Evaluating the Performance of Generative Adversarial Networks for EEG Signal Synthesis (MSc Thesis)

Overview

This MSc thesis project investigates the effectiveness of various Generative Adversarial Network (GAN) architectures for synthesizing realistic EEG (Electroencephalogram) data. The primary goal was to address the critical challenge of data scarcity in neurophysiological research, where the high cost and complexity of data collection limit the performance of machine learning models. This study systematically compares multiple GANs including Standard GAN, DualGAN, DCGAN, and WGAN-GP to generate synthetic EEG signals corresponding to two distinct mental states: "Concentrating" and "Relaxed." The research validates the potential of GANs as a powerful data augmentation tool for enhancing brain-computer interfaces and mental state classification tasks.

The Problem

The scarcity of high-quality, labeled EEG data is a significant bottleneck in neuroscience and machine learning. Acquiring large datasets is resource-intensive and often impractical, leading to models that overfit and fail to generalize well. While traditional data augmentation methods exist, they often fail to capture the intricate temporal and spatial dependencies inherent in complex EEG signals. This project explores GANs as a sophisticated solution, capable of learning the underlying distribution of real EEG data to generate novel, high-fidelity synthetic samples that can effectively expand training datasets.

Approach & Technical Details

The project was structured around a comprehensive comparative analysis of different GAN models, from baseline architectures to state-of-the-art implementations.

• Dataset and Preprocessing: Public EEG dataset from Bird et al. (2018), containing recordings of "Concentrating" and "Relaxed" mental states. Data was preprocessed for normalization, temporal alignment, and channel selection (TP9, AF7, AF8, TP10, Right AUX).


• GAN Architectures Investigated:
  • Standard GAN – baseline model for foundational performance.
  • Deep Convolutional GAN (DCGAN) – uses convolutional layers for spatial features.
  • DualGAN – two generator-discriminator pairs for independent mental state distributions.
  • WGAN-GP – advanced model to prevent mode collapse and improve signal fidelity.


• Evaluation Strategy: Assessed using Pearson Correlation, KL Divergence, Spectral Similarity, PSNR, and Fréchet Inception Distance (FID).

Key Technologies: Python | PyTorch | Generative Adversarial Networks (GANs) | EEG Signal Processing

Results & Outcomes

Heatmap showing similarity between generated EEG signals and real EEG signals.

• WGAN-GP as Top Performer: Achieved the highest combined performance score (0.6623) with superior stability and high-fidelity signal generation.


• Strong Performance from DualGAN: Excels in statistical alignment metrics like KL Divergence and FID.


• Successful Synthesis of Realistic Data: Demonstrated that advanced GANs can generate high-quality synthetic EEG signals.