Bridge-Disease Curriculum Transfer Enables Smartphone-Based Balance Assessment in Rare Neurologic Disorders

Video-based motor assessment could enable scalable assessment of neurological function, but rare-disease applications are constrained by extreme data scarcity and poor generalization from small cohorts. We develop and evaluate a smartphone-based balance-assessment framework in NF2-related schwannomatosis (NF2-SWN), an ultra-rare disorder for which objective balance assessment is clinically valuable yet logistically challenging. At its core, a staged bridge-disease curriculum adapts a skeleton-based spatiotemporal graph model from large-scale action recognition through an intermediate Parkinson's disease domain to the target NF2-SWN cohort ( ${N}={19}$ ), decomposing one large domain shift into two more learnable transitions under leakage-proof, subject-independent evaluation. Three clinically motivated modules, bilateral asymmetry gating, temporal phase attention, and anatomical group saliency, yield interpretable, clinician-facing outputs aligned with balance assessment. Relative to direct transfer, bridge-disease transfer reduces patient-level prediction error by 40% on the Mini-BESTest total score, narrows the gap toward inter-rater agreement, and yields well-calibrated fall-risk stratification. A robustness analysis quantifies sensitivity to simulated acquisition variability, supporting a pragmatic smartphone recording protocol. Together, these results support the feasibility of privacy-preserving video-based balance assessment in NF2-SWN under extreme data scarcity.