Developing An ICF Core Set To Automatically Assess Upper Extremity Motor Function In Stroke Patients Based On Body Sensor Network

The assessment of upper limb motor function is of great importance for strokerehabilitation. While in the traditional clinical practice, too many outcome measuresmake it hard for the communication between professionals from different areas.What’s more, most of traditional rehabilitation evaluation methods rely on clinicians’experience and are characterized by subjective and time-consuming properties.Therefore, a simple, objective and quantitative unsupervised assessment is required,especially in the home-based rehabilitation. The International Classification ofFunctioning, Disability, and Health (ICF)provide a unified and standard language andframework for the description of all aspects of human health and functioning. Theautomated and quantitative assessment can be attained with a Wireless Body SensorNetwork (WBSN) system which includes several motion sensor nodes. To achieve thestandardization and automation of rehabilitation, we developed a minimized ICF coreset for stroke patients’ upper limb motor function using signal processing and machinelearning methods. The automated evaluation was based on sensor data collected byWBSN system. Main contents of this paper are as follows:Firstly, The ICF core set was developed by linking ICF to Fugl-Meyerassessment scale which is most frequently used in the clinical practice. UpperExtremity Fugl-Meyer scale was divided into three sections-shoulder and elbowmovements, wrist movements and finger movements-corresponding to three ICFcategories respectively, which were b760Control of voluntary movement functions,b710Mobility of joint functions and d440Fine hand use. Thus compose theminimized ICF core set for stroke patients’ upper extremity motor function.Secondly, the automatic evaluation system was established.7tasks were selectedfrom the Upper Extremity Fugl-Meyer scale according to expert advice and motiondata was collected from24stroke patients using a WBSN system comprised of twoarm nodes (integrated with accelerometers) and one rehabilitation glove (integrated with six curvature sensors). Using the collected data, three Fugl-Meyer scoreprediction models for shoulder-elbow section, wrist section and finger section werebuilt respectively. The output Fugl-Meyer scores from the models were converted tothe corresponding ICF qualifiers, thus realizing the automated estimation of theconstructed ICF core set for stroke patients’ upper extremity motor function.Thirdly, comparative study of several prediction models was performed.Single-task prediction models were built using Support Vector Regression (SVR) andExtreme Learning Machine (ELM) algorithms. Then three comprehensive modelswere constructed which linearly mapped the relationship between the scores predictedby the corresponding single-task models individually and the actual FMA scoreprovided by a clinician. Cross-validation results showed that ELM models couldachieve similar performance with SVR models but could be implemented moreefficiently. What’s more, the comprehensive prediction models performed better thansingle-task prediction models.Finally, the original prediction models were improved. To reduce the dependenceon data sets of the activation function and hidden nodes in the original ELM algorithm,a hybrid activation based ELM (Hybrid-ELM) was proposed in this paper. Besides,genetic algorithm was applied for feature selection. Results showed that the improvedmodels were more quick and efficient. Using the proposed Hybrid-ELM model, muchbetter prediction results were achieved on all7data sets, thus left out the cumbersomeprocess of selecting proper parameters for each data set. Moreover, the selectedfeatures by genetic algorithm have more clear physical meanings.