INTRODUCTION
Episodic memory impairment is a hallmark symptom commonly seen during the early stages of Alzheimer’s disease (AD; Dubois et al., Reference Dubois, Feldman, Jacova, Dekosky, Barberger-Gateau, Cummings and Scheltens2007). This disorder likely results from encoding and consolidation deficits (Dubois et al., Reference Dubois, Feldman, Jacova, Dekosky, Barberger-Gateau, Cummings and Scheltens2007) and is thought to be due to medial temporal lobe damage, particularly affecting the hippocampus (Sarazin et al., Reference Sarazin, Chauvire, Gerardin, Colliot, Kinkingnehun, de Souza and Dubois2010). While this cognitive feature of AD has long been recognized, current diagnostic criteria (Albert et al., Reference Albert, DeKosky, Dickson, Dubois, Feldman, Fox and Phelps2011) do not recommend specific methods for its evaluation. Recent studies have nevertheless suggested that the Free and Cued Selective Reminding Test (FCSRT; Grober & Buschke, Reference Grober and Buschke1987; Grober, Sanders, Hall, & Lipton, Reference Grober, Sanders, Hall and Lipton2010) is sensitive and specific for diagnosing prodromal AD (Saka, Mihci, Topcuoglu, & Balkan, Reference Saka, Mihci, Topcuoglu and Balkan2006; Sarazin et al., Reference Sarazin, Berr, De Rotrou, Fabrigoule, Pasquier, Legrain and Dubois2007).
However, the most effective FCSRT measures for evaluating AD-related episodic memory impairment are unclear. For instance, total recall (TR; i.e., sum of items recalled across free and cued-recall trials) was superior to free recall (FR; i.e., sum of items recalled across FR trials) for discriminating AD from non-AD dementia patients (Grober et al., Reference Grober, Sanders, Hall and Lipton2010; Pillon et al., Reference Pillon, Deweer, Michon, Malapani, Agid and Dubois1994). However, Saka et al. (Reference Saka, Mihci, Topcuoglu and Balkan2006) demonstrated that, while both FR and TR successfully differentiated AD patients from controls, only the third FR trial accurately discriminated patients with mild cognitive impairment (MCI) from controls. Other studies identified FR as the best measure for predicting dementia (Grober, Lipton, Hall, & Crystal, Reference Grober, Lipton, Hall and Crystal2000; Grober et al., Reference Grober, Sanders, Hall and Lipton2010). Finally, Sarazin et al. (Reference Sarazin, Berr, De Rotrou, Fabrigoule, Pasquier, Legrain and Dubois2007) demonstrated that FR and TR had comparable specificity while TR had the highest sensitivity for discriminating prodromal AD patients from MCI patients who did not evolve to AD within 3 years.
Importantly, these studies focused primarily on groups defined by clinical diagnostic criteria (American Psychiatric Association, 2000; McKhann et al., Reference McKhann, Drachman, Folstein, Katzman, Price and Stadlan1984; Petersen, Reference Petersen2004), without verifying pathophysiological processes associated with MCI or AD. Furthermore, because clinical diagnosis requires objective memory deficits confirmed by neuropsychological testing, analyses of diagnostically determined groups defined by memory-related scores may be biased (Carlesimo, Perri, & Caltagirone, Reference Carlesimo, Perri and Caltagirone2011). Thus, studying memory processes in incipient AD clearly calls for patient classification using the presence of underlying pathology rather than on the basis of objective memory deficits.
Recent progress in the development of in vivo biomarkers presumed to mirror AD underlying pathology may address this issue. These biomarkers include indexes of abnormal amyloid-β (Aβ) and tau protein deposition in the brain, which track underlying AD pathology directly, and surrogate neurodegeneration markers reflecting, for example, temporo-parietal hypometabolism or medial temporal atrophy. Neurodegeneration markers provide only indirect evidence of AD because they are assumed to track sequelae of protein deposition (Albert et al., Reference Albert, DeKosky, Dickson, Dubois, Feldman, Fox and Phelps2011; Trojanowski et al., Reference Trojanowski, Vandeerstichele, Korecka, Clark, Aisen and Petersen2010). Hence, current pathological criteria (Albert et al., Reference Albert, DeKosky, Dickson, Dubois, Feldman, Fox and Phelps2011) require a positive amyloid marker for diagnosing AD with an intermediate or high likelihood. These biomarkers are abnormal in a considerable proportion of MCI and AD patients (e.g., Frisoni et al., Reference Frisoni, Prestia, Zanetti, Galluzzi, Romano, Cotelli and Geroldi2009; Ivanoiu et al., Reference Ivanoiu, Dricot, Gilis, Grandin, Lhommel, Quenon and Hanseeuw2015; Ong et al., Reference Ong, Villemagne, Bahar-Fuchs, Lamb, Chetelat, Raniga and Rowe2013) and appear years before cognitive symptoms (Jack et al., Reference Jack, Knopman, Jagust, Petersen, Weiner, Aisen and Trojanowski2013).
