A Deep Generative Model for Resting-State EEG Synthesis and Transferable Representation Learning

REST-GAN introduces a generative adversarial network framework for synthesizing resting-state EEG signals while learning transferable representations without manual feature engineering. The model demonstrates strong performance in reproducing key EEG properties and outperforms direct raw-signal approaches on demographic classification tasks, offering a computationally efficient alternative to existing EEG analysis methods.

REST-GAN represents a meaningful advance in neuroimaging analysis by applying generative modeling to an underexplored domain. The framework addresses a persistent challenge in EEG research: limited high-quality datasets and dependence on manually engineered features that often fail to capture the full complexity of neural signals. By combining adversarial training with self-supervised learning, the model learns meaningful representations directly from raw time-domain data without explicit instruction on frequency or spatial properties.

The technical achievement is notable because the generated EEG signals reproduce temporal, spectral, and connectivity characteristics of authentic recordings across multiple dimensions—band power features showed precision/recall scores of 0.87-0.91, while spectral coherence matrices remained within 0.01-0.03 of real data. This synthetic data quality opens pathways for augmenting scarce datasets, a persistent bottleneck in clinical and research applications where collecting large-scale EEG cohorts remains expensive and logistically challenging.

From a practical standpoint, REST-GAN's efficiency gains matter significantly. The learned representations transferred effectively to independent classification tasks, outperforming models trained directly on raw signals and matching recent foundation models while demanding substantially fewer computational resources and training data. This democratizes access to sophisticated EEG analysis for organizations lacking extensive infrastructure or datasets.

The broader implication extends beyond neuroscience. Successfully applying GANs to EEG synthesis demonstrates how generative approaches can solve domain-specific data scarcity problems in biomedical AI. Future work should explore whether similar architectures transfer to other biosignal modalities and clinical applications where high-quality training data remains limited.

• →REST-GAN successfully synthesizes realistic resting-state EEG signals while learning transferable representations without manual feature engineering
• →Generated EEG samples reproduced key temporal, spectral, and connectivity properties with high fidelity across eyes-open and eyes-closed conditions
• →The model's learned representations outperformed raw-signal baselines and matched recent foundation models with dramatically reduced computational requirements
• →This approach enables more data-efficient EEG analysis by addressing the persistent challenge of limited high-quality datasets in neuroscience research
• →The framework demonstrates GANs' potential as dual-purpose tools for both signal synthesis and unsupervised feature extraction in biomedical domains