Tablet-based augmented reality test spots mild cognitive impairment in under 15 minutes

A brief, self-administered tablet assessment that asks people to trace shapes, navigate virtual objects in augmented reality, and describe images aloud has shown strong accuracy in identifying mild cognitive impairment, the transitional state that often precedes dementia. In a cross-sectional study spanning the United States, Australia, and Spain, a machine-learning model trained on the multimodal digital assessment distinguished clinically adjudicated mild cognitive impairment from cognitively unimpaired performance with an area under the receiver operating characteristic curve of 0.90, and the result held up in an independent validation cohort. The findings, published as a peer-reviewed study, suggest that a scalable digital tool could help close one of the most persistent gaps in dementia care: the vast majority of people with mild cognitive impairment never receive a diagnosis.

The scale of that gap is striking. Mild cognitive impairment affects roughly 15 to 22 percent of people over 65, yet Medicare-based estimates cited by the researchers suggest that up to 90 percent of cases go undiagnosed. There is no single definitive test for the condition, and diagnosis typically rests on lengthy cognitive batteries administered by trained personnel, which makes routine cognitive evaluation rare in frontline settings where consultation times are short, neuropsychological testing is scarce, and memory complaints are often normalized as an inevitable part of aging. With Alzheimer’s disease among the most common underlying pathologies, and with amyloid-lowering therapies now raising the stakes of early identification, the need for faster and more accessible assessment has become increasingly urgent.

The platform evaluated in the study, the Altoida NeuroMarker, is designed to capture complementary aspects of cognition and function through three task families. Motoric tasks involve tracing and tapping activities that measure fine motor performance, processing speed, coordination, and executive function. Augmented reality tasks, named “Back In Time” and “Day Out,” require participants to place, locate, and recall virtual objects within their physical environment while completing goal-directed activities, engaging learning and memory, spatial navigation, planning, and task sequencing. Speech tasks involve verbal description and recall of visual stimuli. The assessment runs on an iOS tablet and produces an MCI classification, a global composite score, and scores across five cognitive domains aligned with DSM-5 categories: learning and memory, complex attention, executive function, perceptual-motor function, and language.

To train the classification model, the researchers assembled a dataset of 638 participants aged 50 and older, recruited from communities, research cohorts, and memory clinics between May 2022 and October 2024. Of these, 441 were classified as cognitively unimpaired and 197 as having mild cognitive impairment under Petersen and 2011 NIA-AA clinical criteria, with no restriction by cognitive subtype or presumed etiology. Clinical status was established through integrated clinical evaluation, including demographics, medical history, medications, Mini-Mental State Examination scores, and neuropsychological assessment, rather than cognitive test scores alone. The resulting dataset contained 970 assessments, and participants completed either one or two evaluations per study visit.

The machine-learning model itself was an XGBoost binary classifier built from 22 features reflecting performance on the motoric, augmented reality, and speech tasks, plus four demographic variables: age, sex, education level, and dominant hand. Hyperparameters were tuned with Bayesian optimization, and performance was estimated through five-fold cross-validation with participant-level grouping, so that each fold served as an internal holdout set. An optimal classification threshold was determined within each fold, and the mean threshold of 0.315 was then applied in a second round of cross-validation from which the final metrics were derived. A logistic regression model using only the four demographic variables served as a comparator, and it achieved a cross-validated area under the curve of just 0.68.

The headline performance figures were robust. In the training data, the platform achieved an area under the curve of 0.90 (95 percent confidence interval 0.88 to 0.92), with sensitivity of 0.83, specificity of 0.83, and overall accuracy of 0.82. The positive predictive value was 0.68 and the negative predictive value was 0.91. In an independent validation cohort of 75 participants, 50 recruited in Chicago with available plasma pTau217 measurements and 25 in Barcelona, the model showed comparable discrimination, with an area under the curve of 0.87, sensitivity of 0.75, specificity of 0.86, and accuracy of 0.81. The platform also outperformed the demographics-only model in the validation cohort by a statistically significant margin, indicating that its performance was driven primarily by assessment-derived features rather than who the participants were.

