Sleep characteristics and stages detection and analysis using electroencephalogram (EEG)

Sleep characteristics and stages detection and analysis using electroencephalogram (EEG)

An electroencephalogram (EEG) signal is an efficient tool for identifying and diagnosing neurological diseases. In addition, it is very important for assisting patients with a disability to interact with their environment through a brain-computer interface. It can also assist scientists and experts to understand the most complex part of the human body, the brain. However, finding effective techniques to detect sleep characteristics and sleep stages using EEG signals is still a challenging task in sleep research, as visual detection requires advanced skills, as well as time, and effort. For example, visual scoring of sleep characteristics such as sleep spindles and k-complexes is very time consuming, subjective and sometimes does not work accurately because it requires experts to identify the presence or absence of sleep characteristics in EEG recordings. Consequently, automatically detecting and analysing sleep characteristics and stages will help sleep experts and clinical doctors to work more efficiently in diagnosing sleep disorders. This project aims to develop new and more efficient techniques to identify the characteristics of sleep stage 2 in EEG signals.

Two new techniques were proposed in this thesis to detect sleep spindles: firstly, a wavelet Fourier analysis and statistical model was used. In this method, firstly an EEG signal was divided into segments using a sliding window technique. The size of the window was 0.5 seconds with an overlap of 0.4 seconds. Then, wavelet Fourier analysis (WFA) was used to extract statistical features from each 0.5s EEG signal. The extracted features were used as inputs to a Kruskal-Wallis nonparametric one-way analysis variance to select the important features. Finally, four classifiers: a least-square support vector machine (LS-SVM), K-nearest neighbours, a k-means algorithm and a C4.5 decision tree, were used to detect the sleep spindles as well as to evaluate the performance of the proposed approach. The proposed WFA method was tested on two different EEG databases.

Secondly, a novel approach based on a time frequency image (TFI) and a fractal technique (FD) was proposed to identify sleep spindles in EEG signals. This method was employed in this thesis to investigate the main relationships between behaviours of sleep spindles in EEG signals and changes in the nonlinear features. In addition, this method was designed to improve the classification accuracy rate and to reduce the execution time. In this study, a short time Fourier transform (STFT) was applied to obtain a TFI from each EEG segment. Then, a box counting method was then applied to estimate and discover the FDs of EEG signals, as well as to extract the features of interest. Different sets of features were extracted from each TFI after applying a
statistical model to the FD of each TFI. Subsequently, four popular machine learning methods (LS-SVM, Naive Bayes, k-means and a neural network) were employed to evaluate the performance of the suggested algorithm. The obtained results demonstrated that both methods performed well and were effective in detecting sleep spindles in the EEG signals. The FDs algorithm coupled with the TFI technique improved the classification accuracy rate and reduced the execution time compared to the WFA method. The developed methods using fractal dimensions were applied to identify other sleep characteristics such as k-complexes in sleep stage 2.

Additionally, in this research, a new method was proposed for the detection of k-complexes in EEG signals based on fractal and frequency features. A dual-tree complex wavelet transform (DT-CWT) was applied to analyse EEG recording signals into frequency bands for features extraction. To select the most important feature, the extracted features were analysed. Subsequently hybrid features based on fractal and frequency features were employed to detect the k-complexes. The extracted features were then forwarded to an ensemble classifier to detect the k-complexes in EEG signals in addition to evaluating the performance of this method.

Finally, an undirected graph was used to extract the most important features from FDs. The extracted features were forwarded to the LS-SVM and k-means as classifiers to evaluate the performance of the proposed feature extraction technique and to detect k-complexes with high accuracy rate and smaller execution time. The proposed method was tested on whole EEG databases. The methods developed in this thesis aim to effectively score sleep characteristic wave forms and correctly identify the discriminative characteristics of sleep stage 2 such as sleep spindles and k-complexes using EEG signals. Furthermore, the research indicates that the proposed techniques are both practical and effective for identifying and studying the brain behaviour of sleep disorders.

Those methods can assist in the presentation of the most important clinical information about patients with sleep disorders. The outcomes from this project will help sleep experts and clinical doctors to improve their working efficiency and accuracy and will potentially reduce medical costs.