In one of the largest MRI studies to date, we determined the correlates of ventricular enlargement in AD and MCI and ranked them in order of effect size. We found that ventricular enlargement (1) correlates with cognitive impairment (measured using MMSE, global and sum-of-boxes Clinical Dementia Rating, Geriatric Depression, delayed logical memory test and Hachinski Ischemic scores), (2) correlates strongly with lower levels of CSF Aβ1-42 but not with CSF Tau (after adjusting for age, sex and educational level), (3) predicts future cognitive decline (in MMSE, global and sum-of-boxes Clinical Dementia Rating), in all of the AD, MCI, and normal groups (4) ApoE4 carriers versus non-carriers, and (5) ADAS-Cog (tests including word recognition, spoken language and word finding).
One notable aspect of this cohort is that ApoE4 carriers are somewhat over-represented relative to other studies. As noted in , approximately 64% the AD group, 54% in the MCI group, and 27% of the normal group carried one or two copies of the ApoE4 gene (each copy confers increased risk for AD). In a related study (Ho et al., 2010
, in press
), we compared the level of brain atrophy in 587 ADNI subjects with that of another cohort of 113 MCI and AD subjects from the Cardiovascular Health Study-Cognition Study (CHS-CS; see Lopez et al., 2003
, Raji et al., 2010
, and Ho et al., 2010
, for details of the CHS-CS study). The atrophic pattern in MCI and AD was consistent in both ADNI and CHS populations, but the percentage of patients carrying the ApoE4 genetic variant was much higher in ADNI compared to CHS for both AD (ADNI=67.0% versus CHS=23.3%; X21
= 18.8, P-value=1.5 × 10-5
) and MCI subject groups (ADNI=54.6% versus CHS=27.5%; X21
= 16.0, P-value= 6.5 × 10-5
); these numbers differ very slightly from the figures reported in this paper, as Ho et al. (2010)
examined only 587 of the full cohort of 804 ADNI subjects assessed here). Differences in the prevalence of ApoE4 may be due to the fact that ADNI assesses a referral clinic-based population rather than a population-based community cohort (as is the case for the CHS study). There is some evidence that the referral-based cohort, ADNI, may include subjects with more severe symptoms of AD at an earlier age (Ho et al., 2010
), suggesting that even larger studies comparing ADNI data with other cohorts may be useful.
In a subsequent pilot ADNI study (N=240; Chou et al., 2009
), we attempted to correlate ventricular morphology with ApoE genotype and found no effects (in 115 carriers versus 122 non-carriers), supporting the argument above. However, we were concerned that the sample was too small to detect subtle associations so here we used a sample size almost three times greater, and still found no effect. Even so, the expanded dataset allowed us to detect significant differences between MCI and normal, and to rank a large range of influential covariates according to their effect sizes.
When we correlated baseline ventricular morphology with subsequent changes over 1 year, in MMSE, global CDR and sum-of-boxes CDR scores, all maps were highly significant. This is a useful observation, as it shows that all regions of the ventricles, not just selective regions, have characteristic expansion that predicts future decline. Even so, this correlation is to be expected, as subjects who are more impaired at baseline are more likely to have future cognitive decline than subjects who are less impaired. In other words, cognitive impairment measured by MMSE, global CDR or sum-of-boxes CDR scores, predicts (or correlates with) future decline in the same measures. Furthermore, the ApoE4 gene and increasing age are risk factors for developing AD, so that in any sufficiently large group of controls, MCI, or AD subjects, the ApoE4 gene (and age) will also correlate with future cognitive decline.