Calibration, a dimension often neglected in digital biomarker studies, was also assessed. The model achieved a Brier score of 0.18 in the validation cohort, significantly better than the 0.25 of a prevalence-based null model that simply assigns every assessment the observed MCI rate, indicating that the predicted probabilities carried genuine information beyond base rates. Beyond classification, the composite and domain scores differed significantly between groups: participants with MCI scored a mean of 41.27 on the composite measure versus 49.22 for cognitively unimpaired participants, with lower scores across all five domains and the largest group differences in language, complex attention, and perceptual-motor function. Exploratory subdomain analyses pointed to differences in object naming, flexibility, visual perception, sustained attention, and processing speed, suggesting the platform captures impairments that extend well beyond the memory deficits classically associated with amnestic MCI.

Reliability and validity checks added further texture. The MCI classifier showed substantial agreement across repeated assessments, with a Cohen’s Kappa of 0.60 among 40 validation participants who completed two evaluations. The composite score showed good test-retest reliability, with an intraclass correlation coefficient of 0.80 among 220 participants retested two to six weeks apart, while domain coefficients ranged from 0.55 for learning and memory to 0.79 for complex attention, a pattern the authors attribute to greater day-to-day variability and practice effects in memory and executive measures. Convergent validity was supported by moderate correlations between the composite score and established instruments, including the Mini-Mental State Examination and the RBANS Total Index, and by smaller but significant correlations between individual domains and corresponding neuropsychological tests, with the platform capturing related rather than identical aspects of cognition.

Acceptability data offered a glimpse of how such a tool might fare outside the lab. The median completion time was 13.8 minutes, rising to 15.9 minutes for participants with MCI compared with 12.9 minutes for cognitively unimpaired participants. Among 382 survey respondents, 67 percent reported feeling comfortable using digital devices, 66 percent reported frustration or worry about their memory in the past month, and 90 percent said they would be open to completing a short cognitive assessment as part of routine medical care. Older participants and those with MCI were more likely to find the tutorial helpful, and although augmented reality tasks were most frequently rated as the most challenging, staff reported no participants asking to stop because of distress. Most respondents preferred completing such an assessment at home rather than in clinical settings.

The authors are careful to frame the platform as an adjunct rather than a replacement for clinical evaluation. The classifier identifies the clinical syndrome of MCI, not its underlying etiology, and results should be interpreted alongside clinical history, medication review, and other components of a full assessment. Because predictive values depend on the prevalence of MCI in the population being tested, the tool is intended for people presenting with cognitive concerns rather than for population-wide screening of asymptomatic individuals; a negative result may provide reassurance, while a positive result should prompt further clinical evaluation. Limitations include the absence of systematic biomarker data for etiologic characterization, the fact that performance within individual recruitment settings was not established, a cohort that was predominantly White and non-Hispanic with high self-reported digital comfort, and the need for larger prospective validation across diverse healthcare settings before broad clinical implementation. If those steps succeed, the researchers argue, multimodal digital assessments could support earlier identification, more efficient referral decisions, and better integration with the blood-based biomarkers now transforming Alzheimer’s diagnostics.

Subject of Research: A multimodal digital cognitive assessment platform for identifying mild cognitive impairment

Article Title: A multimodal digital NeuroMarker platform with composite and domain-level validity for identification of mild cognitive impairment: a cross-sectional study

Article References: Iulita, M. F., Kolachalama, V. B., Brugada-Ramentol, V., Harrison, J., Griffin, N., Benavides, G. S., Vera-Campuzano, E., Minguillon, C., Alcolea, D., Mackintosh, L., Ward, L., Merino, A. T., Balasa, M., Ishida, S., Rivera, O. G., Calvet, M. S., Robertson, J., Doecke, J., Mas, C. P., … Streel, E. (2026). A multimodal digital NeuroMarker platform with composite and domain-level validity for identification of mild cognitive impairment: a cross-sectional study. eClinicalMedicine, 101, Article 104266. https://doi.org/10.1016/j.eclinm.2026.104266

Image Credits: AI Generated

DOI: 10.1016/j.eclinm.2026.104266

Keywords: mild cognitive impairment, digital biomarkers, machine learning, augmented reality, Alzheimer's disease, cognitive assessment, XGBoost, test-retest reliability, convergent validity, early detection, neuropsychological testing, primary care