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

A new generative model called REST-GAN can synthesize realistic resting-state EEG signals and extract transferable features without manual engineering, according to research posted on arXiv [1]. The framework uses adversarial training with a self-supervised reconstruction objective to serve as both a signal generator and an unsupervised feature extractor [1]. The model was trained solely on raw time-domain signals, without explicit frequency-domain or sensor-topographic supervision, yet the generated time series reproduced key temporal, spectral, and connectivity properties of real EEG [1]. In band-power feature space, generated samples achieved precision and recall of 0.91 and 0.67 in the eyes-open condition, and 0.87 and 0.65 in the eyes-closed condition [1]. Group-average spectral coherence matrices showed mean absolute differences from real data of approximately 0.01 to 0.03 across frequency bands [1]. The representations learned by the model's critic transferred to independent resting-state demographic classification tasks, outperforming models trained directly on raw EEG and showing competitive performance relative to a recent EEG foundation model, while requiring substantially less training data and computational resources [1]. The approach may support more data-efficient EEG analysis while reducing reliance on manual feature engineering [1]. REST-GAN enters a growing field of generative architectures applied to EEG. A separate framework called EEGDM, proposed by a different group, uses a latent diffusion model for EEG representation learning, conditioning the denoising process on channel augmentations to capture global inter-channel spatial synchrony and phase relationships [3]. Another diffusion-based approach, also termed EEGDM, integrates a structured state-space model for diffusion pretraining to capture temporal dynamics and was reported to be approximately 19 times more lightweight than existing EEG foundation models while outperforming them on the Temple University EEG Event Corpus [8]. Earlier work introduced the Generative EEG Transformer, a model pre-trained on motor imagery and alpha wave datasets that generated continuous, context-sensitive neural signals while maintaining the frequency spectrum of the training data [5]. Resting-state EEG provides a non-invasive view of spontaneous brain activity, but extracting meaningful patterns is often limited by scarce high-quality data and reliance on manually engineered features [1]. Generative adversarial networks can synthesize neural signals and learn transferable representations directly from raw data, a dual capability that remains underexplored in EEG research [1]. The implementation code for REST-GAN is available on GitHub [1].