Open Collections

Monitoring stroke rehabilitation of arm movement outside of the clinical setting

University of British Columbia

Stroke is the leading cause of disability in North America. Fifty-four percent of stroke survivors suffer from upper body hemiparesis, a weakness that limits the client’s ability to perform functional tasks with the affected side of the body. Stroke rehabilitation aims to recover limb mobility through thousands of repeated functional movements that lead to neural regeneration. However, time constraints in clinical rehabilitation lead to an average of 32 arm repetitions per session, which is insufficient for optimal recovery. Accurate monitoring of client activity outside of the clinical setting could enable therapists to track what they do, improving recovery. To address this problem, we have designed the Arm Rehabilitation Monitor (ARM), a wrist-worn device that collects movement data in unconstrained environments, and processes it offline to identify reach actions. Reach actions were identified as functionally meaningful tasks that lead to better rehabilitation. We enrolled 15 participants with mild to moderate hemiparesis due to stroke to perform two activities: (1) a functional assessment of the arm, and (2) an activity of daily living (ADL) task that consisted of making a pizza. The data recorded by the IMU on both activities was used to train three different machine learning algorithms (Random Forest, Convolutional Neural Networks and Shapelets) to detect reaching gestures. We found that the ARM obtained the best results with the Random Forest and CNN algorithms. The CNN algorithm had the best F1-score (0.523) for the Clinic-Home inter-subject tests, while the RF algorithm obtained the best score (0.486) in the Clinic-Home intra-subject configuration. We used the ARM to estimate the time spent reaching and the number of reach counts. The CNN algorithm predicted the reach time for the Clinic-Home inter-subject tests to be 1.07x ( 0.55x) the true reach time and the reach counts to be 1.28x ( 0.40x) the true number of reach gestures. In turn, the RF algorithm predicted the reach time for the Clinic-Home intra-subject configuration to be 1.16x ( 0.84x) and the reach counts to be 1.26x (0.40x). Both results have a smaller standard deviation when estimating reach counts than a comparable commercial accelerometer worn on the wrist.

Text

2018-10-18

Attribution-NonCommercial-NoDerivatives 4.0 International

http://hdl.handle.net/2429/67604

Biomedical Engineering

University of British Columbia

UBCV

http://creativecommons.org/licenses/by-nc-nd/4.0/