Deep Convolutional Autoencoder for EEG Noise Filtering

Electroencephalography (EEG) signals may be severely affected by noise originated from various sources due to their low amplitude nature, specially if they are collected from scalp sensors. Several methods have been proposed for EEG denoising in order to facilitate diagnosis and communication in brain-computer interfaces, but such algorithms often have high complexity. This work presents a denoising approach based on deep learning using a deep convolutional autoencoder, which should reduce the effort of projecting denoising filters. Experiments were performed using two types of noise, originated from eye blink and from jaw clenching. Performance was evaluated with peak signal-to-noise ratio (PSNR) and the results showed that all confidence intervals for the proposed approach were superior to those obtained by the baseline bandpass traditional filtering method. Best average PSNR results for eye blink were obtained for Cz channels with $(20.3\pm 2.6)\mathrm{d}\mathrm{B}$ versus $(14.3\pm 2.4)\mathrm{d}\mathrm{B}$. For jaw clenching, best average PSNR results were obtained for Fz channels with $(21.7\pm 3.1)\mathrm{d}\mathrm{B}$ versus $(13.9\pm 2.6)\mathrm{d}\mathrm{B}$. The proposed approach seems to open a promising scope of research for noise filtering in EEG.