Estimating Walking Speed in the Wild

Abstract

An individual's physical activity substantially impacts the potential for prevention and recovery from diverse health issues, including cardiovascular diseases. Precise quantification of a patient's level of day-to-day physical activity, which can be characterized by the type, intensity, and duration of movement, is crucial for clinicians. Walking is a primary and fundamental physical activity for most individuals. Walking speed has been shown to correlate with various heart pathologies and overall function. As such, it is often used as a metric to assess health performance. A range of clinical walking tests exist to evaluate gait and inform clinical decision-making. However, these assessments are often short, provide qualitative movement assessments, and are performed in a clinical setting that is not representative of the real-world. Technological advancements in wearable sensing and associated algorithms enable new opportunities to complement in-clinic evaluations of movement during free-living. However, the use of wearable devices to inform clinical decisions presents several challenges, including lack of subject compliance and limited sensor battery life. To bridge the gap between free-living and clinical environments, we propose an approach in which we utilize different wearable sensors at different temporal scales and resolutions. Here, we present a method to accurately estimate gait speed in the free-living environment from a low-power, lightweight accelerometer-based bio-logging tag secured on the thigh. We use high-resolution measurements of gait kinematics to build subject-specific data-driven models to accurately map stride frequencies extracted from the bio-logging system to stride speeds. The model-based estimates of stride speed were evaluated using a long outdoor walk and compared to stride parameters calculated from a foot-worn inertial measurement unit using the zero-velocity update algorithm. The proposed method presents an average concordance correlation coefficient of 0.80 for all subjects, and 97% of the error is within ±0.2m· s−1. The approach presented here provides promising results that can enable clinicians to complement their existing assessments of activity level and fitness with measurements of movement duration and intensity (walking speed) extracted at a week time scale and in the patients' free-living environment.