Wearable sensors to predict improvement following an exercise intervention in patients with knee osteoarthritis

Abstract

BackgroundMuscle strengthening exercises consistently demonstrate improvements in the pain and function of adults with knee osteoarthritis, but individual response rates can vary greatly. Identifying individuals who are more likely to respond is important in developing more efficient rehabilitation programs for knee osteoarthritis. Therefore, the purpose of this study was to determine if pre-intervention multi-sensor accelerometer data (e.g., back, thigh, shank, foot accelerometers) and patient reported outcome measures (e.g., pain, symptoms, function, quality of life) can retrospectively predict post-intervention response to a 6-week hip strengthening exercise intervention in a knee OA cohort.MethodsThirty-nine adults with knee osteoarthritis completed a 6-week hip strengthening exercise intervention and were sub-grouped as Non-Responders, Low-Responders, or High-Responders following the intervention based on their change in patient reported outcome measures. Pre-intervention multi-sensor accelerometer data recorded at the back, thigh, shank, and foot and Knee Injury and Osteoarthritis Outcome Score subscale data were used as potential predictors of response in a discriminant analysis of principal components.ResultsThe thigh was the single best placement for classifying responder sub-groups (74.4%). Overall, the best combination of sensors was the back, thigh, and shank (81.7%), but a simplified two sensor solution using the back and thigh was not significantly different (80.0%; p = 0.27).ConclusionsWhile three sensors were best able to identify responders, a simplified two sensor array at the back and thigh may be the most ideal configuration to provide clinicians with an efficient and relatively unobtrusive way to use to optimize treatment.Electronic supplementary materialThe online version of this article (10.1186/s12984-017-0309-z) contains supplementary material, which is available to authorized users.

Wearable Sensor Data to Track Subject-Specific Movement Patterns Related to Clinical Outcomes Using a Machine Learning Approach

“…A subset of 8 knee OA patients (Sex: 4F/4M, Age: 58 (5) years, Body Mass Index: 25.3 (4.8) kg/m 2 , walking speed: 1.1 (0.15) m/s) were analyzed from a larger exercise intervention [ 11 ]. These patients were selected for the current analysis as they completed two baseline gait trials before the intervention, as well as one gait trial post-intervention.…”

“…All participants were required to be radiographically diagnosed with knee OA and able to walk without assistive devices. For additional inclusion and exclusion criteria see Kobsar et al, 2017 [ 11 ]. This study was approved by the Conjoint Health Research Ethics Board at the University of Calgary (E-22417: Approved 14 June 2014) and all participants provided written, informed consent prior to participating.…”

“…Participants also completed a Knee injury and Osteoarthritis Outcome Score (KOOS) before and after the intervention to assess changes, if any, in self-reported pain, symptoms, function in daily living and knee related quality of life following the intervention [ 24 ]. The exercise intervention itself was a 6-week therapist-directed, hip-focused muscle strengthening intervention detailed in Kobsar et al, 2017 [ 11 ].…”

“…To do so, data from both baseline gait trials were combined, resulting in an average of 270 gait cycles collected over 5 min of walking data. These data were then standardized to a mean of 0 and a standard deviation of 1, before being transformed into linearly uncorrelated PCs using a principal component analysis [ 11 , 29 ]. The PCs that explained at least 95% of the total variance in the original data were selected as features for the algorithm [ 30 ].…”

“…Machine learning involves the integration of statistics and computer science to identify patterns in large datasets [ 10 ]. These pattern recognition tools provide the opportunity to quickly process and compare large wearable sensor datasets between or within different clinical populations or subgroups [ 11 , 12 , 13 , 14 , 15 ]. While group-based models can be effective in identifying clinically relevant differences in gait between individuals, there remain a number of important limitations.…”