| Epilepsy is a nervous system disease caused by the over discharge of cerebral neurons.Electroencephalogram(EEG)contains a mass of physiological and pathological information,and it is an important tool for the diagnosis of epilepsy.In clinical practice,doctor conducts the diagnosis of epilepsy via visual inspection of 24-hour electroencephalogram of patient as well as the history of epileptic seizure and family medical history.However,the sheer volume of EEG records makes visual inspection of EEG signal time-consuming,and doctor's subjective judgment may also have an impact on the detection result.Research of automated classification algorithm of epileptic EEG signal combining signal processing and machine learning based on the characteristics of EEG signal during epileptic seizure is of important significance.It not only can greatly relieve doctor's workload and improve the quality of life of epilepsy patient,but also provides a new way to study the pathogenesis of epilepsy.For that reason,more and more researchers have paid attention to it.However,the complexity and diversity of neuron firing during epileptic seizure bring challenge to automated classification algorithm.In order to solve the problem of the robustness and generalization ability of the existing automated classification algorithms of epileptic EEG signal,the distribution entropy and fuzzy entropy were fused together to propose three new kinds of embedded entropy algorithms named fuzzy distribution entropy,complex-valued distribution entropy and complex-valued fuzzy distribution entropy in this paper.On this basis,we conducted this study from following aspects: firstly,the computational complexity of automated EEG signal classification algorithms based on embedded entropy is high and the performance of classification algorithms are sensitive to parameters of embedded entropy;secondly,the feature extraction methods put forward based on embedded entropy cannot take both amplitude and phase into account which results in the loss of phase information;Thirdly,the automated EEG signal classification algorithms developed based on traditional wavelet transform have poor robustness and generalization ability.We want to explore the effectiveness of multi-domain joint analysis approach integrating time-frequency transform and nonlinear dynamic analysis for assisted diagnosis of epilepsy.The main research work and innovations of this paper are as follows.(1)Aimming at overcoming the issue of which the computational complexity of embedded entropy-based classification algorithms is high and the performance of classification algorithms is sensitive to the setting of parameters of embedded entropy,the distributed entropy and fuzzy entropy were fused together to put forward fuzzy distribution entropy.On this basis,an automated classification algorithm of epileptic EEG signal combining wavelet packet transform,fuzzy distribution entropy,Mann-Whitney U test and nearest neighbor was suggested.Fuzzy distribution entropy is able to estimate the complexity of real-valued sequence accurately and it possesses lower computational complexity than traditional embedded entropies such as fuzzy entropy.Utilizing the time-frequency local characterization ability of wavelet packet transform and the nonlinear dynamic analysis ability of fuzzy distributed entropy,the suggested classification algorithm successfully achieved the quantitative description of the complexity of EEG subbands.In the seven classification tasks of Bonn EEG dataset,the proposed classification algorithm has achieved at least 99.963% specificity,98.450% sensitivity,99.400% classification accuracy and 0.989 Matthew correlation coefficient,respectively;at the same time,our proposal also attained average classification accuracy of 99.046% and Matthew correlation coefficient of 0.795 in Children's Hospital Boston-Massachusetts Institute of Technology(CHB-MIT)EEG dataset.Experiment results show that our algorithm is more robustness and owns better classification effect than the algorithms developed based on distributed entropy and fuzzy entropy.(2)To solve the problem of which the feature extraction methods based on embedded entropy cannot take both amplitude and phase into account and results in the loss of phase information,the distributed entropy and fuzzy distribution entropy are extended to the complex-valued domain,and the complex-valued distributed entropy and complex-valued fuzzy distribution entropy were further raised based on two rules named equal ring width and equal area.On the basis of complex-valued fuzzy distribution entropy,a fusion method of flexible analytic wavelet transform,complex-valued fuzzy distribution entropy,Mann-Whitney U test and K nearest neighbors was presented for automated classification of epileptic EEG signal.Complex-valued distributed entropy and complex-valued fuzzy distribution entropy are capable of taking information of both amplitude and phase of signal into account,and it overcomes the defect of losing phase information of traditional nonlinear dynamic analysis approaches.The flexible analytic wavelet transform and the complex-valued fuzzy distribution entropy were combined organically.Taking advantage of the characteristics of partitioning the time-frequency plane flexibly of flexible analytic wavelet transform and the ability of nonlinear characterization of complex-valued fuzzy distribution entropy,the complexity of complex-valued coefficients of wavelet subband was estimated,the operation of subband reconstruction was avoided and the computation was reduced effectively.In both Bonn and CHB-MIT EEG databases,the presented classification algorithm at least attained the maximum specificity,sensit ivity and classification accuracy of 99.151%,97.329% and 99.130%,respectively,and the goal of automated recognition of epileptic EEG signal is realized.(3)For purpose of overcoming the issue of which one-basis-based classification algorithms own poor robustness and generalization ability,the multiple wavelet basis analysis was introduced and five rules for wavelet basis selection were brought forward.An automated classification algorithm of epileptic EEG signal was designed based on multiple wavelet basis lifting wavelet packet transform,hybrid features,kernel principal component analysis and least squares support vector machine.The multiple wavelet basis lifting wavelet packet transform and kernel principal component analysis were bonded.By combining multiple wavelet basis lifting wavelet packet transform and ten kinds of parameters like standard deviation,fuzzy distribution entropy and complex-valued fuzzy distribution entropy,the goal of multi-angle and multi-scale feature mining of EEG signal was achieved.The kernel principal component analysis was employed for feature selection to preserve features with high distinction degree so that the redundancy of features was reduced and the computational cost of subsequent pattern classification was further reduced.Experiment results on Bonn and CHB-MIT EEG databases indicated the minimum sum of Pearson correlation coefficient is the best rule for wavelet basis selection.The proposed classification algorithm was insensitive to the change of wavelet basis.Furthermore,our proposal respectively leaded to the least specificity and classification accuracy of 98.484% and 98.445% for different classification tasks and EEG records of different subjects.It possesses better robustness and generalization ability than existing classification algorithms.In this paper,we comed up with three kinds of multi-domain joint analysis approaches integrating wavelet analysis and modified distributed entropy to achieve the goal of accurate classification of non-seizure and seizure EEG signals.The research work of this paper has laid a theoretical foundation for the further development of the epileptic auxiliary diagnostic system and also provides a new way for the study of the pathogenesis of brain diseases. |
PDF Full Text Request