Improving Accelerometry-Based Measurement of Functional Use of the Upper Extremity After Stroke: Machine Learning Versus Counts Threshold Method

Abstract

Background Wrist-worn accelerometry provides objective monitoring of upper-extremity functional use, such as reaching tasks, but also detects nonfunctional movements, leading to ambiguity in monitoring results. Objective Compare machine learning algorithms with standard methods (counts ratio) to improve accuracy in detecting functional activity. Methods Healthy controls and individuals with stroke performed unstructured tasks in a simulated community environment (Test duration = 26 ± 8 minutes) while accelerometry and video were synchronously recorded. Human annotators scored each frame of the video as being functional or nonfunctional activity, providing ground truth. Several machine learning algorithms were developed to separate functional from nonfunctional activity in the accelerometer data. We also calculated the counts ratio, which uses a thresholding scheme to calculate the duration of activity in the paretic limb normalized by the less-affected limb. Results The counts ratio was not significantly correlated with ground truth and had large errors ( r = 0.48; P = .16; average error = 52.7%) because of high levels of nonfunctional movement in the paretic limb. Counts did not increase with increased functional movement. The best-performing intrasubject machine learning algorithm had an accuracy of 92.6% in the paretic limb of stroke patients, and the correlation with ground truth was r = 0.99 ( P < .001; average error = 3.9%). The best intersubject model had an accuracy of 74.2% and a correlation of r =0.81 ( P = .005; average error = 5.2%) with ground truth. Conclusions In our sample, the counts ratio did not accurately reflect functional activity. Machine learning algorithms were more accurate, and future work should focus on the development of a clinical tool.