Automated Implementation and Validation of the Edinburgh Visual Gait Score (EVGS)

Abstract

Gait analysis is an integral component of physical and neurological status assessment in humans. The Edinburgh Visual Gait Score (EVGS) is a reliable and clinically feasible scoring system for visual gait analysis. Clinical use of EVGS relies on manual scoring of video recordings, which is a point where automation can be utilized to gain even higher efficiency and accuracy. In this thesis, an algorithmic implementation of EVGS scoring using patient videos was implemented and evaluated. Videos with EVGS manual scores were obtained from the Sanatorio del Norte medical center dataset, providing sagittal and coronal plane views of people with cerebral palsy walking. Body keypoints representing joints and limb segments were identified using the OpenPose Body 25 pose estimation model. The algorithm used these keypoints to identify foot events, strides, and relevant body angles, which were used by the algorithm to automatically score each EVGS parameter. The stride identification results were compared against the ground truth foot events and EVGS results were compared with expert scorer evaluations. The algorithm was excellent for plane detection and movement direction classification, in both sagittal and coronal views. Stride detection was accurate for the majority of videos. Of the 17 EVGS parameters evaluated, six had an accuracy of 90-94%, five had a high accuracy of 84-89%, three a moderate accuracy of 70-76%, while the final three had a low accuracy of 58-62%. The results support use of the automated EVGS scoring approach to accelerate clinical visual gait analysis so that routine monitoring of patients can be performed without requiring extensive clinician time and enabling remote EVGS analysis at the point of patient contact.