DAT is already an epidemic among the elderly in the US and worldwide. To address the pressing need to better understand and treat AD, many researchers are focused on developing and validating AD-related biomarkers - quantitative AD-associated measures that serve as an indirect metric of disease severity. In the present study, we analyzed the relationship between memory loss - the most pervasive AD symptom - and hippocampal atrophy, the most established AD imaging biomarker. We examined how well hippocampal radial distance, a measure of hippocampal thickness, correlates with one task that assesses delayed recall for short stories (LM II) and two tests of delayed recall of a list of unrelated words (AVLT and the verbal memory test from the ADAScog) across the full sample (N=490, consisting of 148 NC, 245 MCI and 97 mild DAT subjects) and separately within each diagnostic group. We used a newly developed high-throughput hippocampal automated segmentation technique that has been previously applied in MCI and DAT (Morra et al., 2008a
; Morra et al., 2008c
; Morra et al., 2008d
). As this method shows promise as a potential analytic tool for clinical trials, we wanted to explore the associations between hippocampal morphology and several of the memory measures that have been repeatedly used as screening tests or as primary and secondary outcomes in MCI and DAT clinical trials.
None of the three verbal memory measures showed significant associations with hippocampal volume/morphology among cognitively normal elderly. In MCI, all three measures showed significant associations with atrophy in both the left and right hippocampi, with AVLT-DR showing the strongest linkages. One explanation for this finding is that subjects with mild cognitive problems find AVLT more challenging than the other two tasks. AVLT consists of a list of 15 words, whereas the ADAScog list consists of only 10 words. In addition, AVLT requires subjects to learn a distractor list following learning of the to-be-remembered words, which can interfere with consolidation of the first list. It is known that memory consolidation is highly dependent upon hippocampal functions. Further, the two word list tests (AVLT and ADAScog-DR) are comprised of unrelated words, in which it would be more difficult for subjects to utilize memory mnemonics (e.g., drawing associations among the words). In contrast, a short story (as used in LM II) provides both a context and built-in associations. Subjects are more readily able to recall the information content of short stories as opposed to lists comprised of unrelated words as they can more easily utilize memory strategies, such as retaining the theme of the story and using pictorial rehearsal.
The observed stronger associations between LM II-DR and hippocampal atrophy in MCI at follow-up is another interesting observation. It could be due to two separate processes. First for some MCI subjects it could reflect disease progression where the progressive loss of the ability to compensate by means of these memory strategies results in a tighter hippocampal-memory performance association. On the other hand many MCI subjects demonstrated a learning effect (i.e., had improved LM II scores in follow-up, see ). Perhaps the ones benefiting most were the subjects with the least amount of atrophy, which in turn further strengthened the association between memory performance and hippocampal radial atrophy.
The DAT subjects were the only group where differences were detected only on one hippocampus. They showed predominantly left sided hippocampal-memory recall associations. The greatest effect was observed with the cognitive test specifically designed for DAT subjects – the ADAScog. This task is the easiest of the three and showed higher performance levels in DAT subjects, suggesting that the ADAScog would be a better measure for studies hoping to provide above-floor performance levels. DAT subjects showed better performance on LM II-DR at baseline and a significant LM II-DR association with the hippocampal formations. However, this association becomes nonsignificant at follow-up when most DAT subjects can remember only one information unit on average (floor effect, see ). AVLT-DR, the most challenging memory test of the three, showed no associations with hippocampal volume or radial distance in DAT, likely due to a floor effect (DAT subjects recalled an average of less than 1 word, see ).
We used two different criteria to assess whether a map was significant after multiple comparisons correction, as both tests are somewhat prevalent in the brain mapping literature, although not always applied to the same data as they are usually considered as alternatives. First, the total supra-threshold surface area (with p-values more extreme than 0.01) was used and a corrected p-value was given for its rank in a null distribution obtained by randomization. And second, we used FDR theory to see if there was a statistical threshold that could be applied to the map that controlled the false discovery rate at the conventional 0.05 level. These are slightly different criteria and they are not always true in the same situations; they agreed for control and DAT groups but permutation gave slightly more powerful results in MCI. These two tests, which have different definitions, are generally true at the same time (i.e., they generally declare the same maps as significant), but for some effects one of the tests may work and the other may not. There is a point of connection between the CDF plots used in the FDR theory and the permutation tests, in that one could look up whether the map, thresholded at p=0.01, controls the FDR. By contrast, permutation tests find out the null distribution for this suprathreshold area by randomization. For the memory scores in the MCI subjects, the CDF plots show that the supratheshold area for map thresholded at p=0.01 is around 30 times higher than that which would be expected by chance for AVLT-DR, but only 5–10 times higher than that which would be expected by chance for the other memory scores. This number has to exceed 20 for the FDR to be considered to be controlled at 5% when the data are thresholded at p=0.01, i.e., 20% of the map has to be significant at p=0.01. By contrast, the permutation tests establish a non-parametric distribution for the suprathreshold area. As seen in Table 3, it was rare (less than 1 in 20 occurrences) for the suprathreshold area in randomized (null) data to exceed that seen by chance. It has to be concluded that permutation tests at the 0.01 threshold were sometimes more powerful in detecting effects than an FDR test applied at the same threshold. Even so, FDR is adaptive and considers all possible thresholdings of the data, so some effects in MCI, and all effects in DAT, had critical values lower than 0.01. As such they passed FDR as there was some threshold that controlled the false discovery rate.
