Upper extremity assessment and rehabilitation system for stroke patients

Abstract

Stroke is the leading cause of disabilities worldwide. Upper extremity impairments are very common after stroke. To support the recovery process, conventional assessment methods such as Fugl-Meyer Assessment (FMA) and Motor Assessment Scale (MAS) are widely used to assess motor performance of stroke patients. However, the assessments face some limitations such as being subjective and time-consuming. Many research have been done to solve the limitations of conventional assessments by using motion capture sensor or robotics for objective assessment. The main objective of this research is to design and develop a vision-based automated rehabilitation and assessment system to assess upper extremity of stroke patients. A Kinect-based system was used as an upper extremity stroke rehabilitation assessment system with isolated training movement namely Shoulder Abduction-Adduction (SAA). Three experiments were conducted involving a total of eight healthy subjects and three stroke patients. A total of six out of nine collected features have been proved being significantly different using t-test method. The suitable features were selected using three different features selection methods, namely Relief-F, Principal Analysis Component, and Correlation-based Feature Selection. These three feature sets were then trained with four different classifiers: Artificial Neural Network (ANN), Support Vector Machine (SVM), Random Tree and Random Forests in order to achieve the best predictive model. With a total of three feature sets and four classifiers, a total of 12 predictive models were constructed in this thesis. The 12 models were evaluated based on correlation-analysis. The result shows that the combination of ReliefF and SVM achieved accuracy of 91.04%, highest correlation coefficient of 0.9929 and lowest root mean square error of 0.1183 among all the constructed models.