2026 American Academy of Neurology Abstract Website

Objective:

To analyze whether data-driven MoCA score-based subtypes from the ADNI cohort can predict dementia conversion better than conventional MCI classifications and identify the neuroimaging features that differentiate these subtypes in the ADNI population.

Background:

Mild cognitive impairment (MCI) is a cognitive state between normal aging and dementia. However, early detection of cognitive decline is still a challenge. MoCA is widely used for screening, yet the domain-based classification often overlooks the heterogeneity of cognitive decline. These hidden patterns could be found in a large cohort, hence the need to explore data-driven subtypes. This may better capture individual variability, improve precision in prognosis, and link cognitive domains to underlying neuroimaging features.

Design/Methods:

We collect the ADNI cohort data (MoCA, demographics, diagnosis, and imaging features) and the cross-sectional MoCA of the healthy King Chulalongkorn Memorial Hospital cohort (total score 23 to 27) for initial comparison. The Louvain method is used to correlate domain profiles and subtype participants, binned by their total score. Dementia conversion was modeled with Cox regression, and Kaplan-Meier survival analysis was used to compare dementia conversion across subtypes. The subtypes are then compared to the Petersen/Winblad criteria. Neuroimaging features from MRI/PET, such as DTI metrics, are analyzed with recursive feature elimination with cross-validation of the random forest model to identify discriminative regions by subtype.

Results:

Preliminary results showed that data-driven MoCA subtypes exhibit greater heterogeneity in dementia conversion, while performing similarly to the Petersen/Winblad criteria. The median score of diffusion tensor imaging (DTI) metrics within the regions of interest showed that subtypes with a higher proportion of dementia cases exhibited increased mean diffusivity (MD) and radial diffusivity (RD). While axial diffusivity (AD) and fractional anisotropy (FA) were comparable on all subtypes.

Conclusions:

Our data-driven MoCA subtypes can exhibit greater heterogeneity in dementia conversion with neuroimaging correlations, while performing similarly to the traditional Petersen/Winblad classification.