EEG Data Synthesis with WGAN-GP
Model Type: Conditional WGAN-GP
Dataset: EEG Motor Movement/Imagery Dataset (PhysioNet)
Model Summary
This model implements a Wasserstein GAN with Gradient Penalty (WGAN-GP) for generating subject-conditioned synthetic EEG signals based on the PhysioNet EEG Motor Movement/Imagery dataset.
It generates realistic EEG segments that mimic real recordings in both time and frequency domains.
Applications:
- EEG data augmentation
- Privacy-preserving EEG synthesis
- Adversarial and spoofing research in BCI security
Model Architecture
Generator (G)
- Input: Latent vector z ∈ ℝ¹²⁸ + subject embedding (128-dim)
- Layers:
- Fully connected → reshape to (64, 120)
- Dilated ConvTranspose1D stack → (64, 480)
- Gaussian noise injection
- Activation:
tanh
- Output: EEG segment (64 channels × 480 samples)
Discriminator / Critic (D)
- Input: EEG + embedded label
- Conv2D + LeakyReLU + global pooling + linear critic output
- Regularization: Dropout + drift penalty + gradient penalty
Training Configuration
| Parameter | Value |
|---|---|
| Optimizer | Adam (β₁=0.0, β₂=0.9) |
| Learning Rate (G/D) | 1e-4 / 5e-5 |
| Gradient Penalty λ | 10 |
| Drift Regularization | 1e-3 × D(real)² |
| n_critic | 3 |
| Epochs | 300 |
| Mixed Precision | Enabled (GradScaler) |
Dataset
Source: PhysioNet EEG Motor Movement/Imagery Dataset
Subjects: 109
Channels: 64
Sampling Rate: 160 Hz
Segment Length: 480 samples (~3 seconds)
Tasks: Motor imagery (fists, feet) + baseline (eyes open/closed)
Preprocessing:
- Downsampled to smallest subject class
- Per-channel normalization to [-1, 1]
- Integer-encoded subject IDs (0–108)
Training Behavior
Training remained stable across all epochs — no mode collapse or exploding gradients.
| Metric | Mean | Std |
|---|---|---|
| D(real) | −0.43 | ±0.06 |
| D(fake) | 0.38 | ±0.04 |
| GP term | 0.96 | ±0.08 |
| G loss | −0.41 | ±0.05 |
Evaluation & Results
Visual & Spectral Fidelity
- Synthetic EEG signals preserve oscillatory structure and amplitude range (±100 µV normalized).
- Channel correlations match real EEG patterns.
- Spectral energy distribution consistent with 1–40 Hz range.
Quantitative Similarity
| Metric | Mean | Std | Interpretation |
|---|---|---|---|
| MSE | 2.279 | 2.806 | Low reconstruction error |
| MAE | 0.870 | 0.487 | Normalized amplitude deviation |
| Correlation | −0.0014 | 0.075 | Low linear correlation due to stochastic nature |
| MMD | 0.0129 | — | Good distribution alignment |
| Fréchet PCA (32-D) | 40092.3 | — | Baseline EEG-FID |
| Covariance Similarity | 0.673 | ±0.182 | Preserves inter-channel dependencies |
Bandpower Fidelity
| Band | Δ Mean | Δ Std | Interpretation |
|---|---|---|---|
| Delta (1–4 Hz) | 0.391 | 0.675 | Slight underfit |
| Theta (4–8 Hz) | 0.082 | 0.131 | Close match |
| Alpha (8–13 Hz) | 0.032 | 0.038 | Excellent fidelity |
| Beta (13–30 Hz) | 0.012 | 0.016 | Excellent fidelity |
| Gamma (30–40 Hz) | 0.0007 | 0.0014 | Negligible difference |
Discussion
This WGAN-GP model effectively learns to reproduce subject-specific EEG morphology and maintains stable training across complex, high-dimensional signals.
It provides a promising basis for privacy-preserving EEG synthesis and data augmentation in BCI research.
References
- Arjovsky, M., Chintala, S., & Bottou, L. (2017). Wasserstein GAN. arXiv:1701.07875
- Gulrajani, I. et al. (2017). Improved Training of Wasserstein GANs. NIPS
- Lawhern, V. J. et al. (2018). EEGNet: A Compact CNN for EEG-based BCI. J. Neural Eng., 15(5), 056013
- Goldberger, A. L. et al. (2000). PhysioBank, PhysioToolkit, and PhysioNet. Circulation, 101(23), e215–e220
- Downloads last month
- 6