| The disturbance of consciousness is mainly a state in which the patient has an obstacle to the perception of the environment or himself,manifested as the obstacle of awakening or the obstacle of consciousness content.Clinically,the judgment of consciousness disorder mainly depends on expert experience.The judgment based on expert experience is often subjective,and the result is not accurate.And the judgment based on the scale often requires a lot of time to evaluate.In recent years,neuroimaging technology and neuro electrophysiological technology are often used in the diagnosis of brain related diseases,among which electroencephalogram(EEG)is often used in the diagnosis of epilepsy,schizophrenia,depression and so on.This thesis is mainly based on the analysis of EEG data of patients with consciousness disorder,and completes the recognition of abnormal EEG signals and the recognition of consciousness state of patients with consciousness disorder.The main contents of this thesis are as follows:(1)Identification of abnormal EEG data.Because EEG signals show certain randomness and EEG data are easy to be disturbed,there are a large number of abnormal data.The traditional supervised learning recognition method is limited by the recognition accuracy,model robustness and the need for clearly labeled abnormal EEG data.Based on the idea of anomaly detection,uses unsupervised model for the recognition of abnormal EEG data,in which network is used for the unsupervised training of EEG data,so that the network can learn the normal EEG characteristics,minimize the reconstruction loss of normal EEG and maximize the reconstruction loss of abnormal EEG.The threshold is set through the experiment to identify the abnormal EEG data.In order to verify the effectiveness of the algorithm,the accuracy rate,missing rate,F1 score and other indicators are used.In order to improve the recognition rate of the network,the depth convolution self encoder(DCAE)network is improved,the attention mechanism module is introduced,and the deep convolutional self encoder for attention mechanism(DCAE-AT)model is constructed.The experimental results show that the performance indexes of the improved algorithm are greatly improved.(2)Recognition of state of consciousness.The EEG data containing abnormal fragments were removed and used for the recognition of awake and drowsy states of consciousness.In order to recognize effectively,the idea of model fusion is used for reference,and the feature fusion model of convolution and cyclic network fusion(CNN-M-Bi LSTM)is used to recognize the conscious state of waking and sleeping.The comparison of accuracy,retrieval rate and F1 score shows that the recognition rate of the model used in this thesis is higher.In order to further improve the recognition ability of the model,the convolution(CNN)network is improved.By introducing the convolution network of convolution kernels of different sizes,and determining the size of convolution kernels of different sizes through experiments,the multiscale convolution and cyclic network fusion(multi-scale CNN-M-Bi LSTM)model is finally constructed.The experimental results show that the performance indexes of the improved model are better.A joint recognition experiment of abnormal EEG and state of consciousness is designed,finally,88.63% of the comprehensive recognition results are obtained.(3)Design and implementation of telemedicine system.According to the demands of the system,the front and rear end separation technology is adopted.The online consultation for the patients/family members,the intelligent recognition of abnormal EEG data and state of consciousness,remote nursing of the patients are realized. |
