Automatic Assessment Of Upper Limb Motor Function After Stroke Based On Deep Learning

Due to the characteristics of high incidence and disability,stroke has become one of the common diseases in the clinic,and various dysfunctions are often left behind.Among them,upper limb dysfunction is the most widely affected limb dysfunction.Recently,rehabilitation medicine is gradually developing towards precision,remoteness,intelligence,and personalization,and accurate assessment is one of the important directions.In order to make up for the lack of traditional assessment methods and facilitate the rehabilitation assessment in the home and community environment,the integration of sensor technology and artificial intelligence to achieve automatic assessment of motor functions is receiving increasing attention.In this thesis,we conduct a deep learning-based study of the automatic assessment of upper limb motor function in stroke.A unique human motion measurement system is designed from three levels,including arm dysfunction,hand motor dysfunction and upper limb mobility.Three different deep learning methods are established for feature extraction,and classification or regression analysis to realize the clinical scale stage and score prediction.The specific research works are as follows:1.For the evaluation of upper limb arm motor function,this thesis independently designs and develops an inertial measurement unit that can acquire acceleration,angular velocity and magnetic data,obtains inertial movement data of the big arm and forearm during simple upper limb evaluation movement,and proposes time attention recurrent neural network model to extract and classify data features to achieve accuracy of 100% for Brunnstrom staging of upper limb motor function.2.For the evaluation of hand motor function,this thesis combines the IMU and flexible pressure sensors to design a universal data glove to sense the inertial movement of the hand and the finger pressure.Based on the acceleration and angular velocity,the quaternion method is used to calculate the spatial attitude angle of the finger,and a geometric relationship model is established to calculate the range of motion(ROM)in real time.This thesis introduces a filter regularized full convolutional network(RFCN)to extract features from the sensor data and establish a fully connected regression model to predict Carroll's hand Motor function score.The experimental results show that the coefficient of determination between the predictive score and the experienced clinician's score is 0.93.3.In order to evaluate the mobility of the upper limbs,we design a unique assessment action and experimental environment based the clinical Fugl-Meyer scale.The vision-based depth sensor is used to track the spatial position of the upper extremity bone points.We introduce a group-constrained convolutional recurrent neural network to extract the spatiotemporal features of the data,and establish a regression model to predict the upper limb FMA score.Comparative experiments show that the proposed method can complete the FMA assessment process modeling by only 4 assessment tasks,and achieve 0.89 high consistency with the clinician assessment results.4.This thesis uses Eclipse development tools and MySQL database to design and implement a stroke rehabilitation assessment management system prototype,which includes user management,patient management,assessment program management,assessment result management,scale management,assessment report generation and export module,which can be implemented Information management function of stroke clinical rehabilitation assessment process.Based on the research results in this thesis,it can be found that the use of sensor sensing technology and deep learning artificial intelligence analysis method can achieve accurate measurement and quantitative assessment of motor function of stroke patients,which makes up for the shortage of traditional assessment methods and facilitates the development of rehabilitation assessment in family and community environment.