Neuroimaging Insights into Subjective Cognitive Decline: Differential Sensitivity of Cognitive Change Index and Everyday Cognition Scale

2.1 Alzheimer’s Disease Neuroimaging Initiative

Data used in the preparation of this article were obtained from the Alzheimer’s Disease Neuroimaging Initiative (ADNI) database (adni.loni.usc.edu). The ADNI was launched in 2003 as a public–private partnership, led by Principal Investigator Michael W. Weiner, MD. The primary goal of ADNI has been to test whether serial magnetic resonance imaging (MRI), positron emission tomography (PET), other biological markers, and clinical and neuropsychological assessment can be combined to measure the progression of mild cognitive impairment (MCI) and early Alzheimer’s disease (AD). Participants were between 55 and 90 years old at the time of recruitment. The study received ethical approval from the review boards of all participating institutions. Written informed consent was obtained from participants or their study partner. Participants were selected from the ADNI-2 and ADNI-3 cohort because ADNI2 was the first cohort to introduce the CCI. In ADNI the CCI questionnaire was used to define participants with significant memory concerns. For consistency with current research standards, we use the term subjective cognitive decline.

2.2 Participants

ADNI participants were between 55-90 years of age and were included if they were cognitively normal at baseline. That is, they exhibited no evidence of memory decline, as measured by the Wechsler Memory Scale and no evidence of impaired global cognition as measured by the Mini Mental Status Examination (MMSE) or Clinical Dementia Rating (CDR). Participants were included if they completed both the CCI and ECog questionnaire, had an MRI completed within 6 months from the time of the questionnaires, from which WMHs, ventricle, hippocampal, and entorhinal cortex volumes could be extracted or had at least one of the cognitive tests available including the Montreal Cognitive Assessment (MOCA), Mini Mental Status Examination (MMSE), Alzheimer’s Disease Assessment Scale-13 (ADAS-13), or Rey Auditory Verbal Learning Test (RAVLT). A total of 334 cognitively healthy older adults (212 females, 122 males) in ADNI-2 and ADNI-3 who met the inclusion criteria were included in this study.

2.3 Subjective Cognitive Decline Questionnaires

The questions from the two questionnaires related to memory were extracted and examined. These questions are presented in Table 1. The ECog had 8 questions and the CCI had 12 questions related to memory functioning.

2.4 Structural MRI acquisition and processing

All scans were downloaded from the ADNI website (see http://adni.loni.usc.edu/methods/mri-tool/mri-analysis/ for the detailed MRI acquisition protocol). T1w scans for each participant were pre-processed through our standard pipeline including noise reduction (Coupe et al., 2008), intensity inhomogeneity correction(Sled et al., 1998), and intensity normalization into range [0- 100]. The pre-processed images were then linearly (9 parameters: 3 translation, 3 rotation, and 3 scaling) (Dadar et al., 2018) registered to the MNI-ICBM152-2009c average (Fonov et al., 2011).

2.5 WMH measurements

WMHs were generated using a segmentation technique that has been previously validated in ADNI (Dadar et al., 2019), and other multi-center studies such as the Parkinson’s Markers Initiative (Dadar et al., 2020) and the National Alzheimer’s Coordinating Center (Anor et al., 2021). Although FLAIR and T2w images are typically employed for segmentation of WMHs, we used T1w MR images because of their consistent availability across all cohorts and participants, allowing us to maximize the inclusion of the largest number of participants. Our previous work has determined that this T1w-based segmentation method holds strong correlations with the multicontrast T1w and T2w or FLAIR-based WMHs segmentations (r = 0.97, p<0.0001) and have similar relationships with clinical/cognitive scores as the multicontrast WMH segmentations (Dadar et al., 2019, 2018).

WMHs were automatically segmented using T1w contrasts along with a set of location and intensity features obtained from a library of manually segmented scans in combination with a random forest classifier to detect the WMHs in new images (Dadar et al., 2017b, 2017a). WMH load was defined as the volume of all voxels as WMH in the standard stereotaxic space (in mm3) and are thus normalized for head size. The volumes of the WMHs for frontal, parietal, temporal, and occipital lobes as well as the entire brain were calculated based on regional masks from the Hammers atlas (Dadar et al., 2017b; Hammers et al., 2003), and log-transformed to account for non- normality. The quality of the registrations and WMH segmentations was visually verified by an experienced rater (author M.D.), blinded to participants diagnostic group.

