2.1 Participants
We analyzed data from two cohorts, INSIGHT-PreAD and ADNI. The INSIGHT-PreAD cohort [34] is a French mono-centric cohort of cognitively-normal elderly memory complainers, longitudinally followed at the Pitié-Salpêtrière Hospital (Paris). At baseline, the sample included 318 participants aged between 70 and 85 years old, with Mini-Mental State Examination (MMSE) [35] ≥ 27/30, Total Recall of the Free and Cued Selective Reminding Test (FCSRT) ≥ 41/48 and Clinical Dementia Rating (CDR) = 0.
For the present study, some participants were excluded from the original sample (i.e. one participant due to missing metabolic imaging, one due to missing memory scores, 24 due to missing questionnaires). In addition, two outliers (i.e. one on a memory score, the other on brain metabolism) were also removed from the sample. Our final sample included 290 participants from the INSIGHT-PreAD cohort.
The Alzheimer’s Disease Neuroimaging Initiative (ADNI, http://adni.loni.usc.edu) is a multicentric longitudinal study. For this study, we aimed at including only ADNI participants with SCD (i.e. tagged as Significant Memory Concern), in order to be fully comparable to INSIGHT-PreAD participants. These are subjects with normal cognition (MMSE ≥ 24/30; Logical Memory Delayed Recall in standards, CDR = 0) and memory concerns not supported by the informant (Cognitive Change Index, sum of the first 12 items > 12/16) [36]. We identified 277 participants meeting these criteria. We excluded the participants with missing data, and a total of 158 participants from ADNI cohort were retained in the final sample.
For each cohort, we used only baseline data. Our final sample therefore consisted of 448 CN participants with SCD.
2.2 Development of the MMR
2.2.1 Objective Memory Assessment
In this study, we have chosen to focus on memory. Indeed, recent research tends to show that the subtle decline in memory occurring in preclinical AD would be among the earliest to evidence of a transition to a subsequent prodromal AD [1,5–8].
Therefore, we selected three episodic memory scores: the Free and Cued Selective Reminding Test (FCSRT) [37], the Delayed Matched to Sample test 48 items (DMS48) [38] and the Rey-Osterrieth Complex Figure (ROCF) [39]. For the FCSRT, we selected both immediate and delayed Total Free Recalls (FR) and Total Recalls (i.e. FR + Cued Recalls; TR), the number of intrusions and perseverations. For both visual tests (i.e. DMS48 and ROCF), we used immediate and delayed memory measures. Details of the neuropsychological examination proposed in the INSIGHT-PreAD cohort were previously described [34].
For the ADNI cohort, we selected three memory tests, namely the Rey Auditory Verbal Learning Test (RAVLT, “immediate”, “forgetting” and “learning” scores) [40], the Logical Memory II (LMII) test from the Wechsler Memory Scale [41], and the Q4 (memory) score from the Alzheimer Disease Assessment Scale (ADAS-Cog) [42].
2.2.2 Subjective Cognitive Assessment
For INSIGHT-PreAD cohort, we used the cognitive subscale of the Healthy Aging Brain Care – Monitor questionnaire [43]. This questionnaire asks subjects to rate the frequency of occurrence of certain cognitive disturbances during the last two weeks (i.e. from 0,‘not at all (0-1 day)’, to 3,‘almost daily (12-14 days)’). The HABC-M cognitive scale consists of 6 items, the majority of which are related to the memory domain. The total score ranges from 0 to 18.
For the ADNI cohort, we used the memory subscale of the Everyday Cognition questionnaire score [44]. These questions asks the participant to compare his/her current memory ability in everyday tasks to that of ten years ago. The estimate is based on a 4-point scale from 1 (‘Better or no change’) to 4 (‘Consistently much Worse’). A ‘Don’t know’ answer is also possible. The total score then ranges from 8 to 32. Higher scores indicate that the participant perceives a more marked cognitive decline.
2.2.3 The Meta-Memory Ratio
We based our measure of ACD on the model of the “anosognosia index” initially proposed by Dalla Barba and collaborators [45]. This procedure consists in measuring a gap between the subjective complaint and an objective performance. To compute the score, the same procedure was implemented independently in each cohort (Figure 1).
First, since the two samples had different demographic distributions and these variables can be associated with the scores of interest, we started by removing their impact on the scores. Each score of interest (i.e. memory performances and complaint questionnaires) was integrated into a generalized linear model (GLM), as a dependent variable. Demographic variables (i.e. age, gender and socio-cultural level) were included as covariates, to correct for their potential effects. For each measure, the type of model used was selected according to the distribution of each score (i.e. linear regressions for ROCF, ECog, immediate RAVLT, and ADAS-Q4; logistics for FCSRT intrusions and perseverations; and binomial for the other measures mentioned). Subsequently, we extracted the model residuals to obtain objective and subjective measures of decline net of these effects.
