| Limb motor dysfunction caused by stroke,brain trauma,etc.is characterized by high incidence rate and high disability rate.Timely and effective rehabilitation training can help the patients recover certain motor function and reduce disability rate.However,in remote areas and developing countries such as China,due to the large number of discharged patients,shortage of rehabilitation resources and uneven distribution,discharged patients are unable to obtain timely and effective rehabilitation,which may affect their ability to take care of themselves or even lead to lifelong disability.At the same time,the current clinical scale assessment has shortcomings,such as excessive time consumption,strong subjectivity,and "ceiling effect" caused by coarse grading,unable to assess the expected rehabilitation results,which cannot provide accurate and objective motor function assessment results and refined rehabilitation plans.In view of the above clinical pain points,this thesis proposes an intelligent rehabilitation assessment model based on wearable sensors and machine learning,investigates the intelligent quantitative assessment model and algorithm of limb movement function without physicians’ participation,gives system implementation and conducts a clinical application research.The main research contents and results are as follows.(1)Design of wearable sensors and the wireless data acquisition systemWireless wearable inertial measurement units(IMUs)and rehabilitation gloves were designed based on 9-axis sensor chips,flex sensors and ZigBee wireless microcontrollers,to capture the rehabilitation movements of limbs and hands in the rehabilitation process with high accuracy.Sensor calibration and coordinate system transformation methods had been proposed to eliminate the effect of wearing position on rehabilitation assessment results.For the packets loss problem during wireless transmission of wearable sensors,a multi-node wireless motion data modulation and reconstruction method based on heat map attention mechanism was proposed,which contains several Convolutional Neural Networks(CNN)filters for hardware data compression sampling and attention heat map feature extraction for key features of motion data.And the heat map was used for a priori verification of data reconstruction and data packet loss and misalignment verification in wireless data transmission during the multi-node sensor data decomposition.(2)Research on quality assessment of rehabilitation movementsFor the assessment of individual movement quality,the sensor signals were pre-processed by filtering,resampling and normalization,and then the signals were period segmented and valid period extracted.Furthermore,the rehabilitation action features were extracted,and random entropy(ApEn)and dynamic time warping(DTW)distance were used as the deep features.For the multi-source features from multiple sensors,the method of determining weights by information entropy was proposed,and the scores of individual movements were calculated by the product sum of the multi-source features and the corresponding weight matrices.The scores were standardized to a score of 0-100 or 0,1,or 2 corresponding to the clinical Fugl-Meyer assessment(FMA)scale,which can effectively solve the problems of thick grading and "ceiling effect" of clinical scales.(3)Quantitative assessment and intellectual staging of limb motor functionThe single-movement scores of 32 rehabilitation movements were characterized,and the clinical results of rehabilitation physicians were used as labels to establish intelligent rehabilitation evaluation models using a variety of machine learning algorithms suitable for small samples and with advantages in preventing overfitting.The collected data of 232 patients’ rehabilitation assessment movements was divided into training and testing sets by 7:3,and 10-fold cross-validation was used to evaluate the performance of the designed machine learning models.The results showed that the SVR-based intelligent rehabilitation assessment model had the best performance in terms of linearity,root mean square error(RMSE),and mean absolute percentage error(MAPE)with R2=0.962,RMSE=5.38,MAPE=12.6%,and P<0.001.In terms of intelligence staging,SVM model had the best performance in all indicators,among which,for the upper limb intelligence staging,the accuracy was 0.886,the Log-loss was 3.947,the Hamming distance was 0.114,the Jakard similarity coefficient was 0.795,and the Kappa coefficient was 0.853.(4)Research on the expected rehabilitation effect assessment and individualized intervention methodBased on the quantitative assessment of limb motor function,the information related to patients’ basic information,medical history,rehabilitation assessment results,and rehabilitation intervention methods are incorporated,and the features and weights are dynamically adjusted through multi-source information fusion and feature extraction based on the temporal attention mechanism.For predicting the discharge rehabilitation outcome of admitted patients,in terms of feature importance,the total score at admission,upper limb score,rehabilitation intervention mode,onset days,and age are the five most important features for predicting the discharge rehabilitation assessment outcome.Based on the fusion of multi-source spatio-temporal information,a rehabilitation outcome prediction model and an individualized rehabilitation intervention model are established based on knowledge distillation,and continuously optimized by feedback and iterative update against losses,to optimize the expected rehabilitation effect and individualized rehabilitation intervention prescription.(5)Research on clinical applicationBased on the above research,we deployed the model and algorithm,built the remote intelligent rehabilitation assessment system and platform,and completed the system interfacing and test.A clinical application study was conducted in 120 stroke patients in Tangdu Hospital Airforce Medicine University and Xi’an Gaoxin Hospital to verify the reliability and validity of the intelligent rehabilitation assessment model.The results showed that the results of the intelligent rehabilitation assessment were significantly correlated with the results of the Fugl-Meyer assessment for rehabilitation physicians(R2=0.967,P<0.001).The mean of the deviation of the total score of the physician Fugl-Meyer assessment and the total score of the intelligent rehabilitation system was 0.30,95%CI was-0.57 to 1.17,and the relative deviation(%)was 3.49.In addition,the mean time spent by the intelligent rehabilitation assessment system was 35.00%less than the rehabilitation physician’s Fugl-Meyer assessment(P<0.05).Regarding the effect of individualized prescription rehabilitation,after 3 weeks,the test group with individualized prescription exercise rehabilitation training using the intelligent rehabilitation system showed better improvement in rehabilitation scores than the control group with traditional OT training.The above results indicate that the designed intelligent rehabilitation assessment model and system based on wearable devices and machine learning can perform accurate and objective rehabilitation assessment and effective rehabilitation training without physicians’ participation,which provides an effective technical solution and design reference for solving the shortage of rehabilitation medical resources and rehabilitation of discharged patients. |
PDF Full Text Request