EEG-GANet: Simulating Electroencephalogram Data to Address Sample Imbalance in P300 Speller

Objective: The P300 speller is a brain-computer interface (BCI) system that enables character selection using event-related potentials (ERPs). However, the inherent imbalance between target and non-target stimuli biases deep learning models. Reducing non-target samples to mitigate this issue risks underutilizing valuable data. Methods: We propose EEG-GANet, a generative adversarial network leveraging domain transformation to augment target samples and mitigate class imbalance in P300 speller datasets. Additionally, we develop EEG-DBNet-V2, a compact dual-branch network that efficiently extracts temporal and spectral features from electroencephalography (EEG) signals, serving both as a classification model and as the discriminator within EEGGANet. Results: Extensive experiments with ten-fold cross-validation on three public datasets demonstrated EEGDBNet- V2’s superior classification accuracy compared to state-of-the-art models, achieving this with significantly fewer parameters. Integrating EEG-GANet’s augmented data via fine-tuning further improved classification performance. Conclusion: EEG-GANet effectively addresses sample imbalance by generating physiologically accurate augmented data, substantially enhancing EEG-DBNet-V2’s classification capability. Significance: Our study introduces a targeted GAN-based EEG data augmentation framework, enabling balanced model training without discarding valuable non-target samples. The integration of EEG-GANet and EEG-DBNet-V2 offers a lightweight yet robust solution, advancing the practical deployment of EEG-based BCIs under data-constrained conditions. The source code is publicly available at https://github.com/xicheng105/EEG-GANet.