End-to-end EEG artifact removal method via nested generative adversarial network

As physiological artifacts commonly overlap with EEG signals in both time and frequency domains, developing an effective end-to-end EEG artifact removal method is essential for a brain-computer interface (BCI) system. An end-to-end artifact removal method based on nested generative adversarial network (GAN) is proposed, to recover the EEG signals from artifact-contaminated ones. The nested GAN consists of two components: an inner GAN operating in time-frequency domain and an outer GAN functioning in time domain. A light-weighted complex-valued restormer, designed in time-frequency domain, is employed as the generator to reconstruct the denoised EEG signal. Two metric discriminators in the inner GAN and two multi-resolution discriminators in the outer GAN are used, and gradient balance is used to address the partial learning issue during training. The performance of the nested GAN has been evaluated in the realistic EEG dataset and semi-synthetic dataset. Compared to the benchmark methods, the proposed one achieved best average performance evaluation metrics, including mean square error (MSE) = 0.098, Pearson correlation coefficient (PCC) = 0.892, relative root MSE (RRMSE) = 0.065, the percentage reduction of time domain artifacts (ηtemporal) = 71.6%, and the percentage reduction of frequency domain artifacts (ηspectral) = 76.9%. The performance of artifact removal also showed robustness across a wide range of signal-to-noise ratio (SNR) levels.The superior performance of the proposed end-to-end artifact removal method is expected to contribute to the advancement of BCI system development.