| Stroke is an acute cerebrovascular disease in which brain tissue is damaged.The mortality rate of patients with cardiovascular and cerebrovascular diseases in China is on the rise,and the number of stroke patients ranks first.There are about 13 million patients,the incidence rate is 345.1/100,000,and the average age of onset is(66.4±12.0)years old.The economic burden is as high as 40 billion yuan per year.More and more stroke patients require rehabilitation therapy to restore limb motor function.Motor imagery(MI)based brain-computer interface(BCI)technology is a new motor function rehabilitation therapy for stroke patients,which decodes MI intentions by analyzing electroencephalogram(EEG)signals,enables communication and communication between the brain and external devices.Based on EEG signal analysis of stroke patients,this thesis uses Convolutional Neural Networks(CNN),Long Short-Term Memory(LSTM),Gated Recurrent Unit(GRU)and Bidirectional Long Short-Term Memory(BI-LSTM)deep learning network,extracts the EEG signal features of stroke hemiplegia patients and identifies the MI intentions of the left and right hands,and achieved fruitful results.The main research contents and results are as follows:1.A module composed of LSTM and Fully Connected Layers(FC)is designed,and an algorithm including 64 LSTM-FC-LSTM parallel modules is proposed.Each LSTM-FC-LSTM calculates one channel data to independently extract the features of each channel.The algorithm introduces the Spatial Dropout1 D layer to realize the independent selection of channels related to classification tasks for different subjects.Statistically analyze the correlation between channels,and finally obtain the correlation channels between healthy people and stroke patients during left or right-hand MI.It can effectively solve the problem of high-precision selection of corresponding channels in activated brain regions caused by different brain injury regions.2.The application of five classification models in left or right-hand motor imagery BCI was discussed,and the advantages,disadvantages and application scope of each model were compared comprehensively,which provided theoretical and experimental basis for the selection of real-time BCI classification models.Experimental results show: the GRU-GRU model and the BILSTM-BILSTM model can better fit the individual training set and achieve higher classification accuracy with fewer training samples;the LSTM-LSTM model has the shortest offline calibration time,but its classification accuracy is low;in terms of offline calibration time and average classification accuracy,the BILSTM-LSTM model still performs the best.3.In order to reduce the influence of the difference of EEG signals between different subjects on the classification accuracy,this thesis designs a CNN-LSTM algorithm.The MI intent recognition rate of this model reaches 0.89,the Precision value is 0.88,the Sensitivity value is 0.89,and the F1-score value is 0.89.It is confirmed that the model gets rid of individual differences,does not rely on offline calibration for a specific individual,and independently extracts the essential features of left-and right-hand MI.By analyzing the phenomenon of event related desynchronization/synchronization(ERD/ERS),we analyzed the activation of brain regions around C3 and C4 channels when healthy people,stroke patients with left hemiplegia,and stroke patients with right hemiplegia underwent left or right-hand MI.It can effectively improve the classification and establishment of BCI,and provide an experimental basis for the localization of damaged brain regions.MI can activate brain neurons,complete the functional reconstruction of brain nerves,and help stroke patients partially recover or restore their motor abilities.BCI technology utilizes MI features to control external devices.The research work in this thesis strongly promotes the development of MI-based BCI technology for the rehabilitation of stroke patients. |
