Analysis of EEG signals using complex brain networks

Analysis of EEG signals using complex brain
networks

The human brain is so complex that two mega projects, the Human Brain Project and the BRAIN Initiative project, are under way in the hope of answering important questions for peoples' health and wellbeing. Complex networks become powerful tools for studying brain function due to the fact that network topologies on real-world systems share small world properties. Examples of these networks are the Internet, biological networks, social networks, climate networks and complex brain networks. Complex brain networks in real time biomedical signal processing applications are limited because some graph algorithms (such as graph isomorphism), cannot be solved in polynomial time. In addition, they are hard to use in single-channel EEG applications, such as clinic applications in sleep scoring and depth of anaesthesia monitoring.

The first contribution of this research is to present two novel algorithms and two graph models. A fast weighted horizontal visibility algorithm (FWHVA) overcoming the speed limitations for constructing a graph from a time series is presented. Experimental results show that the FWHVA can be 3.8 times faster than the Fast Fourier Transfer (FFT) algorithm when input signals exceed 4000 data points. A linear time graph isomorphism algorithm (HVGI) can determine the isomorphism of two horizontal visibility graphs (HVGs) in a linear time domain. This is an efficient way to measure the synchronized index between two time series. Difference visibility graphs (DVGs) inherit the advantages of horizontal visibility graphs. They are noise-robust, and they overcome a pitfall of visibility graphs (VG): that the degree distribution (DD) doesn't satisfy a pure power-law. Jump visibility graphs (JVGs) enhance brain graphs allowing the processing of non-stationary biomedical signals. This research shows that the DD of JVGs always satisfies a power-lower if the input signals are purely non-stationary.

The second highlight of this work is the study of three clinical biomedical signals: alcoholic, epileptic and sleep EEGs. Based on a synchronization likelihood and maximal weighted matching method, this work finds that the processing repeated stimuli and unrepeated stimuli in the controlled drinkers is larger than that in the alcoholics. Seizure detections based on epileptic EEGs have also been investigated with three graph features: graph entropy of VGs, mean strength of HVGs, and mean degrees of JVGs. All of these features can achieve 100% accuracy in seizure identification and differentiation from healthy EEG signals. Sleep EEGs are evaluated based on VG and DVG methods. It is shown that the complex brain networks exhibit more small world structure during deep sleep. Based on DVG methods, the accuracy peaks at 88:9% in a 5-state sleep stage classification from 14; 943 segments from single-channel EEGs.

This study also introduces two weighted complex network approaches to analyse the nonlinear EEG signals. A weighted horizontal visibility graph (WHVG) is proposed to enhance noise-robustness properties. Tested with two Chaos signals and an epileptic EEG database, the research shows that the mean strength of the WHVG is more stable and noise-robust than those features from FFT and entropy. Maximal weighted matching algorithms have been applied to evaluate the difference in complex brain networks of alcoholics and controlled drinkers. The last contribution of this dissertation is to develop an unsupervised classifier for biomedical signal pattern recognition. A Multi-Scale Means (MSK-Means) algorithm is proposed for solving the subject-dependent biomedical signals classification issue. Using JVG features from the epileptic EEG database, the MSK-Means algorithm is 4:7% higher in identifying seizures than those by the K-means algorithm and achieves 92:3% accuracy for localizing the epileptogenic zone. The findings suggest that the outcome of this thesis can improve the performance of complex brain networks for biomedical signal processing and nonlinear time series analysis.