2.6 Freesurfer Measurements

T1w images were processed using FreeSurfer and quality controlled by the UCSF group, and regional GM volumes for the hippocampal, entorhinal cortices, and lateral ventricles were extracted. These volume estimates were collapsed across hemispheres. 1.5T and 3T data were processed with FreeSurfer versions 4.3 and 5.1, respectively, as appropriate.

2.7 Analysis

As multivariate analyses do not handle missing data well, before analysis, missing values were imputed using a principal components model and the missMDA R package (Josse and Husson, 2016). No variables had more than 10% of their values missing. The data were then analyzed using multivariate partial-least-squares (PLS). This method is a data-driven approach that combines elements of principal components analysis but goes further by maximizing the covariance between two data sets: one representing behavior or design (referred to as the Y matrix) and the other representing brain data (referred to as the X matrix). This analytical approach was based on the work of McIntosh & Lobaugh (2004) and Krishnan et al. (2011) (Krishnan et al., 2011; McIntosh and Lobaugh, 2004). PLS analyses were implemented in R version 4.3.1 using the Two-Table Exposition (TExPosition) package previous described in detail (Beaton et al., 2014). Thus, PLS can generate a set of orthogonal latent variables (LVs) that maximize the relationship between two data sets presented in descending order of explained cross-block covariance. Because PLS tests an entire pattern of relationships in a single step, there is no need to correct for multiple comparisons; instead, the significance of each LV was assessed by comparing the obtained result to a null distribution built with 1000 permutations. The reliability of the contributions of each of the variables to the LV was assessed with 1000 bootstrap repetitions, which were used to estimate standard errors. Salience values (i.e., the contributions for each variable) were divided by the bootstrapped standard errors to obtain bootstrap ratio (BSR) scores, which can be interpreted similarly to Z-scores. BSR scores exceeding thresholds of +/-2 were considered to reliably contribute to the LV. To examine the influence of sex and education on how SCD questions are endorsed we completed the PLS analyses again in males and females separately and then in those with high vs. low education (as determined by a median split).

We then used principal component analysis (PCA) to reduce the dimensionality across the MMSE, RAVLT forgetting, RAVLT learning, RAVLT immediate, and ADAS13 scores and produce a “memory score” from the first component. This memory component was then used as the outcome variable in two linear regressions where the predictors were the behavioral and brain scores from the PLS analysis. This analysis was completed to investigate the extent to which PLS-derived brain and SCD patterns can explain memory performance.

Receiver operating characteristic (ROC) analysis was performed to investigate the capability of the PLS LVs and PCA memory scores in predicting future cognitive decline. Cognitive decline was measured as a change of 0.5 points or greater between baseline and follow- up CDR-SB scores for one and two years following the baseline timepoint, based on which the LVs were calculated. Similar analyses were repeated with a threshold of 1 point increase in CDR- SB to assess the robustness of the results to the choice of cognitive decline threshold.

3.1 PLS Analysis Relating SCD to Brain Changes in Cognitively Normal Older Adults

The first PLS analysis relating SCD scores from the CCI, the ECog, and age with volumetric and white matter hyperintensity scores revealed two significant LVs, p’s < 0.001, which explained 85.89% and 8.30% of the cross-block covariance respectively (see Figure 1). The first LV captures a pattern where the CCI and increasing age load together along with a higher probability of WMH in all regions except occipital, larger lateral ventricle volume, and reduced gray matter volume in hippocampal and entorhinal cortices. The five questions from the CCI that loaded together with increasing age and greater neuropathology were questions 2, 6, 7, 10, and 12, which tend to have a more concrete focus (e.g., memory for faces, new information, objects, memory without notes, and a comparison to the person’s age group) than questions which did not load as strongly (e.g., recalling information, recalling conversations, and remembering what they intended to do).

• Download figure
• Open in new tab

Figure 1:

PLS analyses for full sample.

