The Study Of Electroencephalograph Feature Extration And Clssification In Children With Autism Spectrum Disorders

Autism spectrum disorder(ASD)has been defined as individual with deficits in social interaction and language development,a restricted repertoire of interests,as well as cognitive skills.With the increased population of ASD individuals,EEG has been widely used technology to investigate and diagnose the abnormal brain function,including autism spectrum disorder ASD,attention deficit hyperactivity disorder(ADHD),and Alzheimer etc.Since EEG contains abundant physiological information and responds the change of different physiological status.According,it is one of the most important tools to research on brain function and psychiatry.In this study,we collected the 15 children with ASD and 15 age matched typical development children.And the EEG signal was recorded during an eye-open resting state and decomposed into delta,theta,alpha,beta frequency bands.Then single-channel and two-channel methods were used to extract the features,such as four kinds of entropy(approximate entropy,sample entropy,permutation entropy and wavelet entropy),bio-coherence of different frequency bands,coherence of different frequency bands,phase synchronization etc..Next the effective indicators were selected using maximum relevance and minimum redundancy(mRMR)algorithm and the subjected to the support vector machine(SVM)algorithm to discriminate children with ASD from the typical development children.Finally we compared these features and made a classification model for ASD.The result showed that the ASD had significantly greater relative delta and lower relative alpha in whole regions.The relative power of beta has the highest accuracy which has reached the 90.11%±4.50% in the whole brain.The classification accuracy of wavelet entropy,permutation entropy and sample entropy has reached 85.00%±3.86%,81.78%±1.69%,79.11%±2.13% respectively.And the two-channel methods including coherence,phase synchronization,corr-entropy had the classification accuracy of78.11%±4.38%,77.67%±4.03%,89.00%±3.17%.Despite the performance of each method alone was not well,the combination of all these indicators had high accuracy of 97.78%after feature selection by mRMR.In conclusion,the study aimed at the clinical utility of these indicators and found 13 most effective indicators to make a classification model for ASD.This diagnostic model only need 10 minutes EEG at rest with eyes open,which could be a potential method for assistant diagnosis of ASD.In this study,we found that combining several separable indicators,and then we could improve the diagnostic accuracy.