Comparison of Different Methods on EEG Signal Separating of Stuttering Adult and Child During the Pre-speech Auditory Modulation

Abstract

During the Event-related potential (ERP) study, ideally, the EEG recording only contains the event-related signal. However, there could exist irrelevant signals and noise. Unconscious activities, such as eye movement and muscle movement, and activities caused by the design of the experiment, could occur during the recording sessions. Meanwhile, due to the hyperactive nature of the child, there is more irrelevant signal inside child EEG signals. To solve this problem. there are three methods discussed in this paper, which are averaging, independent component analysis (ICA), and Autoencoder. Averaging is the classical method applying to process data in ERP studies. Two advantages of this method are: 1) preserving the original information of the data most 2) eliminating non-activity-related Gaussian noise. There also are two pitfalls: 1) reducing the number of epoch in each group 2) failing to remove the irrelevant activity-related signals. This method is also unable to get useful information from the child data. And the signal to noise ratio (SNR) of this method is 30.21 for adult subjects. ICA, a linear blind source separation method, is also a common method used by some of the studies. There are two advantages to this method: 1) preserving the number of epoch in each group. 2) removing the irrelevant eye movement and muscle movement signals. One pitfall is that bad rejection choice may cause losing information. This method improves some of the results in child subjects. And the SNR of this method is 33.02 for adult subjects, which is higher than averaging. Autoencoder is a nonlinear dimensionality reduction method. By creating proper loss function, a nonlinear independent feature learning method is applied to the EEG signals. The advantages are 1) nonlinearly learning the feature and linearly reconstructing the data at the same time 2) dimensionality reduction. One pitfall is currently no localization method to validate the features. And the SNR of this method is 22.94 for adult subjects, which is lower than averaging. And Autoencoder also can process part of the child data.