Tell me EXACTLY how any of this gets survivors to 100% recovery.
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.
Introduction
More than 795 000 individuals have a stroke each year in the United States.1 Based on data from multiple cohorts,2-6 up to 77% will have persistent upper-extremity (UE) impairment. Yet the development of markedly effective treatments has been slow7,8 and uncertain.9,10 An important reason for the lack of efficacy findings in clinical trials is the lack of a direct measure of the outcome of interest: productive functional use of the UE in everyday life. Without such a measure, clinicians and clinical trialists simply cannot know for certain if their treatments worked. Current approaches settle for use of proxies: self-report of UE use11,12 (with its attendant biases) and motor performance measures in the laboratory and clinic13,14 (with uncertain correspondence to everyday life). Thus, there is a clinical and research need for a quantitative, objective, and psychometrically sound measure of UE productive use in the community.
There is now a large body of work on wearable sensors that can track UE activity, with the majority of studies using wrist-worn accelerometers and the “counts threshold method” for detecting movement.15-18 This method estimates the total amount of time the UE is in motion. Several studies have reported that the counts threshold method often correlates significantly with other clinical scales.19-22 However, the correlations reported are often weak. For example, the largest sample was collected as part of the EXCITE clinical trial (n = 169), and the reported r value was only 0.52 between accelerometry and the Motor Activity Log (MAL).23 Responsiveness to change is mixed: some studies report changes in accelerometry that parallel change in clinical scores,24-29 whereas other studies have reported no changes in accelerometer-based metrics even as clinical scores improved.30-33
There are significant differences in what is measured by self-report scales of functional use (MAL) and acwcelerometry.34 Accelerometry measures both functional activity (ie, reaching to grasp, gesturing, balancing functions) and nonfunctional movements associated with gait and whole body movements. A bus ride will lead to many instances of acceleration not associated with functional movement. The effects of nonfunctional movements are thought to be eliminated by normalizing duration of movement in the paretic limb by duration in the less-affected limb, referred to as the counts ratio. However, there have been no prior studies to confirm that this normalization is effective. In fact, there have been surprisingly few attempts to quantify if any of the accelerometry based metrics actually measure functional use accurately. Early attempts at validation focused only on confirming that the duration of movement via the thresholding method correlates with video annotation35 or examined the correlation between duration of movement within 15-minute blocks and a video-based metric that combined the amount of functional and nonfunctional movements into a single average score.36
We have developed machine learning algorithms for directly measuring amount of functional movement using accelerometry. In previous work, we determined the accuracy of a single machine learning algorithm during a session of unstructured activities using ground truth from a video-based method that provides frame-by-frame scoring of the activities.37 In this study, using the same data set, we investigated several alternative machine learning algorithms to find generalizable models for our application and have improved accuracies compared with previous reports. We have extended the analysis of this data set by directly comparing the counts threshold method and our machine learning algorithms against ground truth from video annotation. Furthermore, we also analyzed data collected from the dominant limb of controls and the less-affected limb of stroke participants (not analyzed previously), so that ratio metrics could be studied. Most important for neurorehabilitation, we present the first study to separate activity measured by the counts threshold method into functional and nonfunctional categories and present machine learning algorithms to automatically identify periods of functional movement from accelerometer data.
No comments:
Post a Comment