A Residual Attention-Based Algorithm for EEG Artifact Removal

This is the authors' abstract. We don't add key points for this paper.

Background: EEG recordings are readily contaminated by EOG and EMG artifacts, which overlap with neural activity in both waveform morphology and spectral content. Existing artifact-removal approaches may be limited in simultaneously modelling local waveform structures, long-range temporal dependencies, and spectral preservation, while deterministic predictions provide limited information about reconstruction reliability. Method: WeproposeDual-Denoise-A-Net, a single-input EEG artifact-removal framework consisting of a two-layer GRU temporal encoder, a residual dynamic-weight branch, a Transformer branch, learned local–global feature fusion, and a convolutional decoder. The network operates entirely on time-domain sequences, while spectral preservation is encouraged through a composite objective combining temporal RRMSE and normalized PSD consistency. MC dropout is further incorporated to characterize prediction variability and evaluate reconstruction reliability. Results: On a unified 13,200-sample test set containing EOG, EMG, and mixed-artifact conditions, Dual-Denoise-A-Net achieved a CC of 0.8919±0.0011 , temporal RRMSE of 0.4245±0.0028,spectral RRMSE of 0.3785 ± 0.0040, and MSE of 0.0640 ± 0.0008 over five independent training seeds. Ablation experiments demonstrated that the GRU and Transformer components provided the largest complementary contributions, while residual dynamic weighting and learned feature fusion yielded additional consistent improvements. The MC-dropout analysis identified an interpolated empirical reliability boundary of −6.05 dB, at which the mean coverage was 81.63% and the accepted-sample failure rate was 10.00%.Among the evaluated neural baselines, EDGeNet provided the strongest comparison. Relative to EDGeNet, Dual-Denoise-A-Net increased CC by 1.41% and reduced temporal RRMSE, spectral RRMSE, and MSE by 6.34%, 8.56%, and 10.99%, respectively. Conclusions: Dual-Denoise-A-Net effectively integrates sequential temporal encoding, adaptive local waveform modelling, and long-range dependency modelling to improve EEG artifact removal while preserving temporal and spectral characteristics. The validation-calibrated MC-dropout analysis further provides a dataset-specific empirical indication of reconstruction reliability under increasingly severe contamination.