The present study is one of few studies of normal aging where image-derived hippocampal volumes, APOE genotype, and verbal memory performance have been jointly investigated. In our investigation we also included other variables (e.g. age, gender, hippocampus laterality index) that could be important a priori for prediction of verbal memory performance (i.e. CVLT-LD score). Applying the conditional inference model to our multivariate sample, we expected to reveal the sequential importance for each variable in predicting CVLT-LD.
Investigating our results (Figure 6a) we see that gender is the one variable that best predicts CVLT-LD in our sample. This is in concordance with findings in other studies of verbal memory function among elderly normals [52, 53], where women in general, outperform men [1, 52]. For men, there were no other variables that could be used to reject the null-hypothesis of there being no relationship between variable and response. For women however, age provided another node at 70 years, inferring that women over 70 years perform poorer than those at 70 years or less. The lack of predictive variables in men might indicate a power-problem reflecting on the small number of male participants. However, one could also argue that the findings reflect the role of hippocampus in verbal memory. We found a slightly positive relationship between hippocampal volume and CVLT-score, which is in concordance with Walhovd et al., who found a correlation between CVLT free recall after 11 weeks, but not after 5 minutes . As also found, the size of the hippocampus declines by about 0.03 mL/year in both men and women, and given the smaller overall hippocampal volume in women, this would indicate a more prudent role for hippocampal volume in verbal memory in women.
We also found that APOE genotype did not have any important predictive value regarding CVLT-LD score. This was also confirmed by the more simple Wilcoxon rank sum test. This result is in accordance with Smith et al. , who reported that the phenotypical significance of ApoEϵ4 seems to apply only in patients with diagnosed MCI or AD, but not in healthy controls [55, 56]. However, others [11, 12] have reported lower CVLT-LD scores in healthy elderly ApoEϵ4-positive individuals, suggesting that ApoEϵ4-related memory changes precede a clinical MCI or AD diagnosis.
The role of left hippocampal volume in verbal memory function has been reported in several previous studies [57–61], but the exact contribution of this structure in relation to ApoEϵ4-zygosity is still unclear. Hippocampal volumes have, in most cases, been reported second to ApoEϵ4-zygosity in importance at follow-up, or of no importance at all at baseline measurement . It is remarkable that in our sample, the role of ApoEϵ4-zygosity, but also age, is negligible compared to that of left hippocampal volume in predicting verbal memory performance. This suggests a prominent role for the left hippocampal volume in the hierarchy of predictors to verbal memory function. Furthermore, since our analysis did not reveal any significant associations between hippocampal volumes and ApoEϵ4-zygosity, our investigation does not support the notion that changes in hippocampal volumes are, in fact, results of APOE genotype. The lack of findings relating to the APOE genotype may reflect our sample, consisting of healthy, rather well-functioning men and women. If one were to expand the "normality" criteria to include more subjects, one might find stronger APOE genotype correlations.
One other finding was that the hippocampus volume declines relatively more than the cortical volume with age (Figure 3). This is contradictory to the common perception that the cortex is more prone to age related change than the hippocampus. However, due to the high age distribution in our material, we postulate that these findings reflect an accelerated hippocampal atrophy occuring in advanced age [62, 63]. A weakness in our study is the small number of participants. A total of 170 subjects could be too low, and only make for weaker statistical inference than would a larger sample. The sample selected for this study was also found to be rather homogeneous, something which is reflected in the above-normal distribution of education, IQ, and the age range (Table 1). However, in the literature there are indications that the effects of several predictive factors are only surfacing in the older segment of the population [36, 64], thus questioning the significance of including young people in studies of age-related disease. When comparing the results to other aging studies one should be cautious, considering the slightly younger age group in our sample. This age composition produces a smaller variance in cognitive function as compared to other aging studies. However, the selection criteria were motivated by the desire to include only healthy elderly people, and to allow for follow-up studies of the same participants, where cognitive decline, neurodegeneration, and morphological abnormalities are expected to occur on a larger scale.
A possible confounder in our study is the different scanners used to acquire the brain volumes. This is however, a substantial difficulty in multi-center studies, as various scanner vendors provide a wide range of models with different specifications such as field strength, gradient system, coils, and pulse sequence principles and parameters. This might not present as a problem in clinical practice, or when manual deliniations and segmentations are performed, whereas automated methods can be more sensitive to subtle differences in image properties . However, this apparent weakness in the study can also be considered as a strength, as it demonstrates the methodological robustness behind the results. The fact that we obtained similar findings in two independent samples (i.e the Oslo and the Bergen material), also in terms of e.g. hippocampal volumes and its lateralization, is a significant strength of our study.
The automated image segmentation method used is well proven, and correlates well with more conventional, but time-demanding and subjectivity-prone methods [37, 38]. This particular set of algorithms also produce reliable results compared to other, freely available software packages . Multivariate methods based on conditional inference trees were applied in our analysis. These are rather novel methods in the applied statistical community [51, 65], and have to a very little extent been used or known to the medical imaging and aging research community. However, such types of classification and regression trees (CART), e.g. , can provide robust and easily interpretable results.
A follow up study will be required to further investigate our findings. Such a study will provide an opportunity to do follow-up analysis much like Tupler et al. , and thereby make stronger inferences concerning predictive values from our variables. Furthermore, one would benefit from methodological improvements in magnetic resonance imaging and data-analysis made in recent years. It would therefore be interesting (and feasible) to acquire data with different MRI measurement techniques during the same imaging session, such as diffusion tensor imaging (DTI) and functional MRI (fMRI). Diffusion tensor imaging has given valuable biological information regarding 'white matter integrity' in age-related cognitive decline [67, 68]. In addition, a particular kind of BOLD fMRI examination, called resting-state fMRI (RS-fMRI) or task-free fMRI, has shown to be sensitive to neuropsychological/neurodegenerative diseases, such as AD, in that appropriate analysis of such data can reveal a disruption in functional resting-networks in the brain. The RS-fMRI method has a genuine potential as a biomarker of disease and also as an early objective marker of treatment response, but needs to be further investigated [69, 70].