Our regression analyses assessing associations between the change in memory scores and change in hippocampal radius were not significant. This may be due to the relatively short follow-up interval of only 12 months. Over such a short period of time, the changes in measures of disease progression such as memory scores and hippocampal volumes or thickness can be quite noisy. shows that these measures show (1) very small mean changes over 12 month period, and (2) quite large standard deviations. In combination these make it hard to detect significant associations. We plan to repeat these analyses with the 2 or 3-year follow-up data once enough follow-up scans become available.
Once we controlled for baseline hippocampal volume in all subjects we uncovered highly significant regionally specific associations between follow-up memory test performance and hippocampal radial distance predominantly in CA1 distribution.
Most studies that have investigated the associations between memory performance and hippocampal atrophy to date have conducted their analyses on either a mixed samples of demented and nondemented subjects or in DAT subjects alone. Only a few studies have separated DAT and NC subjects in their analyses, and among those several reported a lack of or even a negative correlation between memory performance and hippocampal volume in cognitively normal elderly (Chantome et al., 1999
; Foster et al., 1999
; Kohler et al., 1998
; Ylikoski et al., 2000
). One recent study extracted hippocampal volumes from postmortem MR images and sought possible correlations between hippocampal volume and verbal memory performance within 1 year of death (Mortimer et al., 2004
). The authors reported a weaker association between memory performance and hippocampal volume in nondemented vs. demented individuals. These reports are in line with our findings. Even though the hippocampus plays a crucial role for memory encoding and consolidation in NC, age-associated memory decline depends less on hippocampal and more on extra-hippocampal (e.g., white matter) integrity (Grady, 2008
). Therefore, even though we found associations among the hippocampi and memory in the entire sample of controls, MCI, and DAT subjects, our data indicate the importance of analysing specific subgroups separately, as the correlations for the whole sample were driven by the MCI and DAT data.
The observed bilateral hippocampal/memory association in MCI is in agreement with our previous report (Apostolova et al., 2006b
). We also observed a pronounced left-sided laterality effect in DAT as already documented by de Toledo-Morrell et al (de Toledo-Morrell et al., 2000
). It has also been reported that atrophy of the left but not the right CA1 subfield was predictive of future progression from cognitively normal (CDR=0) to early DAT (CDR=0.5) (Csernansky et al., 2005b
). We do not have a good explanation for this bilateral to unilateral shift from MCI to DAT but one could perhaps hypothesize that in the MCI stage, subjects have greater compensatory capabilities and it takes bilateral disruption of the hippocampal networks for verbal memory impairment to manifest itself. As AD pathology increases and the ability to employ complex, higher order compensatory strategies dissipates (i.e., in the DAT stage), the associations between memory and the hippocampus may become detectable on the left side of the brain where the majority of language processing takes place. Another plausible explanation is that the substantially larger MCI group sample size gives us sufficient power to uncover right-sided verbal memory associations although admittedly our first report of bilateral hippocampal verbal memory associations in MCI came from a much smaller sample size (Apostolova et al., 2006b
To our knowledge this is the largest study investigating the correlations between verbal memory and hippocampal structural integrity. The strengths of this study are its large size, the detailed subject assessment, the unified MRI protocol across multiple sites and its meticulous data quality control. Additional strengths are the advanced preprocessing and 3D modelling techniques used to map discrete structural-functional correlations from normal aging to dementia. One limitation of the study stems from the limited generalizability. The ADNI study uses rigorous exclusion criteria, as it was designed to closely resemble a clinical trial population. As such, it does not necessarily represent the general elderly population and its findings should be generalized with caution. Another weakness is the etiologic/pathologic uncertainty in the MCI stage. At least 30% of amnestic MCI have been found to harbor non-AD pathology (Jicha et al., 2006b
). Such subjects if present in ADNI could be reducing our power to find statistically significant associations. Another limitation is that the link between an inward movement of the hippocampal surface and volumetric atrophy of the underlying subfield is yet to be validated. Such validation is difficult to do with 1.5T magnetic field strength. High-field and ultra-high field structural imaging can potentially provide us with enough resolution for subfield tracing and allow for such validation to take place.
In summary, our findings highlight the importance of the use of hippocampal atrophy as a biomarker by showing significant and robust relationships with the most pervasive symptom of AD, memory loss. Complex measures such as the AVLT-DR seem to be well suited for the MCI population and show strong association with hippocampal radial distance while DAT subjects may require less challenging verbal memory recall measures such as LM II-DR and ADAScog-DR.