Automated objective dystonia identification using smartphone-quality gait videos acquired in clinic

Background: Dystonia diagnosis is subjective and often difficult, particularly when co-morbid with spasticity as occurs in cerebral palsy. Objective: To develop an objective clinical screening method for dystonia Methods: We analyzed 30 gait videos (640x360 pixel resolution, 30 frames/second) of subjects with spastic cerebral palsy acquired during routine clinic visits. Dystonia was identified by consensus of three movement disorders specialists (15 videos with and 15 without dystonia). Limb position was calculated using deep neural network-guided pose estimation (DeepLabCut) to determine inter-knee distance variance, foot angle variance, and median foot angle difference between limbs. Results: All gait variables were significant predictors of dystonia. An inter-knee distance variance greater than 14 pixels together with a median foot angle difference greater than 10 degrees yielded 93% sensitivity and specificity for dystonia. Conclusions: Open-source automated video gait analysis can identify features of expert-identified dystonia. Methods like this could help clinically screen for dystonia.