Gait feature extraction

Three separate methods were used for estimating gait speed during the performance of a gait task in the laboratory. Ground truth gait speed was estimated from data collected using an instrumented mat (GAITRite), using a vendor supplied proprietary algorithm (GAITRite Software version 4.8.5). In addition, six inertial sensors (Opal, APDM) located on the sternum, lower back, and bilaterally on the wrists and feet, were used to estimate gait speed using a vendor supplied proprietary algorithm (APDM Mobility Lab v2.0.0.2018). Lastly, we estimated gait speed from accelerometer data recorded using a lumbar-mounted device (Opal, APDM), using an open-source algorithm (GaitPy v1.6.0), we implemented in Python v3.6³⁵. GaitPy uses a wavelet-based method to enhance patterns that occur in the vertical acceleration signal for first detecting heel strike and toe off events during a gait cycle⁴³. Gait speed is then estimated by integrating the vertical acceleration signal to derive vertical displacement and applying an inverted pendulum model as described by Zijlstra et al.⁴⁷.

GaitPy was also used to estimate gait speed from data collected at home. For at-home data, GaitPy first uses a binary classifier to detect bouts of gait. Bouts of gait <3 s apart are concatenated into a single bout before estimating gait speed on a stride by stride basis.