BWGAN-GP: An EEG Data Generation Method for Class Imbalance Problem in RSVP Tasks.

Area of Science:

• Neuroscience
• Machine Learning
• Signal Processing

Background:

• Class imbalance problems (CIPs) significantly hinder classifier performance in deep learning, particularly in rapid serial visual presentation (RSVP) tasks.
• Existing methods struggle to effectively manage the imbalanced data inherent in RSVP classification.

Purpose of the Study:

• To propose a novel data augmentation method, balanced Wasserstein generative adversarial network with gradient penalty (BWGAN-GP), for generating RSVP minority class data.
• To evaluate the effectiveness of BWGAN-GP in alleviating CIPs and improving classification performance in RSVP tasks.

Main Methods:

• Developed BWGAN-GP, integrating generative adversarial network (GAN) with an autoencoder initialization strategy for accurate class-conditioning.
• Employed BWGAN-GP to generate artificial electroencephalography (EEG) data for the minority class using features learned from majority classes.
• Evaluated classification performance using RSVP datasets from nine subjects, comparing BWGAN-GP with other methods.

Main Results:

• The BWGAN-GP achieved an average Area Under the Curve (AUC) of 94.43% on EEGNet, a 3.7% increase over original data.
• Generated EEG data reduced the required original data for acceptable classification performance to 60%.
• BWGAN-GP demonstrated optimal performance when class data was balanced, effectively alleviating CIPs.

Conclusions:

• BWGAN-GP is a highly effective data augmentation technique for addressing class imbalance in RSVP tasks.
• The findings suggest potential for reducing calibration time in brain-computer interface (BCI) paradigms similar to RSVP through artificial EEG generation.
• This approach offers a promising solution for improving deep learning model performance on imbalanced datasets in neuroscience applications.