Modeling EEG Data Distribution With a Wasserstein Generative Adversarial Network to Predict RSVP Events

Area of Science:

• Computational neuroscience
• Machine learning for biosignal processing
• Deep learning for electroencephalography

Background:

• Acquiring electroencephalography (EEG) data is challenging due to complex setups and participant discomfort, limiting the training of robust deep learning models.
• Limited availability of high-quality EEG datasets hinders advancements in brain-computer interfaces and neurological disorder analysis.
• Computational generation of synthetic EEG data is crucial to augment existing datasets and improve model performance.

Purpose of the Study:

• To develop a novel Wasserstein Generative Adversarial Network with gradient penalty (WGAN-GP) for synthesizing realistic EEG data.
• To address challenges in simulating time-series EEG data, including frequency artifacts and training instability.
• To create a class-conditioned WGAN-GP variant for event-related classification tasks.

Main Methods:

• Implementation of a Wasserstein Generative Adversarial Network with gradient penalty (WGAN-GP) for EEG data synthesis.
• Extension to a class-conditioned WGAN-GP incorporating a classification branch for event detection.
• Training and validation of the proposed networks using data from a rapid serial visual presentation (RSVP) experiment, generating both single and multi-channel EEG signals.

Main Results:

• The proposed WGAN-GP successfully synthesized realistic EEG data, mimicking signals from RSVP experiments.
• The class-conditioned WGAN-GP demonstrated improved event-classification performance compared to EEGNet for RSVP target events.
• The generated EEG samples were validated for their resemblance to authentic signals.

Conclusions:

• The developed WGAN-GP effectively synthesizes high-fidelity EEG data, addressing limitations of data scarcity.
• Class-conditioned WGAN-GP offers a promising approach for enhancing event-related classification in EEG analysis.
• This method holds potential for improving deep learning model training and performance in various EEG applications.