Video-based markerless assessment of bilateral upper limb motor activity following cervical spinal cord injury

Recent advancements in computer vision have positioned pose estimation as a powerful tool for analyzing human movement from video data. However, its application to neurological populations, particularly individuals with spinal cord injuries (SCI), remains largely unexplored.

To evaluate the effectiveness of three video-based pose estimation algorithms in detecting upper limb motor impairments in individuals with cervical SCI, and to determine their reliability and sensitivity in comparison with marker-based motion capture.

Twelve individuals with cervical SCI and twelve unimpaired controls performed four static arm poses from the Arm Stabilization Test (part of the Van Lieshout Arm/Hand Function Test). Three pose estimation methods - MediaPipe 2D, AlphaPose, and MediaPipe 3D - were applied to single-camera videos, and results were compared to kinematics captured by an eight-camera marker-based motion capture system. Motor performance was quantified using a compensation ratio (shoulder vs. elbow movement), and trial-by-trial repeatability was assessed.

All three algorithms demonstrated strong repeatability. MediaPipe 3D, in particular, showed a significant correlation between the compensation metric and impairment severity, effectively distinguishing SCI from control participants. In contrast, 2D methods lacked this sensitivity, likely due to the absence of depth information.

Markerless pose estimation, especially 3D approaches, offers a promising, non-invasive alternative for assessing upper limb motor function in neurological populations. These tools could expand accessibility to motor assessment in both clinical and research settings, although further research is needed to assess their sensitivity to subtle impairments.