FV 23 POSe EstimatoR for Cervical Dystonia (POSER-CD): Automatized assessment of clinical severity and kinematic pathosignatures of Cervical Dystonia using convolutional neural networks

Abstract

Background: Phenomenological description, diagnosis and assessment of disease development, as well as treatment effects, of movement disorders still heavily relies on clinical assessment using scoring scales. However, these scales are inherently prone to rater biases, thereby negatively affecting their reliability and comparability. Objective: To leverage recent advances in computer vision and artificial neural networks to devise a video-based movement analysis, which allows the quantitative measurement of cervical dystonia during clinical scoring of severity using the Toronto Western Spasmodic Torticollis Rating Scale (“TWSTRS-S”) before and after pallidal deep brain stimulation (DBS). Methods: 303 videos documenting the longitudinal TWSTRS-S assessment of 94 individuals who had undergone pallidal deep brain stimulation in the context of three multicentric studies, in addition to 22 standardized videos of healthy subjects (HS) were available for analysis. Clinical videos were retrospectively scored by two independent movement disorder specialists using the TWSTRS-S. HS videos and a subset of clinical videos were used for supervised training of a convolutional neural network for movement classification (MC-CNN) to segment 7 movement states along three axes (pitch, yaw, tilt) mapping to TWSTRS subscore dimensions (head inclination/reclination, turning, tilting). A total of 2178 frames from the video dataset were annotated to train a recurrent CNN for tracking of anatomical landmark points in the videos (R-CNN). Subsequently, coordinate time series were transformed into an anatomically inspired, multidimensional kinematic feature space (i. e. angles, velocities, ratios) to characterize MC-CNN classified movement states relevant to the TWSTRS-S. Results: MC-CNN reached ∼90% accuracy on a held-out test data set. The pose tracking R-CNN achieved train/ test errors of 0.5/ 9.5 mm, corresponding to 1.67/ 7.36 pixels (mean Euclidean distance). Movement classification analysis identified abnormal coherence between face-forward and the rotation and tilt (both left and right) movement states respectively, that are reminiscent of dystonic overflow phenomena and decreased range of motion which could be significantly attenuated by DBS (p<.005 for all comparisons). Furthermore, DBS reversed an abnormal degree of asymmetry in movement representations in 2/3 planes (pitch: p=.009, yaw: p= .011, roll: p= .828). Conclusions: We provide evidence that computer-assisted movement analysis using CNNs is a novel avenue to overcome metrological challenges in the field of movement disorders. CNNs may help derive and analyze patient and/or diagnosis-specific kinematic pathosignatures which can be integrated with multidimensional neural data and may contribute to “precision Neurology”.