Generative Adversarial Networks-Based Data Augmentation for Brain-Computer Interface

Area of Science:

• Neuroscience
• Machine Learning
• Biomedical Engineering

Background:

• Brain-computer interface (BCI) classifier performance relies heavily on training data quality and quantity.
• Real-world BCI applications face performance degradation due to diverted user attention.
• Acquiring extensive EEG data in diverse conditions is impractical.

Purpose of the Study:

• To propose a framework using deep convolutional generative adversarial networks (DCGANs) for artificial EEG data generation.
• To augment training datasets for improved BCI classifier robustness, particularly under attention-diverted conditions.
• To evaluate the effectiveness of DCGANs-based augmentation in enhancing BCI performance.

Main Methods:

• Designed a motor task experiment with focused and diverted attention conditions.
• Utilized an end-to-end deep convolutional neural network for classification.
• Employed DCGANs to generate artificial EEG data for augmenting training sets.
• Conducted leave-one-subject-out cross-validation for performance assessment.

Main Results:

• Baseline accuracies were 73.04% (diverted attention) and 80.09% (focused attention).
• DCGANs augmentation improved accuracy by 7.32% for diverted attention and 5.45% for focused attention.
• On BCI Competition III dataset IVa, the method increased accuracy by 3.57% for motor imagery tasks.

Conclusions:

• DCGANs-based EEG augmentation significantly enhances BCI classifier performance.
• The proposed framework is particularly beneficial for real-life BCI applications with potential attention diversion.
• Artificial data generation offers a time- and cost-efficient solution for improving BCI robustness.