Automated Bradykinesia Assessment with Custom-designed Features from Pose Estimation Algorithms in a Computer Vision Pipeline

Wolters, Niels

Automated Bradykinesia Assessment with Custom-designed Features from Pose Estimation Algorithms in a Computer Vision Pipeline

Files

Wolters, N. s-4633873-MSc-MKI94-Thesis-2023.pdf (9.55 MB)

Authors

Wolters, Niels

Issue Date

2023-11-03

Language

en

URI

https://theses.ubn.ru.nl/handle/123456789/16429

Abstract

Parkinson’s disease (PD) is a worldwide burden; out of all neurodegenerative diseases, the number of deaths and disabilities caused by PD grows the fastest. Regular examinations of the progression of symptoms must be performed to determine treatment. One of the most widely used assessment frameworks is the Movement Disorder Society-sponsored revision of the Unified Parkinson’s Disease Rating Scale (MDSUPDRS). However, a low inter-examiner consistency in MDS-UPDRS assessment exists due to the subjective nature of the assessment criteria and large differences in examiner experience level. In addition, there is a deficit of trained neurologists, especially in low- and middle-income countries, and frequent visits to a hospital are strenuous for PD patients. Computer vision allows for the objective analysis of videos, allowing neurologists to work more efficiently and effectively by automatically assessing the task performance of patients. However, end-to-end machine learning methods with self-learned features often lack transparency. In this research, we propose a three-stage computer vision pipeline to automatically score videos of PD patients performing the Leg Agility (LA) test of the MDS-UPDRS with custom-defined features to get more insight into the automated scoring process. We outline experiments in which we compare two pose estimators and test 32 features and six classifiers. With our experiments, we find that a 61% accuracy can be obtained with our features. While this is lower than the reported state-of-the-art accuracy (70%), we obtained it through custom-defined features, with simple classifiers and on a less heterogeneous and smaller dataset compared to related studies. These results show the potential of the use of custom-defined features for automated LA test assessment, providing a basis for an explainable solution.