AR/ML Tool Flagged All Confirmed PD-MCI Cases in 21-Person Pilot

A small feasibility study paired a 10-minute augmented reality and machine learning tablet test with clinician assessments in Parkinson's disease, catching every case of mild cognitive impairment that specialists confirmed but also flagging nearly half the group who did not have it.

A pilot study published in Frontiers in Aging Neuroscience tested whether a tablet-based digital assessment combining augmented reality tasks and machine learning could identify mild cognitive impairment in Parkinson's disease (PD-MCI) more reliably than routine clinic visits allow. Researchers led by Karolina Poplawska-Domaszewicz and colleagues, working across centers in Sweden, the United Kingdom, Poland and Italy, gave 21 people with Parkinson's disease a 10-minute, self-administered test called the Altoida NeuroMarker, built from thirteen motor, augmented reality and language tasks. Clinicians, using standard tools including the MMSE, ACE-III and Hoehn and Yahr staging, classified 6 of the 21 participants as having PD-MCI. The digital tool identified all 6 of those cases, giving it 100 percent sensitivity (95 percent confidence interval, 54.1 to 100), but it also flagged 11 of the remaining 15 participants as likely impaired, cutting specificity to 26.7 percent (95 percent CI, 7.8 to 55.1).

Cognitive impairment is a common non-motor feature of Parkinson's disease and a factor in caregiver burden and eventual progression to dementia, but researchers describing the diagnostic landscape have noted that standard neuropsychological screening can be time-consuming and can still miss subtler deficits, part of the motivation for testing faster, tablet-based alternatives. The NeuroMarker's builders point to it as a possible screening layer that could be added ahead of, not instead of, full clinical assessment, since it can be run in about ten minutes without a specialist administering it in person.

The study's own numbers temper any early enthusiasm. With a positive predictive value of 35.3 percent (95 percent CI, 14.2 to 61.7) and overall accuracy of 47.6 percent (95 percent CI, 25.7 to 70.2), the tool's classifications matched clinician judgment on fewer than half of participants overall, and agreement measured by Cohen's kappa was 0.17, in the range statisticians describe as only slight. The authors report the study as a feasibility pilot, not a validation trial, and note that group differences in age, cognitive test scores, sex and education existed between the participants classified as impaired and those who were not, factors that could have influenced the tool's output independent of true cognitive status.

The research team frames the results as a first, feasibility-stage look at whether the platform can distinguish PD-MCI from unimpaired cognition in Parkinson's disease, and says larger, better-balanced cohorts are needed before any claim about diagnostic accuracy can be made.