Abstract

Electroencephalography (EEG) is a popular method to detect brain-neuron activities because
of its high temporal resolution. However, very often, various types of biological and non-biological signals contaminate EEG recordings. These non-neural signals create EEG-artifacts, which cause unintentional control in the brain-computer interface related applications and difficulty in the analysis and interpretation of EEG-data. While these artifacts corrupt and mask the underlying neural activity in general, the contaminated EEG data due to the contraction and expansion of the scalp-muscles are called electromyogram (EMG) artifacts. In particular, the frontalis and temporalis scalp-muscles seriously affect the EEG-signals ranging from 0-200 Hz frequency band. This thesis studies the most common EMG artifacts originating from these two brain regions. Its aim is to analyze and remove the EMG artifacts using microstate analysis and randomization statistics.

The thesis first presents a brief literature review of the EEG-artifacts, followed by the preprocessing and analysis of the EEG recordings using EEG signal-power analysis. The purpose of this analysis is to detect the EMG contaminated EEG data-segments or epochs due to frontalis and temporalis scalp muscles (EMG-artifacts). The preliminary step in this analysis includes the transformation of the EEG epochs into the frequency domain through discrete-Fourier transform. Then the signal-powers of the EEG epochs are calculated and compared to some threshold values. These threshold values are selected based on the mean signal-power amplitudes of the EEG-epochs of the highly contaminated EEG data channels representing the frontalis and temporalis brain regions.

Electric potentials from the frontalis and temporalis region of the brain project a set of spatial patterns on the scalp surface. These spatial patterns can be clustered into a set of representative maps called microstates. Using microstate analysis, the EMG-contaminated and non-contaminated EEG epochs, obtained from signal-power analysis are clustered into an optimal number of microstates. This number best explains the data variance of both groups of EEG epochs. The difference between these microstate features can be used to distinguish artifactual and pure EEG epochs. To find the significant-differences, we have calculated the feature-differences of these microstates after a random group-wise shuffling of the EEG epochs many times to generate a distribution of the feature-differences. The research hypothesis of this distribution is that the differences in features have occurred by chance. To reject this hypothesis, we compare the probability of this distribution to the difference in features obtained before group-wise random shuffling of the EEG epochs. This technique is called multivariate randomization statistics. It has a higher statistical power compared to classical statistics to find a statistically significant difference.

In this thesis, we analyze the EEG recordings of four subjects to detect the EMG artifacts by EEG signal-power analysis. We propose a method to remove EMG artifact from EEG recordings in two steps. In the first step, we cluster the EMG contaminated and non-contaminated EEG epochs obtained from signal-power analysis into ten optimal microstates and calculate three temporal features. In the second step, through randomization statistical analysis, we differentiate between the artifactual and pure EEG epochs and reconstruct the EMG-artifact free EEG data. Finally, we validate the proposed method by comparing it with independent component analysis (ICA), a signal processing technique for separating the additive sub-components of a multivariate signal. We have found that our proposed method gives similar results to that of ICA. Our research findings suggest that a combination of microstate analysis with randomization statistics be an effective-method in the removal of EMG-artifacts.