Secondly, we have centered and reduced all the residuals in order to make them comparable (i.e. z-score transformation).
Thirdly, we computed a composite score by averaging all memory scores collected for each subject. In this way, we had two values. The first one represented an objective measure of memory function. The higher this score, the better the memory performance at testing. The second one represented a subjective measure of memory. The higher this score, the higher the memory complaint.
The choice of relying on a composite score rather than choosing a single memory score addressed two needs. To begin with, it allowed us not to select a certain score a priori. In addition, the use of a composite score also allowed to gather variables that are thought to measure the same cognitive construct [46]. We would have used this procedure also for complaint measures, if the cohorts had included more than one questionnaire.
Finally, we added these two scores. By construction, an MMR close to 0 corresponds to a good match between subjective rating and objective performance (i.e. accurate ACD). The higher the MMR is, the more it corresponds to an SCD (i.e. important complaint with correct performance). On the contrary, the lower it is, the more the ACD is low (i.e. lower complaint associated with a poorer performance; Figure 1).
2.3 Brain imaging acquisition and processing
2.3.1 Amyloid PET Imaging
In the INSIGHT-PreAD cohort, participants underwent PET with a florbetapir tracer [18F-Florbetapir, AmyvidTM, Avid Radiopharmaceuticals]. A standardized uptake value ratio (SUVr) was calculated with the CATI pipeline (Centre d’Acquisition et de Traitement d’Images, https://cati-neuroimaging.com), with a focus on selected target regions (i.e. bilateral precuneus, anterior and posterior cingulum, temporal cortex and orbitofrontal). Details of the imaging procedure and threshold calculations have previously been presented [34,47]. For ADNI participants, we selected those with SUVr values calculated using the same radiotracer (i.e. florbetapir). The details of both imaging procedures are presented in supplementary materials.
To make the two cohorts neuroimaging features comparable, we normalized the SUVr using the respective amyloid positivity threshold of each cohort. To do so, we divided the SUVr of each participant by the positivity threshold, i.e. 0.79 for the INSIGHT-PreAD cohort and 1.11 for ADNI. Thus, any normalized SUVr above 1 could be considered significantly pathological (i.e. amyloid-positive patients).
2.3.2 FDG-PET Imaging
For each of our cohorts, we calculated a mean metabolism index using Fluorodeoxyglucose Positon Emission Tomography (FDG-PET) by averaging the regions of interest (ROIs) of AD, namely the posterior cingulate cortex, inferior parietal lobule, precuneus and inferior temporal gyrus [48]. Then, since the FDG-PET did not have an established cut-off, we normalized this meta-ROI using a centered-reduced method (i.e. intra-cohort z-score transformation). As for amyloid PET imaging, details are available in supplementary materials.
2.4 Statistical Analysis
The different scores of interest (MMR, complaint and memory) and demographic variables were compared between the two cohorts using Welch’s t-tests for the numerical variables, and a Chi2-test for the gender variable.
MMR scores were normally distributed. In order to evaluate the influence of AD biomarkers on awareness, we computed a linear regression model with the MMR as dependent variable. To account for the specific effect of each biomarker, amyloidosis (AV45-PET) and metabolism (FDG-PET) were both included in the model. We also included interactions between the “cohort” effect and biomarkers to determine whether the effect of biomarkers varied across cohorts. Finally, we adjusted the results including demographic (i.e. age, gender and education) and the cohort variables as covariates.
Looking at the scatterplot between MMR and AV45-SUVr, we identified a non-linear effect of amyloid on the MMR. Therefore we added a quadratic effect of amyloid to the models. The main effects and interactions (both with linear and quadratic effect) were tested via the likelihood ratio test type II. Normality of residuals and heteroskedasticity were checked visually. Cook’s distances and hat values were computed to investigate potential influencers and outliers. We also performed these computations with an additional group of cognitively normal (CN, that is normal cognition without cognitive complaint) from ADNI without anosognosia (data not shown). Finally, the same analysis was performed with an MMR calculated from a single rather than a composite memory score. These results can be found in supplementary material (Additional File 3).
Statistical analyses were performed using R 3.5.2 (https://www.R-project.org/). An R package was developed for the calculation of MMR in various cohorts (https://github.com/GagGeo/MMAD).