AI Cameras Catch Parkinson-Like Gait Changes When Dementia Drugs Are Cut

Some of the most consequential changes in a person with dementia never show up on a scan or a lab report. They show up in the way they walk. A new study published in PLOS Digital Health suggests that ordinary ceiling-mounted cameras, paired with machine learning, can detect those changes automatically—capturing subtle shifts in parkinsonian gait when residents in dementia care units have their psychotropic medications reduced or stopped.

The research, led by Benjamin Filtjens and colleagues with senior authorship from Babak Taati and Andrea Iaboni, tackles a stubborn clinical problem: drug-induced parkinsonism. Certain medications, antipsychotics chief among them, can produce movement symptoms that mimic Parkinson’s disease—slowed, shuffling steps, reduced arm swing, stiffness. The condition is potentially reversible, but only if clinicians notice it and act. In residential dementia care, where residents may not be able to report what they feel and staffing is stretched thin, those signs are easy to miss or to misattribute to dementia itself.

Clinical movement disorder scales such as the Unified Parkinson’s Disease Rating Scale, or UPDRS, are the traditional tool for grading these symptoms, but they depend on trained raters performing structured examinations. That makes them poorly suited for tracking day-to-day fluctuation in a long-term care setting. The study’s central idea was to move the assessment out of the clinic and into the hallway—continuously, unobtrusively, and without anyone wearing a sensor or being filmed up close in a way that identifies them.

The team ran a two-center cohort study involving 61 residents with dementia who were receiving psychotropic medication. Gait was recorded with ceiling-mounted cameras positioned in hallways the residents walked through as part of ordinary daily life. Computer vision algorithms converted the raw video into three-dimensional body pose—essentially a moving skeleton for each person—without storing identifiable imagery. From each recorded walk, a machine learning model called TabPFN, a tabular foundation model that has attracted attention for performing well on modest-sized clinical datasets, estimated a gait severity score corresponding to item 29 of the UPDRS-III, the scale’s gait examination item.

The researchers then asked a specific question: what happens to daily mean predicted gait severity when residents undergo an acute reduction or cessation of antipsychotic or antidepressant medication, defined as at least a 50 percent dose reduction or complete stop within zero to 21 days after the change? To answer it, they used linear mixed-effects models, a statistical framework well suited to repeated measurements nested within individuals, analyzing participants who had ever been exposed to the medications in question.

The results pointed in one direction for one drug class. Acute antipsychotic dose reduction or cessation was associated with lower predicted gait severity—a coefficient of −0.127 with a p-value of 0.022, a statistically significant association. In plain terms, when antipsychotics were cut back, the algorithm’s estimate of parkinsonian gait severity tended to fall. Antidepressant changes, by contrast, showed no such association: the coefficient was −0.060 with a p-value of 0.210, well short of conventional significance thresholds. The pattern is pharmacologically plausible, since antipsychotics—particularly first-generation agents—block dopamine receptors and are the classic culprits in drug-induced parkinsonism, while antidepressants are generally weaker offenders.

But the story is more complicated than a single headline number suggests. When the team ran a complementary analysis restricted to participants who had been observed in both medication states—on the full dose and after the reduction—they weighted the two states equally and looked for within-person change. Here, the participant-average effects vanished. For antipsychotics, the mean change was 0.031, with a 95 percent confidence interval spanning −0.114 to 0.176 and a p-value of 0.653. For antidepressants, the mean change was −0.021, with a confidence interval of −0.098 to 0.056 and a p-value of 0.575. Neither came close to statistical significance.

That discrepancy between the two analytical approaches is not a contradiction so much as a window into biological reality. The mixed-effects models, which borrow strength across the entire cohort, detected a cohort-level trend. The paired analyses, which ask whether the average individual improved, found nothing—consistent with highly heterogeneous individual responses. Some residents presumably improved markedly when their antipsychotics were reduced; others barely changed, or their gait worsened for unrelated reasons. Averaging across such a mixed picture can wash out a real signal even when the trend is genuine at the population level. This is a familiar tension in clinical research, and the authors are candid that their findings are associative rather than causal.

The technical achievement is nonetheless substantial. Estimating a standardized clinical rating from passive video, in a real care facility, at scale, and then tracking its longitudinal movement around medication changes is exactly the kind of unobtrusive monitoring that long-term care has lacked. The pipeline—ceiling cameras, pose estimation, a classifier producing a UPDRS-III item 29 equivalent for each walk—could in principle run continuously, flagging when a resident’s gait deteriorates after a dose increase or improves after a taper, without a single scheduled clinic visit. For a potentially reversible condition like drug-induced parkinsonism, that kind of early, objective feedback could matter enormously.

The authors are equally clear about the limits. The study did not systematically document the reasons for medication changes, so it cannot rule out confounding—clinicians may have reduced doses precisely because they noticed movement problems, which would make the association partly circular. The cohort was modest, the observation windows were short, and the findings require replication. Larger prospective studies with documented reasons for medication changes are needed before such monitoring could guide actual medication decisions. Still, the study offers a glimpse of where dementia care may be heading: hallways that watch, algorithms that grade, and a continuous, invisible record of how the body responds when the chemistry changes.

Subject of Research: Vision-based machine learning assessment of drug-induced parkinsonism in daily-life gait during psychotropic medication changes in residential dementia care

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

Article References: Filtjens, B., Adeli, V., Mehraban, S., Sabo, A., Taati, B., & Iaboni, A. (2026). Assessment of change in drug-induced parkinsonism in daily-life gait using vision-based machine learning. PLOS Digital Health, 5(10), e0001780. https://doi.org/10.1371/journal.pdig.0001780

Image Credits: AI Generated

DOI: 10.1371/journal.pdig.0001780

Keywords: drug-induced parkinsonism, dementia care, gait analysis, machine learning, computer vision, antipsychotics, psychotropic medication, UPDRS, long-term care, PLOS Digital Health, TabPFN, pose estimation