Assessment of Parkinson gait through digital signal processing and machine learning

It would be of both patients’ as well as clinicians’ interest, if diagnosis of Parkinson’s disease (PD) as well as following check-up methods were perfectly sensitive, accurate, reproducible and feasible of objectively classifying motor symptoms of PD. This is an arduous task due to the possible subjectivity of clinical evaluations. In the past decade, attention turns into a multitude of technology based measures (TBMs) to address this need, among which the method of this research is positioned. Author hopes to contribute with a motor assessment method that addresses not only the issue of subjectivity of measurement, but also does not require extensive installments and is easy to use. For this study, data from a clinical trial conducted at Uppsala University Hospital, Sweden in 2015 are used. 7 PD patients and 7 healthy controls each performed 7-13 times each the same motoric gait test, which has been was video recorded. These recordings were showed to clinicians, who rated subjects’ gait and possible dyskinesia on the unified Parkinson's disease rating scale (0-4 rating). Thus the aim of this research was to imitate and automate the tasks of clinicians when diagnosing PD and its symptoms through motoric ratings, using various gait features. These gait features were obtained through quantification of signals from different body parts while patient performs walking motoric test, using image processing. Diagnosis of PD and its symptoms was twofold, as to firstly identify whether the subject has PD and to secondly predict the severity of PD patients symptoms. When classifying subjects into healthy controls and PD patients, classification trees and support vector machines have been deployed, while these achieved 76- 85% accuracy depending on features selected. Following focus was to diagnose severity of PD among patients, while using UPDRS ratings by clinicians as a target variable for supervised learning. Herein, linear regression has been deployed, while average absolute prediction error was 0.25 and correlation of UPDRS ratings with predicted values was 0.84