Deep Learning Technique to Denoise Electromyogram Artifacts from Single-Channel Electroencephalogram Signals

Abstract

The adoption of dependable and robust techniques to remove electromyogram (EMG) artifacts from electroencephalogram (EEG) is essential to enable the exact identification of several neurological diseases. Even though many classical signal processing-based techniques have been used in the past and only a few deep-learning-based models have been proposed very recently, it is still a challenge to design an effective technique to eliminate EMG artifacts from EEG. In this work, deep learning (DL) techniques have been used to remove EMG artifacts from single-channel EEG data by employing four popular 1D convolutional neural network (CNN) models for signal synthesis. To train, validate, and test four CNN models, a semi-synthetic publicly accessible EEG dataset known as EEGdenoiseNet has been used the performance of 1D CNN models has been assessed by calculating the relative root mean squared error (RRMSE) in both the time and frequency domain, the temporal and spectral percentage reduction in EMG artifacts and the average power ratios between five EEG bands to whole spectra. The U-Net model outperformed the other three 1D CNN models in most cases in removing EMG artifacts from EEG achieving the highest temporal and spectral percentage reduction in EMG artifacts (90.01% and 95.49%); the closest average power ratio for theta, alpha, beta, and gamma band (0.55701, 0.12904, 0.07516, and 0.01822, respectively) compared to ground truth EEG (0.5429; 0.13225; 0.08214; 0.002146; and 0.02146, respectively). It is expected from the reported results that the proposed framework can be used for real-time EMG artifact reduction from multi-channel EEG data as well.