The failure to detect a correlation between Tau measures and ventricular morphology does not mean that there is no such association, and the effects in the maps are borderline. In , ventricular expansion correlates well with A-beta levels in the CSF, and somewhat less well with Tau effects after controlling for age sex and educational level, but visual inspection of the maps in the full sample of 397 subjects shows that the A-beta effects are quite robust, and the Tau effects are also formally significant but cover less of the ventricular surface. This suggests that either the effects are more anatomically selective for Tau, or, more likely, they have weaker effect sizes across the entire surface and so do not pass the significance threshold in so many places on the surface. Due to a peculiarity of the false discovery rate method, a map is only declared significant overall if there is some statistical threshold (called the critical P value) that can be applied to the map, that successfully controls the proportion of false positives in the map to be no more than 5%. This criterion is satisfied for Tau uncontrolled for age, sex and educational level, but only just, as the critical P value is very low (0.0029). In FDR, perhaps confusingly, low critical P values denote weaker effects than higher critical P values, as a low critical P value means that only stringent statistical thresholds can control the false discovery rate (if a high threshold controls the false discovery rate, generally all lower ones do). When the Tau effects are controlled for age, sex and educational level (), the map is not much different from the uncorrected map, but it is marginally weaker and just falls below the threshold for FDR, so is declared not significant. The most reasonable interpretation is that Tau effects are not as robust as those of A-beta, which pass the FDR threshold easily (the critical P value is 0.0361 in , showing that much of the surface shows a detectable effect). Most likely, if the sample size were expanded, both effects would be robustly detected. This scenario has been noted in other papers relating CSF biomarkers to morphometry. Weak correlations are detected in some studies but not others. Any null findings do not necessarily imply that the biomarkers are not causally related (as both are sensitive to the ongoing progression of AD, but the CSF markers tend to fluctuate over time).
In this study, we chose to analyze a radial distance measure (i.e., distance from a central curve threading down the hippocampus) instead of a surface distance measure, i.e., the distance from one surface to another. In very early work (Thompson et al., 1996
), we did in fact quantify differences in anatomy using a distance between the surface mesh points across subjects after aligning all the subjects' brains to a standard coordinate system. This can be useful, and a series of early computational anatomy papers focused on modeling the mesh displacements as chi-squared or Hotelling's T-squared distributed random fields (Thompson and Toga, 1997
). Even so, a limitation of the surface displacement measure is that the relative shifting of the surfaces in stereotaxic space can be due to atrophy occurring elsewhere in the brain. This means that effects mapped on the surfaces may be disease-related but may not be occurring in the structure modeled. Subsequently, we switched to a method based on fitting a central line down the medial axis of the structure (as in related work by Yushkevich et al., 2009
, Styner et al., 2005
, Gerig et al., 2001
, Pizer et al., 2003
, Bansal et al., 2000
and many other authors). This has the advantage that the distances to this central line do reflect atrophy that is intrinsic to the structure – the resulting atrophy measures would not be altered by a shifting of the structure in stereotaxic space. Alternatively, surface invariants may be used (Gutman et al., 2009
), although they do not provide spatial detail on the pattern of effects. Radial distance maps have been used in over 30 studies and occasionally allow better group discrimination than simple volumetric measures, although both measures are useful. Alternatively, it is possible to analyze parcellated subvolumes, but again they provide less spatial detail on the pattern of effects. For a very detailed comparison of many different surface metrics for disease discrimination, please see Wang et al. (2010)
Also, the computation of group anatomical differences relies on a computed correspondence derived from a surface-based parameterization method that stretches a grid over the surfaces. Even so, stretching a grid over the surface does not mean that the points match up either anatomically or in the best possible geometric way. Ongoing research in computational anatomy is focusing on how to align features within surfaces to provide higher order correspondences between regions that may correspond across subjects. This may lead to the reinforcement and better detection of systematic effects, especially when differentiated cellular fields lie within surfaces (Zeineh et al., 2001
). Current work on surface reparameterization includes alignment of explicitly identified internal landmarks that lie within the surfaces (Thompson et al., 2004
; Durrleman et al., 2008
), and alignment of curvature fields or other differential geometric features such as Riemannian structures using flows within surfaces (Lui et al., 2010
). Active work is focusing on which method boosts power the most for detecting statistical effects on brain structure (Wang et al., 2010
The maps reported here assessed residual anatomical differences after an initial 9-parameter global scaling of all AD, MCI, and control subjects' images to match an anatomical template. This scaling was performed in the automated registration step, and, in our cohort, the degrees of scaling (mean global expansion factors) for groups of controls, MCI and AD patients were 1.020 (SD=0.031), 1.019 (0.031) and 1.018 (0.026) respectively, and there was no significant difference among the three groups (single factor ANOVA p-value=0.773). As such, we did not adjust for group differences in overall brain scaling in our analyses, as no such differences were detected.
In general, our studies of ventricular differences show bilateral statistical effects if they show effects at all, and only occasionally, when effects are borderline, effects are picked up on one side but not the other. There are some natural asymmetries in the anatomy of the ventricles: the occipital horn extends around 5 mm further back on the left than the right. This asymmetry, which is present in most but not all subjects, emerges early in embryonic development due to the tendency for the perisylvian language areas, such as the planum temporale, to expand more in the left hemisphere. This expansion has a mild but systematic torquing effect on subcortical anatomy. One limitation of this study is that we did not test relationships between the degree of ventricular asymmetry and cognitive decline; this is because the primary biological process of atrophy is pervasive in both hemispheres. As such, we do not expect there to be strong hemispheric differences in the relation between ventricular expansion and cognitive decline.
