EOG Artifact Removal from EEG Using a Gated WGAN with Multi-Scale and Multi-Domain Feature Extraction

Background: Electroencephalogram (EEG) signals provide important information for brain-computer interfaces (BCIs) and the diagnosis of neurological disorders but are susceptible to electrooculographic (EOG) artifacts. Traditional methods often rely on multichannel data or relatively long windows, limiting online applicability. Deep learning enables single-channel end-to-end artifact removal, but existing methods may not fully exploit features in the time and frequency domains and may unnecessarily modify relatively clean EEG. Methods: We developed a Wasserstein generative adversarial network (WGAN) integrating reconstruction loss, multi-domain and multi-scale feature extraction, and a gating unit. Reconstruction loss improved waveform reconstruction, yielding a reconstruction-loss-assisted artifact-removal WGAN (rarWGAN). Parallel branches with different kernel sizes extracted multi-scale features, while a multi-domain module combined temporal features with spectral features extracted using complex-valued convolutions. Integrating both modules into rarWGAN produced MDS-rarWGAN. Based on discriminator scores, gating routed inputs classified as contaminated to the generator and retained the others unchanged. Results: On EEGdenoiseNet, MDS-rarWGAN yielded relative root mean square errors (RRMSEs) of 0.189 ± 0.028 and 0.183 ± 0.027 in the time and frequency domains, respectively, with a correlation coefficient of 0.955 ± 0.013. Compared with rarWGAN, RRMSEs decreased by 23.2% and 32.5%, respectively, and the correlation coefficient increased by 1.2%. The gating unit yielded false positive and false negative rates of 1.47% and 0.12%, respectively. For EEG containing artifacts, no statistically significant differences in time domain RRMSE were detected with and without gating (all p > 0.05). On BCI Competition IV 2a, reconstruction metrics measured agreement with a reference derived from independent component analysis (ICA), as artifact-free EEG ground truth was unavailable. Using the same EEGNet classifier, EEG processed by gated-MDS-rarWGAN achieved the highest mean motor imagery classification accuracy among the evaluated deep learning methods. Conclusions: The method suppresses EOG artifacts, reduces unnecessary processing of signals without evident artifacts, and supports motor imagery decoding, offering a feasible preprocessing approach.