MINDGAN: EEG-Based Motor Imagery Neural Decoding with GAN-Augmented Hybrid Deep Learning

MINDGAN is a unified deep learning framework designed to address two major challenges in motor imagery (MI) EEG classification:

• Data scarcity in EEG recordings
• Inter-subject variability that limits generalisation

The framework integrates:

• A conditional DCGAN (cDCGAN) for class‑conditioned EEG data augmentation
• A hybrid CNN–Transformer classifier for local–global spatiotemporal feature extraction
• A three‑phase curriculum training schedule for stable optimisation

---

Key Features

• EEGNet‑inspired depthwise separable convolutions
• Six‑block pre‑LayerNorm Transformer encoder
• Wasserstein loss with gradient penalty
• Spectral normalisation + class‑conditional batch norm
• EMA‑based synthetic sample quality filtering
• Class‑balanced replay buffer
• Ablation study showing LSTM layers degrade performance
• Real‑time prototype using Emotiv EPOC X

---

Performance

Evaluated on BCI Competition IV datasets:

• 2A (4‑class MI): 81.17% mean accuracy
• 2B (binary MI): 86.87% mean accuracy
• GAN‑generated EEG achieves r = 0.9923 spectral fidelity
• Augmentation benefit strongly depends on per‑class data availability

---

Included in this Repository

• Full implementation of MINDGAN
• Training scripts for Datasets 2A and 2B
• Preprocessing pipeline
• Synthetic EEG generation code
• Results, figures, logs, and trained models

---