Notes: PT_Memory 1 through 8 are the ECOG questions. CCI1 through 12 are the CCI questions. PLS Analysis results for healthy older adults from the ADNI database and links subjective cognitive decline, as measured by the Cognitive Change Index (CCI) and Everyday Cognition (ECog) scores, with volumes of various brain regions, including white matter hyperintensity, hippocampus, entorhinal cortex, and ventricles. White matter hyperintensity volumes are specifically parcellated into Hammer’s Atlas lobes, encompassing frontal, parietal, occipital, and temporal regions. The visualization includes point-range plots showing the loadings, or brain scores, for each of the questions on each latent variable. Brain images within the figure highlight bootstrap ratio values, with areas exceeding absolute values of +/-2 outlined in black for emphasis. The color coding is critical for interpretation: blue-colored regions in the brain, as well as blue point-range values on the charts, indicate loadings in the same direction, whereas red-colored brain regions and point-range values signify loadings in the opposite direction. Brain images are displayed with NILEARN.

The second LV from this first analysis was associated with lower entorhinal gray matter volume in younger individuals with higher scores on CCI questions 3, 7, 9, and 11, suggesting that these questions (focusing on remembering the recent past, object location, names of family members and friends, and insight into how peers might assess their memory), might be particularly useful for detecting early decline associated with entorhinal cortex vulnerability.

3.2 Subgroup Analyses Examining Sex and Levels of Education

Two separate PLS analyses were fit to examine how the initial model was impacted by having a high or low education, determined by median split, or being male or female (see Figure 2). Each PLS model yielded a single significant LV, p’s < 0.001, and explained 86% and 76% of the cross- block covariance for education group and sex, respectively.

• Download figure
• Open in new tab

Figure 2:

PLS analyses grouped by education and sex.

Notes: PT_Memory 1 through 8 are the ECOG questions. CCI1 through 12 are the CCI questions. PLS analyses again comparing subjective cognitive decline with brain change; however, this time, the data were grouped by education (A) or sex (B). High versus low education level was determined by median split. Individuals with higher levels of education reliably expressed more white matter hyperintensities in the frontal regions than individuals with lower education, and this was part of a larger (non-reliable) pattern where more WMH and lower hippocampal volumes tended to be associated with greater SCD in the high education group. In Panel B, females who expressed higher levels of subjective cognitive decline and who were older expressed lower entorhinal cortex volumes.

In the education model (see Figure 2A), individuals who are older, and report higher levels of subjective cognitive decline on both the ECog and the CCI have a greater probability of white matter hyperintensities affecting the frontal lobes. A pattern of higher temporal lobe WMH burden and lower hippocampal volume was also associated with this LV, though the bootstrap ratio values for these regions were not reliably different from zero. Thus, more highly educated individuals may have greater insight into their own neural decline than individuals with less education.

In the sex model (Figure 2B), females who were older and had higher values of subjective cognitive decline, measured by either the ECog or the CCI, tended to have reduced volume in the entorhinal cortex. These findings indicate that report of SCD in females may be more strongly associated with structural brain changes than reports of SCD in males.

3.3 Brain-Behaviour Correlations

To relate the representativeness of the PLS LVs to behavior, we first extracted latent scores for each individual for the brain and behavioral matrices (i.e., brain and behavioral scores). These scores represent how strongly each individual loaded on the dimension being considered. Next, we used PCA to reduce the dimensionality of six measures of memory (MMSE, RAVLT forgetting, RAVLT learning, RAVLT immediate, MoCA, and ADAS13) to derive a “memory component” (see Figure 3). The first dimension of the PCA explained 39% of the covariance and was primarily driven by the MOCA, ADAS13, and RAVLT immediate scores.

• Download figure
• Open in new tab

Figure 3:

Brain-behaviour correlations with the first latent variable.

Notes: The brain-behavior correlations observed between the first latent variable (LV1) scores from Partial Least Squares (PLS) analysis of both brain and behavior data, and a composite memory measure. This composite memory measure is derived from a Principal Component Analysis (PCA) of various cognitive assessments, including the Mini-Mental State Examination (MMSE), Rey Auditory Verbal Learning Test (RAVLT) focusing on Forgetting, Learning, and Immediate recall scores, the Montreal Cognitive Assessment (MOCA), and the Alzheimer’s Disease Assessment Scale-Cognitive Subscale-13 (ADAS-13). The first dimension in this analysis captured 39% of the covariance between these variables. Notably, positive scores on this dimension are indicative of better cognitive performance.

