EEG-GAN: A Generative EEG Augmentation Toolkit for Enhancing Neural Classification

Williams, Chad; Weinhardt, Daniel; Hewson, Joshua; Płomecka, Martyna Beata; Langer, Nicolas; Musslick, Sebastian · 2025 · Crossref

DOI: 10.1101/2025.06.23.661164

archive: archived pipeline: cataloged verified

Get this paper ↗ (DOI — opens at the source; we link to it, we don't host it)

Summary

This paper introduces EEG-GAN, a novel generative adversarial network (GAN) architecture designed to augment electroencephalography (EEG) data for improved neural classification. The study addresses the challenge of limited sample sizes in EEG research, which often hinders the training of robust machine learning classifiers. By generating realistic synthetic EEG samples, the authors aim to enhance classification performance, particularly in low-data regimes. The motivation stems from preliminary evidence suggesting that GAN-based augmentation can improve brain decoding, but this work seeks to validate the approach across diverse experimental contexts and compare it against traditional augmentation methods. The researchers developed EEG-GAN by modifying standard GAN architectures to address training instability and mode collapse. Key innovations include replacing the traditional objective function with a Wasserstein distance, incorporating a gradient penalty, and implementing transformer models for both the generator and discriminator to capture temporal dependencies. Crucially, the architecture embeds an autoencoder (AE) to decouple feature extraction from generation; the GAN operates on the AE’s embedded latent space rather than raw time-series data, resulting in faster and more stable training. The model is conditional, allowing it to generate data specific to experimental conditions. The study evaluated EEG-GAN using four open-source datasets covering distinct cognitive tasks: reinforcement learning, anti-saccade, face perception, and visual search. These datasets varied in participant counts, EEG systems, and processing pipelines. The authors assessed the quality of generated data by comparing synthetic samples to empirical data across trial-level waveforms, grand-averaged event-related potentials (ERPs), and frequency transforms. They also benchmarked EEG-GAN against a variational autoencoder (VAE) and five traditional transformation techniques (oversampling, Gaussian noise, time reverse, polarity reverse, and smoothing). Classification performance was tested using five classifiers (neural networks, SVMs, logistic regression, random forests, and K-Nearest Neighbors) across varying sample sizes (5 to 100 participants). Results demonstrated that EEG-GAN significantly enhanced classification performance compared to training on empirical data alone, with improvements of up to 16% in two-way classification accuracy. These gains were consistent across all four datasets and varied among classifiers. The augmentation was most effective for smaller sample sizes (30 or fewer), where it improved 70% of classification analyses and impaired only 4%. In contrast, traditional transformation methods and the VAE benchmark showed less consistent or inferior performance. The synthetic data generated by EEG-GAN closely matched empirical distributions, as evidenced by high correlations in grand-averaged difference waveforms. The significance of this work lies in providing a robust, generalizable toolkit for EEG data augmentation. By demonstrating consistent improvements across diverse paradigms and classifiers, EEG-GAN offers a viable solution for enhancing neural decoding in studies with limited data. This approach can facilitate more reliable machine learning applications in neuroscience, particularly where collecting large-scale EEG datasets is impractical. The open-source availability of the toolkit further supports its adoption and validation by the broader research community.

Provenance

The full processing record for this entry. Every stage of this paper's journey through the pipeline is logged — what ran, with which tool and model, how many attempts it took, and when it last completed.

Summary generated by qwen3.6-27b-nvidia on 2026-08-10; verification: verified.

Topics

Ranked by relevance to this paper. Hover a topic for its definition.

• drowsiness detection algorithms