Feature Extraction And Brain Network Analysis Of MEG

Schizophrenia is a psychiatric disorders, and the number of patients accounts for about 1% of the world's total population. The clinical symptoms of schizophrenia vary, and exhibit symptoms with significant differences even at different stages of disease in the same patient. At present, the etiology and pathogenesis of schizophrenia are not clear, and the treatment and severity can be assessed according to the studies of the patients' brain signals. The MEG has advantages of capturing resolution images and accurately positioning. In this paper, the MEG of normal subjects and schizophrenic patients are analysed from the perspectives of feature extraction and functional brain networks. The main work is as follows:Firstly, the paper proposes a new method of feature extraction and classification, in which the optimal wavelet package decomposition and the energy entropy are combined. The MEG data are decomposed by wavelet packet after the dimensionality reduction with PCA, and the optimal wavelet packets basis can be selected from the obtained wavelet packet library based on wavelet entropy. Then the amplitude modulation(AM) is applied to the optimal wavelet coefficients to acquire the energy entropy of the envelope, and the statistical properties of energy entropy are considered as the features to be classified. This method not only takes the larger dimension characteristics of MEG into account, and also retains the characteristic information effectively via the amplitude modulation.Secondly, according to the theory of complex networks, this paper presents a new method to build the brain functional network. For the resting MEG signals of schizophrenic patients and healthy subjects with eyes closed in left temporal and frontal regions, the dynamic function connectivity matrix can be obtained by calculating the Pearson correlation coefficients with the use of the sliding window and short-time Fourier transform. After that, extracting parameters of the constructed weight and binary network based on graph theory and threshold selection to study the small-word properties of brain networks.Finally, it is found that the accuracy of classification of the MEG automatic classification method based on wavelet packet and energy entropy is up to 97.5868%. At the same time, the experimental results of the functional networks construction method show that patients with schizophrenia have smaller absolute clustering coefficient and larger shortest path length for different functional networks compared to healthy subjects'. Furthermore, the brain networks ofhealthy subjects show more significant small-world properties, and information can be more efficiently transmitted between brain regions.