Computer vision-derived gait-state tracking of acute treatment response in Parkinson’s disease: external and prospective validation

Abstract Continuous, treatment-responsive measurement of motor state is an unmet need in Parkinson’s disease and a broad challenge for digital biomarkers: clinical gait scales are coarse and ordinal, and most digital measures are never validated beyond their derivation setting. We present a digital biomarker derived from full-mesh 3D human-pose reconstruction in standard monocular video, built from high-contrast within-patient treatment transitions and applied to out-of-sample patients. In a retrospective cohort (53 patients) it discriminated treatment states (cross-validated AUROC 0.88) and defined a continuous motor-state axis. Across six independent cohorts (266 subjects, 3,591 walks, eight sites, three continents) it showed consistent medication effects in the two medication-labelled cohorts and tracked clinical gait impairment in the four cohorts carrying clinical gait ratings; in a prospective cohort (21 patients) it preserved ordered transitions for medication and deep brain stimulation. On identical paired observations it was more responsive than the ordinal MDS-UPDRS Part III item 3.10 (standardized response mean 1.29 versus 0.83 internally, 0.92 versus 0.54 externally); its finer resolution also registered directional within-patient improvement more often, an advantage of measurement resolution that reaches parity once change is required to exceed the minimal detectable change, supporting a scalable, generalizable measurement tool.