Electroencephalography Data Denoising with Deep Neural Networks

Abstract

Electroencephalography (EEG) is frequently utilized in investigations including neuroscience, neural engineering, and biomedical engineering, because of its benefits, such as non-invasiveness, high temporal resolution, and relatively low cost, Nevertheless, the raw EEG data are frequently corrupted by physiological artifacts and many noises, such as cardiac artifacts, myogenic artifacts, and ocular artifacts. Artifacts can have a negative impact on an EEG signal and can be reduced or eliminated using existing deep learning-based EEG signal denoising techniques. However, when the acquired EEG data has significant artifacts, they typically experience performance degradation. In order to address this problem, we introduced the 2x3R-CNN and compared it to four deep learning-based methods: fully connected neural network (FCNN), basic convolution network, and deep reinforcement learning (simple CNN), complex convolution network (complex CNN), and recurrent neural network (RNN) to remove muscle and ocular artifacts from EEG signal. The experimental results demonstrate that, on the EEGdenoiseNet dataset, the proposed model outperforms the existing approaches.