly generate subject-, session, and class-specific EEG data that resembles real data for each subject, session, and class is introduced.

Abstract

Data scarcity in the brain-computer interface field can be alleviated through the use of generative models, specifically diffusion models. While diffusion models have previously been successfully applied to electroencephalogram (EEG) data, existing models lack flexibility w.r.t.~sampling or require alternative representations of the EEG data. To overcome these limitations, we introduce a novel approach to conditional diffusion models that utilizes classifier-free guidance to directly generate subject-, session-, and class-specific EEG data. In addition to commonly used metrics, domain-specific metrics are employed to evaluate the specificity of the generated samples. The results indicate that the proposed model can generate EEG data that resembles real data for each subject, session, and class.

Synthesizing EEG Signals from Event-Related Potential Paradigms with Conditional Diffusion Models

TL;DR

A novel approach to conditional diffusion models that utilizes classifier-free guidance to directly generate subject-, session, and class-specific EEG data that resembles real data for each subject, session, and class is introduced.

Abstract

Data scarcity in the brain-computer interface field can be alleviated through the use of generative models, specifically diffusion models. While diffusion models have previously been successfully applied to electroencephalogram (EEG) data, existing models lack flexibility w.r.t.~sampling or require alternative representations of the EEG data. To overcome these limitations, we introduce a novel approach to conditional diffusion models that utilizes classifier-free guidance to directly generate subject-, session-, and class-specific EEG data. In addition to commonly used metrics, domain-specific metrics are employed to evaluate the specificity of the generated samples. The results indicate that the proposed model can generate EEG data that resembles real data for each subject, session, and class.
Paper Structure (3 figures, 1 table)

This paper contains 3 figures, 1 table.

Figures (3)

  • Figure 1: Figures (\ref{['fig:timecourse']}) and (\ref{['fig:SCM']}) provide two different comparisons between real and generated data. The figures are based on the averages of EEG data (522 target and 2875 non-target examples), which are sampled for the combination of subject and session combination that resulted in the worst ABA metric (subject 52, session 1). The model with the highest ABA metric over all subjects and sessions (i.e. 600k training steps) was used to generate the samples.
  • Figure 2: Scores of every metric measured on the generated data of the model at multiple training steps. The "real: within-session" baseline is computed by taking the ABA metric on the real data within a session of the same participant. The "real: between-session" baseline displays the variance between two sessions of the same participant on that particular metric. The band shows the 95 % confidence interval.
  • Figure 3: The PLD between the P300 peak latency in channel "O1" in the evoked real and generated data.