We predicted memory performance (using the first component score) with two PLS models using the brain and behaviour scores from the first PLS analysis (see Figure 3). The first model, predicting memory performance from brain structural scores, was significant but only accounted for 7% of the variance in memory performance (R² = 0.07, F(1, 332) = 24.02, p < .001, adjusted R² = 0.06). The positive and significant effect of brain score on memory performance was quantified with a β coefficient of 0.40 (95% CI [0.24, 0.56], t(332) = 4.90, p < .001), indicating that as brain score increases (i.e., a person expresses a lower probability of white matter hyperintensities and greater hippocampal and entorhinal cortex volume), so does memory performance, albeit to a modest extent.

The second model shifted focus towards behavioral scores, presenting a more robust relationship with the memory component. This model accounted for a moderate 13% of the variance (R² = 0.13, F(1, 332) = 51.68, p < .001, adjusted R² = 0.13). Similar to the first model, the influence of the behavioural score was both positive and statistically significant (β = 0.56, 95% CI [0.41, 0.72], t(332) = 7.19, p < .001), suggesting that younger individuals with fewer subjective memory complaints had stronger memory scores.

3.4 ROC Analysis

Eighty-seven and 230 individuals had completed one and two year clinical follow up assessments, respectively. Out of those, 15 (17.24%) and 45 (19.6%) individuals experienced cognitive decline in year one and two follow up visits. Figure 4 shows the performance of the first two brain and behaviour LVs as well as the memory PCs in predicting future cognitive decline at years one and two, respectively. The first memory PC was the best differentiator of future cognitive decline, both at year 1 (AUC = 0.80) and year 2 (AUC = 0.75). Brain and behaviour LV scores were also able to differentiate the individuals that demonstrated cognitive decline in follow visits from those that remained stable, albeit to a lesser extent. The best differentiations were observed for LV2 brain score and LV1 behaviour scores at year 1 (AUC = 0.63) and LV1 and LV2 brain scores at year 2 (AUCs = 0.62 and 0.66). Overall, brain scores were better predictors of future cognitive decline (all AUCs > 0.61). Similar results were obtained for a cognitive decline threshold of 1 point increase in CDR (Year 1: N decline= 5; AUC PC1-memory=0.92; AUC LV1-brain=0.78; AUC LV2-brain=0.62, and Year 2: N decline= 22; AUC PC1-memory=0.82; AUC LV1-brain=0.64; AUC LV2-brain=0.65).

• Download figure
• Open in new tab

Figure 4.

ROC analysis

Note: Results are comparing the AUC of PLS LVs in predicting future cognitive decline at one (left panel) and two (right panel) year follow up timepoints.

3.5 Secondary Exploratory Analysis

A secondary analysis was also performed with the data regression out age before completing the PLS analysis. This analysis was completed to determine the relationship independent of age. However, it should be noted that both models are important to consider because SCD is strongly associated with age.

In this secondary analysis, one LV was significant, p < 0.01 explained 24% of the cross- block covariance. For this LV, 8/12 (67%) CCI variables, but none of the ECoG scores, negatively covaried with entorhinal cortex volume. This finding suggests that above and beyond the effect of age, the CCI is sensitive to early abnormal changes in a region highly sensitive to changes observed in early AD.

Group analyses by education revealed two significant LVs, p’s < 0.01, which explained 39% and 24% of the cross-block covariance, respectively. The first LV showed that people with higher education tended to have stronger CCI responses (8/12 CCI scores loaded positively and reliably for this group), than people with less education, (1/12 CCI scores loaded positively for this group). As above, positive scores indicating higher subjective memory complaints were associated with reduced entorhinal cortex volume. The second significant LV was associated with reliably higher frontal and temporal WMH load (BSR values > 2), with generally higher CCI scores in the more educated groups, though these values did not exceed +2, indicating they were less reliable.

After removing the effect of age, re-running the analysis grouped by sex revealed two significant, ps < 0.01, LVs that accounted for 65% and 41% of the cross-block covariance, respectively. The first LV revealed that males who had higher CCI scores on questions 1 and 2 and higher ECoG scores on questions 4, 7, and 8 were more likely to have higher WMH burden in frontal, parietal, and occipital lobes, along with larger ventricular volumes. The second LV revealed that CCI scores were reliably associated with entorhinal cortex volume loss in both males (67% of questions) and females (50% of questions).