Assessment of change in drug-induced parkinsonism in daily-life gait using vision-based machine learning

Abstract

Drug-induced parkinsonism is a potentially reversible movement disorder that may be subtle and difficult to monitor in residential dementia care. Vision-based machine learning analysis of gait may provide unobtrusive longitudinal monitoring around medication changes. We conducted a two-center cohort study of 61 residents with dementia receiving psychotropic medication. Hallway gait sequences were recorded using ceiling-mounted cameras and converted to three-dimensional body pose. A TabPFN classifier estimated UPDRS-III item 29 gait severity for each walk. We evaluated associations between acute antipsychotic (AP) and antidepressant (AD) dose reduction/cessation ( ≥ 50% dose reduction or cessation, 0–21 days after change) and daily mean predicted gait severity using linear mixed-effects models in ever-exposed participants, with complementary equal-weight paired analyses among participants observed in both medication states. In mixed-effects models, acute AP dose reduction/cessation was associated with lower predicted gait severity ( β = − 0.127 , p = 0.022), whereas AD dose reduction/cessation was not ( β = − 0.060 , p = 0.210). Equal-weight paired analyses found no statistically significant participant-average within-person change among paired participants (AP mean change = 0.031, 95% CI [ − 0.114 , 0.176], p = 0.653; AD mean change = − 0.021 , 95% CI [ − 0.098 , 0.056], p = 0.575), consistent with heterogeneous individual responses. Vision-based monitoring captured within-participant gait variation around psychotropic dose-reduction periods, but findings are associative. Larger prospective studies with documented reasons for medication changes are needed before using such monitoring to guide medication decisions.