Research On EEG Signal Based On Entropy Measures And Entropy Estimators

Brain is a typical nonlinear complex system that is effected by different factors.Entropy measure is widely used to quantify the complexity of dynamic systems in different fields.Therefore,three basic entropy measures are selected: entropy,conditional entropy and mutual information processing EEG signals.However,due to the diversity of entropy measurements and estimators,the paper will use three different methods of entropy estimators(linear estimator,kernel estimator,k-nearest-neighbor estimator)to calculate entropy measures,which will be used to extract three different types EEG signals include digital features of healthy EEG,seizure-free EEG,and EEG signals during seizures.Finally,the T-test calculation results are used to measure the effect of the entropy measure under different entropy estimators on the treatment of EEG signals.First,the performance of entropy under different entropy estimators when dealing with EEG signals.It can be concluded from the calculation that the entropy can correctly extract the static digital features of different types of EEG signals.The experimental results of the three methods are roughly the same in distinguishing the performance of EEG signals under different states.EEG signals from seizures can be distinguished from the other two EEG signals,but the EEG signals between seizure-free and healthy people cannot be correctly distinguished.In addition,the linear estimator assumes that the signal to be processed obeys the Gaussian distribution in the calculation,so the type of the signal to be processed needs to be considered in the application;the entropy of the kernel estimator is different from the threshold in the kernel function when distinguishing different signals.The effect of differentiation is gradually significant until stable;the difference is that the k-nearest-neighbor estimator has no significant influence on the entropy when calculating the entropy.Second,using three different entropy estimators to calculate conditional entropy,processing EEG signals in turn and comparing the complexity of different types of EEG signals.The experimental results show that the conditional entropy of seizures is the largest,the complex conditional entropy of healthy EEG is second,and the conditional entropy of seizure-free is the smallest.In distinguishing the performance of different types of EEG signals,linear estimator can distinguish EEG signals during seizures,and can not correctly distinguish EEG signals between seizure-free and healthy people;the conditional entropy of kernel estimator is calculated in three different ways.The distribution of entropy values on EEG signals is consistent with linear estimator,but with the increase of threshold r of Heaviside kernel function,the difference between conditional entropy of EEG signals and conditional entropy entropy of healthy EEG is gradually reduced.By calculation,when the dimension value dim=3,when r=0.5,the discrimination effect of the three entropy values is the best.The k-nearest-neighbor estimator has the best discrimination effect in dealing with three different EEG signals,but the k-nearest-neighbor estimator has a great influence on the extraction of the digital features of the signal when calculating.After comparison with the other two entropy measures It is concluded that the discrimination effect is best when the dimension value dim=8 and the neighbor number k=15.Third,using different estimators to estimate mutual information to process EEG signals in different states.Mutual information under three different entropy estimator can correctly distinguish the EEG signals between seizure-free and healthy people,but can not correctly extract the digital features of EEG signals during seizures.The calculation results show that the mutual information of the EEG signals in the seizure-free is greater than the mutual information value of the healthy EEG.The linear estimator distinguishing effect is similar to the kernel estimator,and the threshold r and the dimension value of the kernel estimation have less influence on the mutual information result of the kernel function estimator;the k-nearest-neighbor estimator calculation mutual information result is consistent with the conclusions of the first two estimator methods,but in the calculation,the variable selection has a greater influence on the result.