Analyzing EEG with Machine Learning

Abstract

A brain research pursues an ultimate vision for `Improvement of Life Quality', consequently, the brain research aims to achieve better understanding of human-being and to prepare the aging society. An electroencephalography (EEG), defined as the electro-physiological monitoring method to detect and record electrical activity of the brain along the scalp, has gained considerable attention for the brain research. To date, a considerable number of EEG researches have been reported covering biomedical engineering, clinical neurology, neuroscience, and neuro-rehabilitation.

The EEG signals can be analyzed by experts, however, EEG-based applications have been improved with the help of machine learning algorithms. The goal of machine learning is to automate the process of learning. The domain knowledge in machine learning refers to all auxiliary information about the learning task. In the EEG signal processing with machine learning, the understanding of the EEG domain knowledge provides two practical benefits: First, as the part of noise elimination, it enables to remove EEG artifacts and to add complementary features from the training examples. Furthermore, it increases the transparency of the machine learning models and learning processes. The brain research, especially, requires scientific rationale accompanying reliability of performances and interpretation of the results in neuro-science aspects. In this dissertation, therefore, we aim to propose several machine learning techniques considering the EEG domain knowledge for EEG applications.

The first major part of this dissertation is devoted to the quantification of the EEG signal changes for the event detection. In general, peak amplitudes or latency features are analyzed to interpret the event related potential responses. However, these typical features have limitations on the expression as it only possesses temporal domain. We aim to apply the quantification measure by reflecting both temporal and spectral domain of the brainwave signals. We apply Kolmogorov distance measure adapted to the two different signal epochs represented in the time-frequency distribution manner. We confirm the effects of the applied distance measure to the visual evoked potentials and validate through statistical tests.

In the second part of this dissertation, we present the EEG functional networks for the personal biometric system. Recently, the functional network (FN) which fully exploits the physiological information representing functionally coupled brain regions using multi-channel EEG signals show high performance for the EEG-based biometrics. However, the FN computed from the whole channels distributed on the overall scalp is easily affected to the partially activated brain region or artifacts. We propose a method to extend the spatial domain features through the concatenation of multiple local FNs as we refer the EEG domain knowledge regarding the FN for intra- and inter- individual variability.

In the third part of this dissertation, the spectral feature fusion and exploration of the spectral features for the early diagnosis of the Alzheimer's disease is described. One of the EEG biomarkers for the Alzheimer’s disease is a slowing. In order to quantify the slowing, previous researchers have designed slowing-related EEG features in spectral domain. Although many related works have investigated critical spectral features based on the diverse spectral ranges, a comprehensive comparison between different feature sets has not been conducted yet. We first compare the various spectral feature sets, in addition, we propose the spectral feature fusion to improve the diagnostic performance. Finally, we investigate the crucial spectral ranges through comprehensive investigation using various feature selection methods.

In the last part of our dissertation, we propose a solution for early diagnosis of Alzheimer's disease based on the deep learning architecture which automatically learns data representations. Specifically, we propose the deep learning architecture fully utilizing temporal, spectral, and spatial EEG traits. The Alzheimer's disease experts, already, have reported EEG-based biomarkers: slowing and functional networks. In order to make the deep learning model learn biomarkers, we reshape the EEG signals and design the deep learning network architecture facilitating learn the EEG features in temporal-spectral-spatial domain simultaneously.