Diagnosis and grading of spasticity with electrophysiological and kinesiological data

Spasticity is a clinical condition caused by movement disorder induced by lesion in the upper motor pathways. There is no generally accepted method in diagnosis, measurement and evaluation of spasticity. In this thesis, a reliable clinical evaluation method is developed for spasticity based on signal processing and machine learning techniques. For this purpose, a measurement system that records the EMG data from RF and BF muscles when Patella T-reflex was triggered is designed. The data that is recorded by the system is passed through the preprocessing step. In the following feature extraction step, five features in time domanin and frequency domain of the short time EMG signal that measured from aech of two muscle groups are determined. These features are combined with six features of the Pendulum movement triggered by the Patelle T-reflex to generate the feature vector that characterizes the spasticity. Feature vector that includes 26 elements is obtained eventually. After feature extraction step, Fisher Score is used in the feature selection process that is developed to improve the separation ability of the feature vector. Preprocessing, feature extraction and feature selection steps constitute the signal processing part of the thesis. ANN, k-NN and SVM machine learning algorithms, which produce acceptable and reliable results in the grading of spasticity, were applied in the classification phase of the system. Furthermore, in this thesis AdaBoost algorithm is utilized to improve the performance of the classifiers. Neurologists reported that 80% and above accuracy scores in clinical trials are absolutely acceptable and feasible. Thus, the proposed method that includes Fisher Score, 3-fold cross validation and SVM or k-NN techniques can effectively be used for spasticity assessments. It was determined that the accuracy values of these combinations which are used to determine the normal, Ashworth 1 and Ashworth 2 spasticity levels were 86.66% and 80.33%, respectively. The proposed thesis performs the grading of spasticity in an authomatic and reliable way by using signal processing and machine learning methods and thus it is an important study that provides valuable information to the neorulogists.