Standard statistical methods were used to analyze the data, including descriptive statistics, t-test, analysis of covariance (ANCOVA), repeated measures analysis of variance, partial correlation, and linear discriminant analysis (LDA) with equal priors. The IBM-SPSS statistical package (version 30) was used for statistical analyses. All reported P values are two-sided.
Descriptive statistics for the various measures used are given in Table 1.
The overall age of study participants was 70.35 ± 0.43 y (mean ± SEM; N = 460). Figure 1 shows the frequency distribution of age in the three CDR groups (Table 1). Pairwise age differences among the three CDR groups were assessed using a univariate ANOVA where age was the dependent variable, and Group (CDR) and Sex were independent variables. We found the following. (a) The Group (CDR) main effect was statistically significant (F = 30.02, P < 0.001); mean age was significantly smaller in the Control group than the CDR 0.5 or CDR 1 groups (P < 0.001, Bonferroni corrected for multiple comparisons, whereas mean ages did not differ significantly between CDR 0.5 and CDR 1 groups (P = 1.0, Bonferroni adjusted). (b) The Sex main effect was not significant (F = 1.46, P = 0.228). And (c) the Group x Sex interaction was not statistically significant (F = 1.32, P = 0.267).
Descriptive statistics for Education are given in Table 1. The Control group (CDR 0) had ~ 2 years more education, whereas the years of education for CDR 0.5 and CDR 1 were very similar. We assessed the CDR-education relation by performing a univariate ANOVA, where Education (years) was the dependent variable and CDR was a fixed group factor. We found the following. (a) The CDR main effect was statistically significant (F = 13.12, P < 0.001). (b) Education in the CDR 0 group was significantly higher than that in the CDR 0.5 and CDR 1 groups (P < 0.001 for both comparisons, Bonferroni adjusted). (c) Education did not differ significantly between CDR 0.5 and CDR 1 groups (P = 1.0, Bonferroni adjusted). (d) The results of the Education CDR group comparisons above remained the same when Sex and Race were added as fixed factors and Age as a covariate in the ANOVA.
The Race distribution among the three CDR groups is shown in Table 1. This did not differ significantly among the CDR groups (Pearson Chi-Square test, χ = 4.8, P = 0.308).
The distribution of apoE genotype among the three CDR groups is shown in Table 1. This did not differ significantly among the CDR groups (Pearson Chi-Square test, χ = 11.79, P = 0.299).
Descriptive statistics for MMSE are given in Table 1. Mean MMSE decreased with increasing CDR: MMSE (CDR 0) > MMSE (CDR 0.5) > MMSE (CDR 1). All pairwise comparisons were statistically significant in a univariate ANOVA (P < 0.001 for all, Bonferroni corrected). The negative partial correlation between MMSE and CDR (controlling for Sex, Age, Education, and Race) was also statistically significant ( = -0.641, P < 0.001, N = 460).
The effect of various factors on cortical gray was assessed by performing a univariate ANCOVA where was the dependent variable, CDR and Race were fixed (categorical) factors, and Sex, Age and Education were covariates. The results are shown in Table 2. It can be seen that only CDR, Sex and Age had statistically significant effects.
Normalized cortical gray matter volumes () decreased systematically with increasing CDR score: (CDR 1) < (CDR 0.5) < (CDR 0). More specifically, we found highly significant negative partial correlations () between cortical volumes and CDR scale (controlling for sex and age), as follows. Left cortical gray (Fig. 2A), = -0.342 (P < 0.001, N = 460); (b) right cortical gray (Fig. 2B), = -0.317 (P < 0.001, N = 460). These correlations did not differ significantly (z = 0.424, P = 0.672).
We investigated further the decrease in cortical gray matter by computing partial correlation coefficients between normalized volumes and CDR score (controlling for age and sex) for each of the 70 cortical areas analyzed (35 for the left and 35 for the right hemisphere, Table 3). - CDR partial correlations were highly correlated between left and right hemispheres ( = 0.861, P < 0.001, N = 70; Fig. 3) and did not differ significantly between the two hemispheres (P = 0.442, paired-samples t-test), hence were averaged across hemispheres. decreased with CDR score most prominently in the hippocampus and least so in the pericalcarine cortex (Fig. 4). The volume decrease was most prominent in the temporal lobe, followed by parietal, cingulate, frontal, and occipital lobes (Fig. 5); volume decreases differed significantly among lobes (P < 0.001, F-test, ANOVA).
The - CDR partial correlations (controlling for age) were significantly correlated ( 0.594, P < 0.001, N = 70 areas) between men and women. Partial correlations were significantly more negative for men (= -0.185 ± 0.010, mean ± SEM) than for women (= -0.160 ± 0.012) (P = 0.018, paired samples t-test).
