Research On Decision-making Support Of Stroke Rehabilitation Based On Deep Neural Network

Stroke is an acute cerebrovascular disease,which leads patients to having physical dysfunction and inconvenience in their daily life.Rehabilitation treatment can improve the pathological conditions of stroke patients,reduce the degree of physical disability of patients,improve the ability of daily living of patients,and promote patients to better integrate into the family and society.The effect of traditional rehabilitation treatment is affected by subjective factors of physicians,and the application of artificial intelligence in rehabilitation treatment can provide effective auxiliary decision support for physicians in the treatment process.Due to the lack of selection and analysis of assessed actions and the failure to consider the correlation of treatment options in auxiliary decision-making of stroke rehabilitation.In this thesis,auxiliary decision-making research is carried out from the three aspects of automatic assessment of stroke,recommendation of rehabilitation treatment plan and prognosis prediction,and the auxiliary diagnosis and treatment system for stroke rehabilitation is designed and implemented.The specific work are as follows:1.The automatic assessment of stroke rehabilitation using a single action did not consider the selection of action assessment and effect comparison.In this thesis,a Bidirectional Enhanced Long Short-Term Memory with Clustering（BELSTMC）model is proposed.The BELSTMC model takes Bi-ELSTM（Bidirectional Enhanced Long Short-term Memory）as the basic network,and the Rough Fuzzy C-means（RFCM）is fused.In this thesis,accelerometer is used to collect the data of the patient’s forearm,forearm and shoulder during the performance of the assessed movements,and experimental analysis is conducted on the five assessed movements.The Brunnstrom scale staging accuracy is 100.00% for the internal and external rotation assessed movements.2.The strong correlation of rehabilitation treatment plan is not considered in the current recommendation of rehabilitation treatment plan for stroke,Convolutional Neural Network based on Class Correlation Learning（CNNCCL）is proposed.The CNNCCL uses a two-layer Network.The main network learns features independent of categories,and the sub-network learns the correlation between classes.Residual links are used to encode the attention mapping learned by the sub-network into the convolution features of the main network,so as to realize the class correlation learning of the treatment plan.Experimental results show that the recall rate and accuracy of CNNCCL model are 83.30% and 71.40%,respectively.3.A Recurrent Neural Network based on Attention Mechanism（RNNAM）is proposed in view of the existing prognosis prediction of stroke rehabilitation without considering the prediction effect of different scales.It adopts Gate Recurrent Unit（GRU）as the basic network of RNNAM,introduces LSTM network to extract text features,and uses self-attention mechanism to locate key features.The three scales are predicted in the experiment.The accuracy of Fugl-Meyer Upper Limb Rating Scale is 90.70%,and the determination coefficient is 0.89.4.On the basis of the research on the related algorithm of intelligent assisted decision of rehabilitation treatment,the stroke rehabilitation assisted diagnosis and treatment system was further designed and implemented.According to the business logic of clinical stroke treatment,the functions of patient management,prognosis prediction,evaluation plan management and treatment plan recommendation are realized respectively,providing effective auxiliary decision support for clinicians.