AI-based Digital Assessment of Gross Motor Function in Metachromatic Leukodystrophy

Gross motor function in metachromatic leukodystrophy (MLD) is routinely assessed using the GMFC-MLD scale, but expert video-based scoring is resource-intensive and difficult to scale across sites and time. We propose a privacy-preserving, skeleton-based framework to estimate visit-level GMFC-MLD scores from long, heterogeneous motor-assessment videos. Using recordings from the multi-site EMBOLDEN phase-2 trial (NCT03771898), we curate MELD-2D comprising 44 children and 303 rated visits, and convert restricted RGB videos into de-identified 2D pose sequences. To handle weak visit-level supervision and within-visit variability, we formulate scoring as ordinal multiple instance learning with an ST-GCN backbone and a clinician-aligned ordinal noisy-OR pooling operator that captures an "at-least-one-good-clip" decision rule. On visits with full rater agreement, the proposed model achieves 82.1% accuracy and 0.806 quadratic weighted kappa against adjudicated labels, with improved class-balanced performance over clip-supervised baselines. The model runs in real time on 2D skeletons and supports streaming inference, enabling potential ondevice deployment without transmitting identifiable video. De-identified data and code are released for reproducibility.