Electroencephalography artifact removal based on an autoencoder deep network

The electroencephalography (EEG) signal acquisition process is inevitably affected by a variety of physiological noise signals, including electrooculogram (EOG), electromyography (EMG). The traditional methods of removing EOG and EMG rely heavily on the subjective experience and prior knowledge of the user. However, the ambiguity of artificial judgments can lead to erroneous and misleading interpretations that are insufficient for qualitative analysis. This inaccurate denoising may affect the true information of the signals in the time domain and spectral domain, leading to a decline in the accuracy of the BCI system. In recent years, a variety of EEG denoising methods based on deep learning have been proposed, but their denoising performance needs to be further improved. In this paper, we design a novel autoencoder (AE) neural network to remove artifacts in EEG. The network includes an encoder and a decoder module. The encoder contains five convolutional layers with increasing feature dimension as depth increases, which are responsible for detecting and suppressing artifacts. The decoder contains five deconvolution layers, whose feature dimension decreases gradually, and is used for EEG reconstruction after denoising. The experimental results on semi-synthetic EEG datasets demonstrate that the proposed algorithm outperforms the four benchmark models.