Monitoring symptom progression in Parkinson’s Disease using Least Squares Support Vector Regression

Abstract

Parkinson’s Disease is a neurodegenerative disease that has a decline in motor behaviour as one of its main symptoms. This decline is currently monitored using subjective measures, such as questionnaires and clinical observations. More detailed and objective tracking of this decline can improve treatment of the disease and allow for earlier identification. Earlier work in this area has lead to the development of a proof-of-concept for detecting behavioural changes in motor performance, but the linear regression model used was limited in its accuracy and had a high latency. This paper uses a non-linear Least Squares Support Vector Regression (LS-SVR) model to increase performance in those areas. LS-SVR was chosen for its ability to regress complex non-linear relationships and because it is highly adaptable and scalable, yet easy to understand and work with. Test data were simulated based on earlier measured data which resulted in a complete and varied data set that allows for exploring many different situations.
Results from the machine learning model when applied to the test data are promising. With the right combination of hyperparameter settings an improvement of 80% in accuracy was reached, and latency could be reduced by 30%. Additionally, a sensitivity analysis of the hyperparameters revealed further room for improvement with more careful tuning. Finally, it was shown that by using LS-SVR to make predictions on data from future trials a further reduction in latency can be achieved. Overall, the new model shows definite improvement over the earlier model and can be developed further in subsequent steps towards a clinical applicable method.