A general dual-pathway network for EEG denoising

Figures

• 

  Figure 1. The coupled system model of scalp EEG contaminated by artifacts. The endogenous signals contain clean scalp EEG signals and physiological artifacts dominated by electrooculogram (EOG), electromyogram (EMG), and electrocardiographic (ECG). The endogenous signals antagonize each other and are co-transmitted through the skull to the scalp electrodes to be acquired. The exogenous signals, which contain physiological artifacts dominated by industrial frequency interference and motion artifacts, are transmitted via the scalp to the scalp electrodes, and together with the endogenous signals, constitute the observed scalp EEG. The exogenous signals comprised primarily of industrial frequency interference and motion artifacts are conveyed through the scalp to the scalp electrodes. These signals, in conjunction with the endogenous signals, collectively form the observable scalp EEG.

• 

  Figure 2. The dual-pathway denoising autoencoder (DPAE) structure diagram. The DPAE model first extracts multiscale abstract features by stacking network layers of varying scales. Subsequently, it simulates the coupling of multiple source signals through the fusion block. Following this, it employs residual connection structures to extract scale-independent generic encoding features of the target source signals. Finally, it reconstructs the target source signals based on the decoder. It is worth noting that to make the network structure look more intuitive, we ignore the joint representation layer. FN denotes a fully connected hidden neural layer, BN denotes a batch normalization layer, and the widths of the individual layers indicate the relative proportions of the layer neuron sizes. The encoder layers can be replaced by MLP, CNN, and RNN layers.

• 

  Figure 3. Typical examples in the EEGdenoiseNet dataset. The waveforms of clean scalp EEG, EOG, and EMG with a sampling rate of 256 Hz and a duration of 2 s are given here, respectively.

• 

  Figure 4. Example waveforms of the denoising results of EMG (left) and EOG (right). It is evident that the time-domain waveforms of the denoised EEG signal closely resemble those of the clean scalp EEG signal, irrespective of the presence of contamination from EOG or EMG.

• 

  Figure 5. Example PSD plots of denoising result of EMG. We presented power spectral density (PSD) plots for scalp EEG signals contaminated by EMG across various frequency bands both before and after noise reduction. The delineated frequency intervals for each band are as follows: delta [1–4 Hz], theta [4–8 Hz], alpha [8–13 Hz], beta [13–30 Hz], and gamma [30–80 Hz]. The efficacy in eliminating artifacts in the higher-energy EEG bands, specifically delta, theta, and alpha, exhibited greater strength and displayed more consistent curves. Conversely, in the beta and gamma bands, the ability to remove artifacts was less pronounced due to the comparatively lower energy levels inherent in the clean EEG signals within these bands.

• 

  Figure 6. Example PSD plots of the denoising results of EOG. We present power spectral density (PSD) plots for scalp EEG signals contaminated by EOG across various frequency bands both before and after noise reduction. The delineated frequency intervals for each band are as follows: delta [1–4 Hz], theta [4–8 Hz], alpha [8–13 Hz], beta [13–30 Hz], and gamma [30–80 Hz]. The frequency range of EOG typically falls within 0–13 Hz, which is relatively narrow. As depicted in the figure, EOG primarily introduces interference in the delta, theta, and alpha frequency bands. The results demonstrate that our denoising method consistently exhibits exceptional performance across all frequency bands.

• 

  Figure 7. Evaluation metrics curves of the denoising results from different artifacts of the EEGdenoiseNet dataset. The figure presents the denoising results of different neural network models based on time-domain and frequency-domain root relative mean square error (RRMSE) and Pearson correlation coefficients. These results are obtained for scalp EEG signals contaminated by varying signal-to-noise ratios of EOG and EMG.

• 

  Figure 8. Ablation study on loss comparison. We conducted the ablation study on two separate instances: one involved the removal of the fusion block module, and the other involved the removal of the residual connection. Both instances utilized the same random seed and the same training set (utilizing the EMG dataset). The observed metric is the training losses.

• 

  Figure 9. Denoising results of myoelectric artifacts on multichannel scalp electroencephalogram at different signal-to-noise ratios. The figure presents the denoising results of different DPAE models based on time-domain and frequency-domain root relative mean square error (RRMSE) and Pearson correlation coefficients. These results are obtained for multichannel scalp EEG signals contaminated by varying signal-to-noise ratios of EMG.