Magnitude Squared Coherence Research Of Eeg Signals In Mild Cognitive Impairment Of Diabete Patients

Electroencephalogram is the general reflection of electrophysiological activities by nerve cells of the cerebral cortex.The research on synchronization between different brain regions or neurons of electroencephalogram is of great importance to explore the pathogenesis,early diagnosis and intervention treatment of brain disorder.This paper aims at studying coherence algorithm and applying the coherence algorithm to real EEG signals of diabetic patients.Firstly,three basic coherence algorithms in frequency domain are studied,which include Welch's averaged periodogram based magnitude squared coherence,minimum variance distortionless responce based magnitude squared coherence and canonical correlation analysis based magnitude squared coherence.The property of the three methods are analyzed and compared by theoretical analysis and simulation experiments.Secondly,a new method named weighted canonical correlation analysis based magnitude squared coherence is proposed on the basis of the canonical correlation analysis based magnitude squared coherence and reduce rank canonical correlation analysis based magnitude squared coherence.The simulation results indicate that the proposed method behaves better in the influence of noise and relative amplitude of frequency components and provides more accuracy.Finally,weighted canonical correlation analysis based magnitude squared coherence is applied to real EEG signals of 35 diabetic patients(18 aMCI and 17 normal controls).The results of statistic analysis indicate a significant increase in Delta and Theta coherence and a significant decrease in Alpha coherence in aMCI patients.The results of correlation analysis show that each significant coherence values has significant correlation between neuropsychological test scores.The result of classification by support vector machine shows that the averaged accuracy,sensitivity and specificity are 74.23%,82.44% and64.86%,in which the maximization of area under ROC curves acts as the feature selection criterion.Based on the above analysis,the proposed weighted canonical correlation analysis based magnitude squared coherence is able to develop the coherence analysis of EEG data effectively and it is important in the early diagnosis of aMCI for diabeticpatients.