ERP-WGAN: A data augmentation method for EEG single-trial detection.

Area of Science:

• Neuroscience
• Computer Science
• Signal Processing

Background:

• Brain-computer interfaces (BCIs) using electroencephalography (EEG) show significant promise but are hindered by limited data.
• The oddball paradigm commonly used in EEG experiments leads to imbalanced datasets, negatively impacting BCI system performance.
• Scarcity of EEG samples is a major bottleneck for developing robust BCI systems.

Purpose of the Study:

• To propose a novel data augmentation method for EEG signals using generative adversarial networks (GANs).
• To improve the performance of EEG signal classification and BCI systems by addressing data scarcity and class imbalance.
• To enhance the quality and diversity of synthetic EEG data.

Main Methods:

• Developed a Wasserstein Generative Adversarial Network (WGAN) framework, termed ERP-WGAN, specifically tailored for EEG data.
• Incorporated resting noise, smoothing, and random amplitude techniques within the WGAN to improve synthetic data quality and address model collapse.
• Evaluated the quality of generated data based on verisimilitude, diversity, and accuracy.

Main Results:

• The ERP-WGAN framework significantly improved the performance of both subject-specific and general EEG classifiers.
• Accuracy of general classifiers trained with fewer than 5 features improved by 20-25% compared to existing methods.
• Demonstrated that ERP-WGAN can reduce real subject data requirements by at least 73%, significantly cutting acquisition costs and test cycle times.

Conclusions:

• ERP-WGAN effectively addresses the challenge of limited EEG data and class imbalance in BCI research.
• The proposed method generates high-quality, diverse synthetic EEG data, leading to substantial performance gains in classification tasks.
• ERP-WGAN offers a cost-effective solution for BCI development by reducing the need for extensive real-world data collection.