Quantitative Pose-Based Analysis of Movement Disorders in Pediatric NGLY1 and SLC13A5 Patients

Movement disorders have long relied on subjective clinical observation for diagnosis and monitoring. By contrast, computer vision tools such as OpenPose can turn video recordings into precise, time-resolved measurements of a patient’s posture and movement. In this work, we apply a fully markerless, pose-based pipeline to classify abnormal movements in children with NGLY1 or SLC13A5 mutations. Our primary focus is on simple, physician-informed pose features that can be interpreted in clinical terms and used with conventional classifiers (Random Forest, SVM, etc.) on a very small dataset. We show that these handcrafted features capture clinically meaningful differences between movement-disorder phenotypes and can achieve useful classification performance. In addition, we include an exploratory comparison with a transformer model that is pre-trained on large-scale action-recognition data and then fine-tuned on our pose data. This experiment illustrates the potential performance ceiling of deep learning with extensive pretraining, but we emphasize that such models are less transparent and more data-hungry than the traditional approaches that form the core contribution of this study.