Brain volumes reflect consequences of both aging and disease. Consequently, structural MRI has emerged as an important tool for predicting cognitive function in the aged. Automated parcelation programs, such as FreeSurfer, make anatomically detailed measures of brain volume feasible in large study cohorts, but at the same time create the analytical challenge of how to model appropriately the effects of a large number of correlated predictor variables. This study examines two questions. First, is a latent-variable model (LVM) approach to modeling regional brain measures feasible in the sense of yielding models that fit well and that correspond well to the brain’s anatomy? Second, does application of this LVM approach improve our understanding of how brain volumes correlate with cognition in the aged?
Our data derive from the Alzheimer’s Disease Neuroimaging Initiative (ADNI) project, a multisite study of Alzheimer’s disease. Data from the initial ADNI sample (termed ANDI1 in this abstract) were obtained on 819 older adults with dementia, mild cognitive impairment (MCI), or normal cognition, we used factor analytic methods to create a latent variable measurement model of FreeSurfer brain volumes. The final model had 55 manifest variables loading on 12 region of interest (ROI) factors (e.g., Frontal, Hippocampus, Entorhinal, Cingulate, etc.). We used these ROI factors as predictors of cognitive performance on four dimensions: Verbal Memory, Digit Span, Verbal Fluency, and Perceptual Speed. Compared to simpler approaches for modeling brain atrophy, such as measuring hippocampal volume and total cortical gray matter, the LVM approach increased the percent of variance explained by 2-3 fold (e.g., explaining 39% of the variance of Verbal Memory, 36% of variance in Verbal Fluency, and 43% of the variance in Perceptual Speed). The Temporal/Parietal ROI factor was the strongest predictor of each cognitive dimension, and the Entorhinal factor was a strong predictor of all dimensions except Digit Span.
Multiple-group modeling was conducted by dividing the sample into three groups: cognitively normal, MCI, and demented. The 12-factor model fit the data very well at the level of strong factorial invariance across the three groups (cognitively normal, MCI, demented). Mean differences across groups in brain volumes differed by region, with the demented group having mean brain volumes from 0.50 to 2.0 SDs below those of the cognitively normal group, with the MCI sample falling between the other two. Relations with individual differences measures revealed that brain volume factors were not significantly related to measures of cognition for the cognitively normal, weak relations were found for the MCI group, and strong relations were found for the demented.
We conclude that LVMs of brain volumes are both feasible and reflect a useful approach to incorporating high-resolution volumetric MRI information in models of cognitive function in aging. Interestingly, neuropsychological research has long stressed the importance of the hippocampus for memory. Work by Quamme et al. (2004) and Yonelinas et al. (2007) supports the existence of two processes underlying memory performance: (a) Recollection, supported by the hippocampus; and (b) Familiarity, supported by the entorhinal cortex. Our study underscores the need to look at brain areas beyond the hippocampus when studying memory, given the strong results for the Entorhinal ROI factor as a predictor of the cognitive dimensions. This study also contributes to research tying particular brain regions to specific dimensions of cognition and declines in brain volumes underlying notable cogntive declines.