The effect of apoE group on - CDR partial correlations was assessed by performing a repeated measures ANOVA where the partial correlations were the repeated measures factor for each cortical area (N = 70). We found the following. (a) The apoE effect was statistically significant (P < 0.001, Greenhouse-Geisser test). (b) The mean - CDR partial correlation ( = -0.202) for the apoE[ε4ε4,ε4ε3] ( = -0.202) group was significantly more negative than that of the apoE[ε3ε3] ( = -0.147, P < 0.001, Bonferroni adjusted) and that of the apoE[ε2ε2,ε2ε3] ( = -0.150, P = 0.002, Bonferroni adjusted). (c) The mean - CDR partial correlations did not differ significantly between the apoE[ε3ε3] and the apoE[ε2ε2,ε2ε3] groups (P = 1.0, Bonferroni adjusted).
We computed partial correlation coefficients between normalized subcortical volumes and CDR score (controlling for age and sex) for each of the 17 subcortical areas analyzed (8 for the left hemisphere, 8 for the right hemisphere, and the brainstem, Table 6). Volume-CDR partial correlations were highly correlated between left and right hemispheres ( = 0.971, P < 0.001, N = 8) and did not differ significantly between the two hemispheres (P = 0.153, paired-samples t-test), hence were averaged across hemispheres. Volume decreased with CDR score most prominently in the amygdala and least in the pallidum (Fig. 7); paradoxically, caudate volume increased (not significantly) with CDR score.
The - CDR partial correlations (controlling for age) were significantly correlated between men and women ( 0.761, P < 0.001, N = 17 subcortical nuclei) and did not differ significantly between the two groups (P = 0.099, paired-samples t-test).
The effect of apoE group on subcortical - CDR partial correlations was assessed by performing a repeated measures ANOVA where the partial correlations were the repeated measures factor for each cortical area. Although correlations were least negative in the control group (CDR 0; Fig. 8), none of the pairwise comparisons in the ANOVA attained statistical significance (P ≥ 0.243).
In these analyses we used all averaged 44 brain volumes (35 cortical and 9 subcortical), adjusted for age and sex, to evaluate whether CDR staging can be predicted by a linear combination of the 44 brain volumes (adjusted for sex and age). We found the following.
The Control (CDR 0) and Dementia (CDR 1) groups are defined unequivocally, in contrast to the "undetermined" group (CDR 0.5) which is a heterogeneous, transitional group. Therefore, we first performed a LDA between the Control and Dementia groups, the results of which are shown in Table 7. It can be seen that the discrimination was excellent, with an overall correct classification rate of 88.7% (88.9% sensitivity and 86.1% specificity). These results were statistically highly significant (Wilk's Λ = 0.661, P < 0.001). The ROC curve is shown in Fig. 9; area = 0.875, SE = 0.035, P < 0.001).
The bar graph in Fig. 10 shows the correlations in the LDA structure matrix where the magnitude of the values (bar length) indicates the relative contribution of each area to the group discrimination; in general values > 0.3 are considered important contributors to the discrimination. It can be seen that areas with high values were also strongly associated with CDR (Figs. 4 and 7).
The results above provide a solid framework to test the hypothesis that, based on brain volumes, the "undetermined" group (CDR 0.5) may actually consist of 2 distinct cluster, Cluster A (closer to the Control group) and Cluster B (closer to the Dementia group). We tested this hypothesis by using the classification functions yielded by the Control vs. Dementia LDA above to classify participants in the "undetermined" group. This analysis yielded 2 clusters; Cluster A comprised 51 (70.8%) brains that were assigned to the Control group, whereas Cluster B comprised 21 (29.2%) brains that were assigned to the Dementia group. A LDA using them as the dependent, grouping variable and the 44 brain volumes as predictors gave excellent classification results shown in Table 8. The overall correct classification rate was 98.6%, with 98.0% specificity and 100% sensitivity. The ROC curve is shown in Fig. 11; area = 0.990, SE = 0.011, P < 0.001). Remarkably, the probabilities of classification of a brain to either Cluster A or B were very high, with 66/72 (91.7%) been 1.0, and the rest 6 > 0.97 (Table 9). This high reliability of dichotomous classification was further documented by calculating the Squared Mahalanobis Distance (SMD) of a case from the center of each Cluster. The mean SMD (± SEM) from the center of the Cluster to which a case was classified was 0.874 (± 0.126, N = 72), as compared to the SMD from the center of the other Cluster 31.12 (± 1.213, N = 72); this 36x difference was statistically highly significant (P = 8.8 × 10, t = 27.87, paired t-test, N = 72). Altogether, these results document the high reliability of dichotomous classification of brains in the "undetermined" CDR 0.5 group to 2 Clusters (Control-like and Dementia-like).