High-Fidelity EEG Generation: Generative Adversarial Network Highlighting Time-Frequency-Spatial Features Regulated...

Area of Science:

• Neuroscience
• Machine Learning
• Biomedical Signal Processing

Background:

• Electroencephalogram (EEG) analysis relies heavily on machine learning.
• Limited availability of diverse EEG datasets necessitates effective data augmentation.

Purpose of the Study:

• Introduce HiFi-EEG-GAN, a novel Generative Adversarial Network framework for high-fidelity EEG data augmentation.
• Improve the fidelity and diversity of synthetic EEG data generation.

Main Methods:

• Developed a framework with a supervisor, generator, and discriminator for EEG generation.
• Employed Global Dynamics Supervision using Kullback-Leibler (KL) divergence.
• Utilized a composite architecture generator for replicating EEG characteristics.
• Incorporated a discriminator for microscopic detail regulation.

Main Results:

• HiFi-EEG-GAN demonstrated superior performance in data augmentation fidelity and diversity compared to state-of-the-art methods.
• Achieved notable metrics: r1NNC (0.88), FID (13.97), and MMD (0.09).
• Augmented EEG data improved classification accuracy by 3.2%-6.8% in ASD and 2.5%-8.2% in mental arithmetic tasks.

Conclusions:

• HiFi-EEG-GAN effectively generates high-fidelity synthetic EEG data.
• The framework significantly enhances machine learning model performance in downstream tasks.
• HiFi-EEG-GAN offers a reliable solution for EEG data augmentation challenges.