An Automated Digital Biomarker of Mobility
Abstract
The Timed Up and Go (TUG) test is a common clinical endpoint whose usefulness is limited by the need to perform it in the presence of a trained evaluator, usually a clinician. The objective of this study was to propose and validate a sensor-agnostic automated pipeline based on a machine learning to predict TUG score using day-to-day walks captured by commonly used wearable sensors, generating a passive and continual stream of mobility biomarkers without the need of conducting scripted TUG tests. This pipeline is then validated against data from 303 subjects in three cohort datasets, each with a different primary focus population of healthy elderly adults, Parkinson’s disease patients, or dementia patients. In addition to TUG scores, the three datasets include walking data collected by different wearable sensors, i.e., a lower-back-worn accelerometer, wrist-worn accelerometer, or in-sole pressure gait sensor, respectively. Our leave-one-subject-out validation using subjects from all cohorts shows the random forest based predictive model is able to achieve an accuracy of 1.7 ± 1.7 seconds (Mean absolute error ± standard deviation) and 84.8% predictions within the minimal detectable change (± 3 seconds), with reasonable generalization across cohorts. Through the validation on data collected using three types of commonly used wearables, we demonstrate the ability of our proposed pipeline to leverage heterogeneous inputs for predicting TUG scores from walking data, suggesting the feasibility to generate a continual stream of TUG estimations as a novel digital biomarker of mobility by leveraging naturally occurring walks in free-living scenario. Our investigation also suggests that, for certain cohorts (e.g., Parkinson’s disease population), applying a cohort-specific model instead of using model trained with mixed cohorts might further improve performance