JMIR Publications

ABSTRACT

Background:

Traditional neuropsychological assessments for cognitive decline are lengthy in-clinic evaluations by a specialist, with typical wait times of 6-8 months. This creates a substantial patient burden and prolonged diagnostic and treatment timelines. Digital cognitive assessments (DCA) offer a scalable solution to these challenges, but their validation is challenged by the scarcity of large, high-quality datasets with established ground truth.

Objective:

To develop a model to identify mild cognitive impairment (MCI) and probable dementia using metrics from the Digital Assessment of Cognition (DAC), a brief, remote-capable DCA. A secondary objective was to conduct a preliminary assessment of the model's validity.

Methods:

We applied a semi-supervised model-based clustering method to combine a large dataset (N=1189) of DAC assessments alone, with a smaller dataset pairing DAC assessments with ground-truth neuropsychological diagnoses (N=248). We examined the model's predictive validity by comparing its predictions with diagnoses on a held-out test set. We examined congruent validity by testing associations with traditional analog assessments and demographic variables.

Results:

We identified a 6-cluster model with 3 MCI clusters and 2 probable dementia clusters. The model identified cognitively unimpaired, MCI, and dementia groups with high accuracy (78.7%) on the held-out test dataset, and showed excellent ability to identify cognitive impairment (AUROC=0.985) and dementia (AUROC=0.932). We identified strong associations with traditional analog assessments and demographic variables. An exploratory analysis showed evidence that clusters correspond to clinically meaningful subtypes of MCI.

Conclusions:

These results validate prior exploratory work and demonstrate the potential for more nuanced, holistic, and scalable cognitive assessments in non-specialist settings.