Movement Based Classification of People with Stroke Through Automated Analysis of Three-Dimensional Motion Data

Abstract

Active recovery of motor function following a stroke requires intense and repetitive physical therapy. The effectiveness of such therapy can be greatly improved if it is fully customized for each patient. Motion tracking and machine learning algorithms can assist therapists in designing the therapy regimens, thereby saving valuable time. In this study, three-dimensional upper body movements both of people who had suffered a stroke and of healthy subjects were recorded as they performed a reaching task. A support vector machine with a five-dimensional feature space was used to automatically distinguish between the movements of people with stroke and those of healthy subjects. The success rate for this task peaked at over 95%. While this specific task is trivial for a clinician, it provides proof of concept, and a foundation for further work in developing classifiers that can locate more specific problems. Such a classifier may help clinicians treat the root cause of a complicated movement deficiency with a patient specific rehabilitation program. The results indicate that using machine learning approaches for analyzing stroke patient movement data shows potential for reducing clinician’s workloads while providing improved treatment to specific subjects.