Eye Blink Artifact Detection In Epileptiform Assisted Analysis

It is inevitable that Electroencephalogram(EEG)will be interfered by wink artifact during the acquisition process.In intelligent medical-assisted diagnosis,EEG of epileptic patients is usually accompanied by epileptiform pathological signals,whose morphological characteristics are very similar to blink artifacts.Therefore,eye blink artifacts can be misleading to the intelligent diagnosis of EEG signals.At present,the mainstream eye blink artifact detection methods are mainly aimed at the EEG of healthy people,which lacks the processing of pathological signals,greatly increasing the possibility of signal misjudgment,and thus seriously affecting the diagnosis of diseases.At the same time,due to the low spatial resolution of EEG,the details of EEG may be blurred,thus affecting the accuracy of eye blink detection result,which has been less mentioned in the field of eye blink artifact detection.In addition,there are significant individual differences in EEG amplitude,spectrum and other aspects,but most of the current methods lack specific treatment for this problem,resulting in poor generalization of the model,the performance of the complex real EEG data can not meet the expectations.To solve the above problems,this paper combined EEG analysis and artificial intelligence,mainly focused on the eye blink artifact detection algorithm of EEG signals in epilepsy patients.The main research work is as follows:1.Aiming at the problem that the current mainstream algorithms lack of spatial domain filtering for low spatial resolution EEG signals,individual differentiation processing and lack of feature characterization ability,which lead to the degradation of blink artifact detection performance in the background of epilepsy,a eye blink artifact detection algorithm based on multi-dimensional EEG feature fusion and optimization(ESP-PVM)was proposed.In order to solve the problem of low spatial resolution of EEG signals,the two-dimensional frontal pole channel signals were firstly upgraded to multi-channel signals by empirical mode decomposition(EMD)and autocorrelation coefficient screening.Combined with the common spatial pattern(CSP),the obtained signals are filtered in the spatial domain,and the variance features of the filtered signals are extracted.At the same time,in order to improve the characterization ability of background signals and blink artifacts containing pathological signals,the proposed method extracted the 16-dimensional key features of the original signals covering the time domain,frequency domain and statistics,and fused with the spatial domain variance features to form a feature library.In addition,compared with the variance filter method,a wrapped feature optimization method was designed based on particle swarm optimization(PSO)and support vector machine(SVM)to solve the problem of individual difference.The method was verified on real patient data from the Children’s Hospital Zhejiang University School of Medicine(CHZU),and the average F1 value of all data sets reached 98.03%.2.In order to solve the problems of high time cost of ESP-PVM and its inadaptability to high frequency pathological signals,a multi-dimensional feature fusion and optimization algorithm(MMF-PCC)based on deep learning and signal standardization was proposed.Firstly,a lightweight network with pre-trained weight Mobile Net V3 is used to extract depth features from EEG time-frequency graphs.Secondly,EEG signal standardization and smooth nonlinear energy operator filtering were carried out to enhance the differentiation between pathological signals and blink artifacts,and then time-frequency domain features were extracted.The feature database is built by integrating depth feature and time-frequency domain feature.In order to solve the problem of high time complexity of the feature optimization part of ESP-PVM,the evaluation index of PSO-SVM was improved to filter the combination of features The average correlation coefficient(ACC)is the fitness of the optimization algorithm,which can be iteratively selected the best solution.Compared with the previous method,although the average F1 value of the proposed algorithm on all CHZU data sets is reduced by0.27%,the average F1 value on some data subsets containing high-incidence disease signals is increased by 4.61%,and the time consumed by the feature optimization part is reduced 107 times on the individual data set and 624 times on the mixed data set.