| Schizophrenia has always been a central topic in the study of psychiatric diseases due to its unknown etiology,varied symptoms and characteristics that have a great impact on patients,families and society.Currently,the diagnosis and treatment process mainly depends on clinical symptomatological criteria such as ICD-10 and DSM-5,the behavior of patients and the clinical experience of doctors.Therefore,the research on the auxiliary diagnosis of schizophrenia based on electroencephalogram(EEG),functional magnetic resonance imaging(f MRI)and computed tomography(CT)is very important and urgent.In this paper,the EEG data of first episode schizophrenia patients(FES)and health control(HC)were mainly fucused.Besides the previous research in the medical field,machine learning and deep learning tools were also used to explore the assisted diagnosis of schizophrenia in a gradual and in-depth way.First of all,EEG data was processed into 3 forms: frequency domain-2d time domain-3d time domain to classify FES and HC.In order to exploit the characteristics of EEG,integrated learning,attention mechanism and residual network were used.A comprehensive framework classification based on EEG of schizophrenia was built.Secondly,on the basis of completing the classification task of EEG data with high accuracy,the concept of EEG channel weight was introduced to represent the ”importance” of64 channels in the classification task.A two-stage method was designed to obtain the EEG channel weight ranking adaptively and make the EEG channel weight more sparse after network dynamic pruning.The weight ranking of EEG channels is conducive to improving the accuracy of classification tasks.Meanwhile,the visual results of weight ranking show that the changes in frontal lobe and parietal lobe are closely related to schizophrenia,which is consistent with the research results in the medical field.Finally,channel selection method was proposed based on EEG channel weight ranking and analysis of the correlation between channels.Using selected 32 channels derives higher accuracy compared with 64 channels of EEG.And maybe the cause is that the redundant information is eliminated.At the same time,due to channel selection,the time consumption of the classification model is optimized to a great extent,which is more conducive to the practical application and promotion of the method.In addition,it is also helpful to the research of EEG acquisition device,so that it not only has the spatial advantage of multi-lead device,but also does not lose the accuracy of classification.This paper is organized as follows: 1)the classification of EEG from three perspectives;2)classification optimization with EEG channel weight ranking;3)EEG channel selection.We hope the research is helpful to the specific diagnosis of schizophrenia and the cost control and promotion of EEG device. |
