The present study investigated the ability of cortical thickness from CSC, empirically defined by their correlation with domain-specific cognitive factor scores, to predict clinicalconversion to AD and accelerated worsening of clinical severity. Cortical thinning in each CSC was associated with faster progression to AD and with faster rates of decline in CDR-SB score. The analyses converge on three main findings. First, domain-specific cortical signatures of cognition can be estimated which are largely independent of the cortical signature of AD.Second, these cortical thickness measurements and cognitive performance account for unique variance in conversion to AD and accelerated worsening of clinical severity ( and ).Third, latent factors representing performance on neuropsychological measures of memory,executive function, and language and their corresponding CSC are the best predictors of conversion to AD and clinically relevant decline in nearly all models.
These results may provide clinicians with the ability to use the ADNI neuropsychological battery in conjunction with a structural MRI scan to provide a more accurate estimate of the risk of conversion to AD. Importantly, in current clinical practice an MRI, which can be conducted in a few minutes, is almost universally performed in dementia evaluations while neuropsychological testing, which can take an hour or more, is much less common. As seen in the results, a 1 SD loss in thickness in the memory CSC and 1 SD decrease in memory function more than doubles the risk of developing clinical AD. Knowledge of AD risk is important for both the treatment of patients and for identifying potential candidates for novel therapeutic interventions.
The present study utilized a large, well-characterized sample of participants to empirically define cortical signatures of cognition. Although the focus of this paper was notabout the detailed significance or implications of any particular region or domain, we briefly discuss the CSC and cortical signature of AD in relation to previous research. We leveraged advantages of rigorously constructed factor scores from a confirmatory factor analysis (Park et al., this issue) to identify structure-function relationships. This approach is in contrast to testing the relationship of the many variables in the ADNI neuropsychological battery; however, sincethe goal was to identify regions associated with domains of cognition for use in subsequent analysis, we were less interested in differences within each domain (e.g. encoding versus retrieval) (Walhovd et al., 2010
; Wolk & Dickerson 2011
We can use results from other imaging studies to confirm using examples below that the cortical signatures are measuring each domain accurately (e.g., content validity). For example, the memory CSC is dominated by areas of the mesial temporal lobe that have been shown to be related to memory both in function and structure (Buckner et al. 2004
; Burggren et al., 2011
; Dickerson et al. 2008
; Fjell et al. 2008
; Johnson et al. 2006
). The executive function/processing speed, language, and visuospatial CSC all had significant overlap, but also encompassed unique brain areas. Importantly, our findings emphasize that executive function is not synonymous with frontal lobe functioning and suggests that successful task performance also relies on non-frontal brain regions responsible for other fundamental skills. The language CSC corresponds to regions shown to correlate with the Boston Naming Test and the Controlled Oral Word Association Test (COWAT), both measures of language ability (Anh et al., 2011). The attention CSC had the fewest vertices, a finding that could be due to acute demands of the tasks.The cortical signature of AD we used is consistent with regions that have previously been shown to be atrophic in AD (Buckner et al. 2005
; Dickerson et al. 2009
; Fjell et al. 2009
). At lower thresholds, most of the brain is atrophic in large AD samples. We chose a high threshold to constrain the signature to be the approximate size of each CSC, but note that alternative thresholds could have been chosen.
Survival plots suggest that there at least 2 distinct patterns of cortical atrophy in Alzheimer’s disease (). Specifically, the middle quartiles of the memory CSC and AD pathology show similar survival curves. In contrast, the visuospatial, executive function/processing speed, and language CSC revealed that the top two quartiles were similar and the bottom two quartiles were similar. These groupings are not surprising given the overlap between CSCs. However, the finding is also consistent with the notion of multiple etiologies of AD (Buckner, 2004
). Future studies should probe covariance patterns in longitudinal change in cortical thickness to better capture potentially separable processes. It is likely that a combination of behavioral and structural change metrics will be ideal for identifying those at highest risk of conversion to AD and accelerated worsening in clinical severity.
