Use of Computer Vision to Estimate Spatiotemporal Gait Measures From Consumer Grade Videos in People With Multiple Sclerosis
Megan McCune1, Yoni Ackerman1, Nikki Sisodia1, Kyra Henderson1, Jaeleene Wijangco1, Shane Poole1, Abel Torres Espin2, Matthew Miller1, Valerie Block1, Riley Bove1
1University of California San Francisco, 2University of Waterloo
Objective:

To assess the validity of an algorithm that estimates spatiotemporal gait parameters in people with MS (PwMS) using only a single consumer-grade video.

Background:

Gait analysis in PwMS typically requires expensive tools, such as motion capture or pressure mats. In contrast, pose estimation software provides a low-cost, accessible method to collect kinematic data. 

Design/Methods:

153 adult PwMS were recorded with a single consumer-grade camera in the frontal plane while walking as fast as safely possible; among these, videos from a second timepoint were available for 73 participants. Videos were processed using MediaPipe Pose pose estimation software. A custom-written algorithm then calculated the following gait parameters: an estimate of walking speed, stride time, cadence, and stride width. Additionally, Timed-25 Foot Walk (T25FW) was collected.

Linear mixed-effect models were used to assess the cross-sectional association of FW video-derived gait parameters with log-transformed T25FW. Descriptive statistics were used to evaluate the longitudinal validity of video-derived parameters compared to T25FW.
Results:

In a multivariate model, older age (Estimate 0.0048 [0.0019, 0.0078], p=0.002), slower walking speed estimate (-0.26 [-0.33, -0.19], p<0.001), and longer stride time (0.97 [0.75, 1.2], p<0.001) were associated with slower walking by T25FW. These three variables predicted a majority of variance in log-transformed T25FW (conditional R2 0.79, marginal R2 0.63).

In preliminary longitudinal analysis, when evaluated categorically (worsening defined as: ≥ 20% increase in T25FW and worsening in either video-derived parameter ≥ smallest real difference), the change in these two video-derived parameters agreed with the change in T25FW in 83% of participants.

Conclusions:
This study demonstrates the initial validity of pose estimation to extract valid, longitudinally informative gait parameters from consumer-grade videos in PwMS. Computer vision shows promise as an accessible, low-cost evaluation of patient gait without the requirement of high-end equipment or trained personnel.
10.1212/WNL.0000000000217408
Disclaimer: Abstracts were not reviewed by Neurology® and do not reflect the views of Neurology® editors or staff.