Research On Automatic Classification Algorithm Of Epileptic EEG Signals Based On Unsupervised Learning

Epilepsy is a disease of human nervous system.Intractable epilepsy accounts for about 30% of all patients.This kind of patients cannot be effectively controlled by drugs.This kind of epilepsy needs surgical resection of epileptic foci to control seizures.Therefore,it is very important to accurately locate epileptic foci before operation.Recently,high frequency oscillatory signals（HFOs）have attracted more and more attention.Studies have shown that there are certain differences in HFOs inside and outside the epileptic focus.By analyzing HFOs in EEG signals,the epileptic focus can be located,which is conducive to the implementation of resection.Traditional methods rely on manual screening of HFOs in EEG signals.However,due to the huge amount of EEG data,visual screening is time-consuming,laborious and subjective.This study combines the idea of unsupervised learning and deep learning method to explore the application of deep unsupervised learning algorithm in automatic detection of HFOs in EEG signals.The main contents and contributions of this study are as follows:（1）A total of 538 hours of intracranial electroencephalogram（i EEG）signals were collected from 5 patients with intractable epilepsy.The i EEG was prescreened by shortterm energy method,and a total of 4042 candidate high-frequency oscillation（c HFOs）signals were obtained.Based on the graphical user interface（GUI）of MATLAB,the calibration system of c HFOs was designed,and the calibration of all chfos was completed,including 1611 real HFOs（r HFOs）and 2431 non HFOs（non-HFOs）;（2）An unsupervised HFOs automatic detection method based on domain countermeasure network（DANN）is proposed.The idea of transfer learning and deep clustering method are combined to automatically detect HFOs in i EEG.It is discussed that the ventricular fibrillation data set is used as the source domain data,and the c HFOs database constructed in this thesis is used as the target domain data to train and test DANN.The results showed that the accuracy,sensitivity and specificity of this method were73.51%,71.76% and 74.96% respectively.（3）An unsupervised HFOs automatic detection method based on convolutional variational automatic encoder（CVAE）is proposed.The unsupervised learning idea and deep learning method are combined to automatically detect HFOs in i EEG.Firstly,CVAE is used to extract the features of HFOs time-frequency map.Then,K-means algorithm and fuzzy c-means algorithm are used for feature clustering analysis.Finally,the database constructed in this thesis is used to evaluate the performance of this method.The results showed that the accuracy,sensitivity and specificity of this method were 93.39%,94.41%and 92.71% respectively.Combined with the results of the above research,the deep clustering model designed in this study can be used for the automatic detection of HFOs,and can assist doctors to analyze HFOs to locate epileptic foci,which has a certain clinical value.Due to the limited research time and the limited number of clinical patients,this thesis only makes a preliminary exploration and Analysis on the application of deep unsupervised learning in the automatic detection of HFOs.In the follow-up,it is still necessary to improve the network model and add data sets to further evaluate and verify this method.