Our results can be contextualized in a hypothetical model of biomarker and cognitive change in pathological AD proposed by Jack and colleagues (Jack et al. 2010
). The Jack model proposes that AD pathology begins with abnormal buildup of amyloid beta, which subsequently results in irregular processing of tau protein, leading to neurofibrillary tangles and cellular apoptosis, impaired function in brain systems, cortical thinning in certain brain regions, and eventually cognitive decline and functional disability characteristic of clinical AD. Although the timing and relative order of biomarkers that measure these signs and symptoms is an active area of research, their importance in AD is not disputed. The present results suggest that amount of atrophy predict rate of functional decline, and thus takes place beforehand. This inference is drawn from the finding that, after controlling for neuropsychological factor scores as indices of behavior that were used to define the CSC, the CSC still predicted conversion and significantly increased CDR-SB trajectories. Additionally, the memory CSC and neuropsychological score together better predicted conversion from MCI to AD than the cortical signature of AD alone. Although it is not known how amyloid (either measured by PET or CSF samples) would affect the predictive value of the CSC in the present study, PET scans and lumbar punctures are notperformed as routinely as MRI at this time. Thus, utilizing a more widely-applied technique to inform the risk of conversion may be preferable.
Although the aim of this work was to investigate conversion from MCI to AD, it is also necessary to predict conversion from cognitively normal to MCI and to AD. At present, this involves identifying individuals with preclinical AD (Sperling et al. 2011
), usually with a PET scan for amyloid. However, our study suggests that cognitive function, in combination with cognition-defined signatures of structural or functional MRI, could provide sensitive measures toidentify individuals with an increased risk of developing AD that may complement PET imaging. Future research using CSCs is needed to investigate which CSCs best predict conversion among healthy controls to MCI and AD.
This study probed the relationship of behavior and morphometry in predicting conversion to MCI and accelerated worsening in dementia severity. Importantly, we chose to define our morphometry metrics from structure-function relationships based on a validated factor analysis of the ADNI neuropsychological battery (Park et al., this issue), rather than pathological differences. In defining morphometry metrics based on behavior, we were able to target several aspects of the pathological AD disease process. Although memory, executive function, and language provided the dominant effects, thickness in other regions can be used to compute a cumulative odds ratio across CSCs or factor scores.
Several caveats merit attention. First, the ADNI sample is more highly educated than theUS population and represents a self-selected sample of volunteers who present to AD research centers. Thus, findings should be replicated in other more educationally and culturally diverse samples. Second, factor scores for particular cognitive domains are, by design, a generalization of performance on cognitive tasks designed to measure very specific aspects of cognitive function. CSC derived from these scores are a further abstraction, which may obscure specific forms of neuropsychological impairment that might predict AD early in its course (Wolk & Dickerson, 2011
). However, clinical AD defined in ADNI, as predicted in our survival models, is a disease of global impairment. A third caveat is that successful performance in any domain of cognition is not independent of other cognitive domains, and likely entails contributions from several neural networks (Wolk & Dickerson, 2011
). The present study’s CSCs were based on empirically defined brain regions correlated with neuropsychological performance, not neural networks. Thus, domain-specific CSCs in this study are also correlated with each other (Table S3
), which has implications for statistical models where all the CSCs are considered at the same time. Finally, the present study did not investigate cortical volumes that were empirically associated with cognitive function. Volumetric analyses, whether using Freesurfer volumes or voxel-based morphometry to identify voxels that are correlated with neuropsychological factors scores, should be explored in future studies.
The ability to accurately identify risk for developing clinical AD while an individual has normal cognition or MCI will enhance selection of participants for treatment trials and enable clinicians to decide on optimal management at an earlier stage. Establishing CSCs is a promising approach that integrates structural imaging with the gold standard of clinical disease stage, neuropsychological measures, and thus enables researchers to track change in multiple modalities over time. The present study indicated that factor scores and CSCs for memory and language both significantly predicted risk of conversion to AD and accelerated deterioration in dementia severity. We conclude that predictive models are best when they utilize both neuropsychological measures and imaging biomarkers.