Automated Video-based Characterization of Movement Quality in a Phase III Clinical Trial of Troriluzole in Subjects with Spinocerebellar Ataxia
Melissa Beiner1, Evangelos Oikonomou2, Gilbert L'Italien1, Rohan Khera2, Michele Potashman1, Jeremy Schmahmann3, Susan Perlman4, Vladimir Coric1
1Biohaven Pharmaceuticals, Inc., 2Yale School of Medicine, 3Massachusettes General Hospital, 4UCLA School of Medicine
Objective:
To compare the effects of troriluzole versus placebo on gait quality among subjects with spinocerebellar ataxia (SCA) through video-based assessment of gait dispersion.
Background:
Traditional methods for assessing disease progression often rely on subjective scales and clinical expertise. This brings challenges for assessing patients with SCA, where changes in mobility can be subtle and varied. We hypothesized that a machine learning analytic system may provide a novel perspective, complementing traditional clinician-rated measures of gait.
Design/Methods:
This study analyzed videos from participants with SCA who underwent gait assessment in a phase III double-blind placebo-controlled trial of troriluzole (NCT03701399). Videos were analyzed using a deep neural network architecture that tracked coordinates of key body parts and was used to define interpretable kinematic features of dynamic pose dispersion, through the integrated Pose Dispersion Index (PDI) that tracks symmetry, balance, and stability during normal and tandem walk tasks. The effects of troriluzole on PDI were assessed in mixed linear models using fixed effects, subject-level grouping, and ‘treatment group’-by-‘visit week’ interaction, adjusted for age, sex, baseline f-SARA, and time since diagnosis.
Results:
From 218 randomized subjects, 67 and 56 subjects had interpretable videos of a tandem and normal walk attempt, respectively. Over 48-weeks, individuals assigned to troriluzole exhibited significant improvement in tandem walk PDI versus placebo (adj. interaction coefficient: 0.58 (95% CI: 0.14 to 1.03)). A similar but not statistically significant trend was observed in the normal walk assessment (coef. 1.20 (95%CI: -1.07 to 3.46)). Lower baseline PDI during normal walk was associated with a higher risk of subsequent falls (adj. Poisson coef. -0.36 [95% CI -0.70 to -0.01]).
Conclusions:
Machine learning applied to video-captured gait parameters introduces a quantitative approach to motor assessment in cerebellar ataxia. The metrics captured in the PDI show that troriluzole-treated individuals with SCA demonstrated improvement in cerebellar motor function compared to controls.