It is interesting to determine the possible contribution of vascular disease burden to the ventricular expansion noted here, especially in the light of recent reports that the level of atrophy in elderly normals is associated with cardiovascular risk factors such as high body mass index (Raji et al., 2010
) and carrying the obesity risk gene, FTO (Ho et al., 2010
). Salerno et al. (1992)
and others have argued that otherwise healthy, but hypertensive elderly subjects have significantly larger mean ventricular volumes. In our study, however, the Hachinski ischemic scale showed no significant differences among the three diagnostic groups (single factor ANOVA p
-value=0.717), suggesting that vascular burden is unlikely to be the primary contributor to the effects. Even so, subtle vascular insufficiencies may contribute to neuronal atrophy and may not be readily detectable on T1- or T2-weighted MRI. In a recent study of obesity and brain structure in an independent sample (Raji et al., 2010
), the effects of body mass index on brain atrophy were quite strong, but could not be explained by conventional measures of white matter vascular burden, such as the volume of white matter hyperintensities on T2-weighted MRI. It is therefore possible that the ventricular expansion seen here is somewhat independent of vascular disease burden, or that microvascular damage may contribute to it but occurs at a finer anatomical scale than is readily detectable on T1- or T2-weighted MRI.
In this paper, we report correlations between atrophy and cognitive or CSF-derived measures in the pooled ADNI sample (combining patients with AD, MCI and controls), yet we also report other correlations within groups split by diagnosis (“disaggregated” analyses). Both types of analysis are complementary, and each has limitations. When analyzing a mixed cohort of subjects with AD, MCI, and controls, it is important to determine (1) the cognitive correlates of atrophy in the entire study, and (2) whether the chosen biomarker of disease burden is linked with decline across the full spectrum of controls, MCI, and AD subjects. As the whole cohort is arguably a continuum, it is vital to look beyond the diagnostic categories and see if the level of atrophy seen is related to function, and if so, which functional scores it relates to. This same correlation may be missed if it is assessed within one group only (e.g., MCI) due to a “restricted range” effect. Similarly, true correlations may be missed if the range of cognitive performance is restricted to include only healthy normal subjects. Furthermore, it is a fallacy to pre-select a diagnostic group based partly on cognitive domains, and then later test if a correlation is maintained with a cognitive subscale that is correlated with tests used to select the group. By running split analyses only, many important correlations will be missed. For instance, the level of brain atrophy correlates well with CSF-derived measures of pathology across the continuum from aging to MCI to AD. But if one sub-selects a group such as MCI, or a group of subjects with a very narrow range of disease burden, it is possible that no such correlation will be detected, due to the restricted range. If groupings are made based on the measure whose correlation is being tested, results may be uninterpretable. If the selection criterion for the group correlates with the variable of interest, nearly all the maps would be false negatives due to the truncated range.
Pooled analyses also have limitations. First, correlations with cognitive scores in a pooled cohort will tend to show similar patterns to a direct comparison of AD and healthy controls, if the cognitive measures are correlated with diagnosis. Second, if correlation analyses are performed across the full diagnostic continuum in a pooled cohort such as ADNI, then any correlations detected may depend somewhat on the proportion of subjects with each diagnosis - in ADNI, this is approximately 1:2:1 for AD:MCI:controls. In other words, part of the range of cognitive decline may be over-sampled. In ADNI, the over-sampling of MCI is deliberate, but it may not reflect a representative sampling of all subjects of a certain age. As such, any correlations with the atrophy in ADNI may not be detected in the same degree in other population studies with different proportions of subjects, or within diagnostic subcategories. For that reason, both pooled and split analyses have value for understanding the cognitive and pathological correlates of atrophy.
In summary, we examined the clinical and pathological correlates of ventricular expansion, in a very large sample of AD, MCI and healthy subjects. Although the ventricles are not the site of pathology deposition, and are at best an indirect measure of brain atrophy, they are nevertheless easier to measure than hippocampal and cortical structures, due to their high contrast on standard MRI. The resulting maps and measures show promise as a biomarker of AD, and provide a useful measure for combination with other more direct measures of pathology or neuronal loss.