Monitoring physical function in patients with knee osteoarthritis using data from wearable activity monitors

TL;DR

This study develops a method to objectively monitor physical function in knee osteoarthritis patients using wearable activity data, accurately predicting clinical performance measures and enabling remote assessment over time.

Contribution

The paper introduces a novel approach to represent physical function as daily activity profiles from wearable data, linking them to clinical performance measures for remote monitoring.

Findings

Activity profiles predict clinical physical performance with AUCs up to 0.79.

Models maintain predictive accuracy on data two years into the future.

Remote monitoring of physical function is feasible using wearable activity data.

Abstract

Currently used clinical assessments for physical function do not objectively quantify daily activities in routine living. Wearable activity monitors enable objective measurement of routine daily activities, but do not map to clinically measured physical performance measures. We represent physical function as a daily activity profile derived from minute-level activity data obtained via a wearable activity monitor. We construct daily activity profiles representing average time spent in a set of activity classes over consecutive days using the Osteoarthritis Initiative (OAI) data. Using the daily activity profile as input, we trained statistical models that classify subjects into quartiles of objective measurements of physical function as measured via the 400m walk test, the 20m walk test and 5 times sit stand test. We evaluated model performance on held out data from the same calendar year as that used to train the models as well as on activity data two years into the future. The daily activity profile predicts physical performance as measured via clinical assessments. Using held out data, the AUC obtained in classifying performance values in the 1st quartile was 0.79, 0.78 and 0.72, for the 400m walk, the 20m walk and 5 times sit stand tests. For classifying performance values in the 4th quartile, the AUC obtained was 0.77, 0.66 and 0.73 respectively. Evaluated on data from two years into the future, for the 20m pace test and the 5 times sit stand tests, the highest AUC obtained was 0.77 and 0.68 for the 1st quartile and 0.75 and 0.70 for the 4th quartile respectively. We can construct activity profiles representing actual physical function as demonstrated by the relationship between the activity profiles and the clinically measured physical performance measures. Measurement of physical performance via the activity profile as described can enable remote functional monitoring of patients.