Despite the significant interest in using the FCSRT for AD diagnosis, few studies have investigated relationships between FCSRT performance and AD biomarkers. Rami et al. (2011) reported that TR was correlated with cerebrospinal fluid (CSF) levels of Aβ in persons with subjective memory complaints (SMC) and CSF t-tau and p-tau in MCI patients. When persons with SMC were classified into subgroups according to their CSF Aβ levels, the Aβ positive group performed lower on several FSCRT scores (i.e., learning, total learning, total recall) than the Aβ negative group. Moreover, both FR and TR were significantly impaired in 88% of prodromal and mild AD patients presenting with memory complaints and who were Aβ+(Xie et al., Reference Xie, Gabelle, Dorey, Garnier-Crussard, Perret-Liaudet, Delphin-Combe and Krolak-Salmon2014). Furthermore, Wagner et al. (Reference Wagner, Wolf, Reischies, Daerr, Wolfsgruber, Jessen and Wiltfang2012) demonstrated that TR was superior to FR for distinguishing patients with a low CSF Aβ/tau ratio (compatible with AD) from persons with a normal ratio.
Most of these studies used TR in addition to FR in their analyses. TR is, however, composed of both FR and cued recall and, therefore, does not permit an unambiguous assessment of the cueing effect, which typically differentiates AD patients from controls or patients with non-AD pathologies, even in the early disease stages (Dubois et al., Reference Dubois, Feldman, Jacova, Dekosky, Barberger-Gateau, Cummings and Scheltens2007). Furthermore, TR does not capture as precisely as cueing efficiency (CE) measures the associative aspects of episodic memory performance, which are particularly impaired in early AD (Atienza et al., Reference Atienza, Atalaia-Silva, Gonzalez-Escamilla, Gil-Neciga, Suarez-Gonzalez and Cantero2011; Hanseeuw et al., Reference Hanseeuw, Dricot, Kavec, Grandin, Seron and Ivanoiu2011; Sperling, Reference Sperling2007; Troyer et al., Reference Troyer, Murphy, Anderson, Craik, Moscovitch, Maione and Gao2012, Reference Troyer, Murphy, Anderson, Hayman-Abello, Craik and Moscovitch2008).
Therefore, we first investigated whether CE measures derived from FCSRT performance can discriminate between participants classified according to the presence or absence of significant amyloid burden using [F18]-flutemetamol positron emission tomography (PET), compared to the typically used FCSRT FR measures. We hypothesized that the former would be superior to FR for discriminating between amyloid positive and negative individuals.
Second, using surface-based morphometry (SBM) to quantify cortical thickness (CT) across the entire brain surface, we explored whether FCSRT CE and FR indexes were associated with the integrity of different anatomical substrates. Previous studies have focused on the cortical correlates of a single composite memory score (Dickerson, Feczko, et al., Reference Dickerson, Feczko, Augustinack, Pacheco, Morris, Fischl and Buckner2009; Nho et al., Reference Nho, Risacher, Crane, DeCarli, Glymour and Habeck2012) or contrasted only immediate versus delayed recall measures (Ahn et al., Reference Ahn, Seo, Chin, Suh, Lee, Kim and Na2011). Assuming that CE measures may be more specific than FR measures to underlying AD-related pathology, we predicted that the former scores would rely on cortical regions that are the most susceptible to atrophy early in AD.
Finally, we investigated whether FCSRT scores can be predicted by biomarkers of AD-related neuropathology (amyloid load and hippocampal volume) and CT of selected regions of interest (ROIs). The relative contributions of each biomarker to prediction of FCSRT performance could reveal whether specific indexes are differentially sensitive to distinct aspects of the disease process. We hypothesized that AD biomarkers would account for a greater proportion of variance in CE measures than in FR measures, and that thickness of cortical regions affected most in early AD would account for additional unique variance in CE.
METHODS
Participants
Participants included 31 control subjects and 60 non-demented patients. The recruitment and assessment procedures have been previously published (Ivanoiu et al., Reference Ivanoiu, Dricot, Gilis, Grandin, Lhommel, Quenon and Hanseeuw2015). The study received Ethical Committee approval (Eudra-CT number: 2011-001756-12). Each participant provided informed consent.
The patient group included persons who: (1) attended the Memory Disorders Clinic of Saint-Luc Hospital in Brussels for memory complaints, (2) presented globally preserved daily living skills, and (3) did not meet DSM-IV-TR criteria for dementia (American Psychiatric Association, 2000). To avoid circularity in the analyses, memory performance was not used for differentiating patients from controls. Patients represented persons presenting memory complaints who were further classified into two subgroups, depending on whether their amyloid load exceeded a threshold (see the Amyloid Imaging section below). We identified 32 amyloid negative patients (Aβ− P) and 28 amyloid positive patients (Aβ+P). Older controls were recruited by advertisements and were included if they denied memory complaints and their amyloid status was negative (Aβ− C). Three controls were excluded because they were amyloid positive, resulting in 28 older controls. Participants with substance abuse or a known neurological or psychiatric condition were excluded.
FCSRT
As part of a larger neuropsychological evaluation, participants were administered the French word version of the FCSRT using the standardized instructions and including an immediate cued recall phase (Van der Linden et al., Reference Van der Linden, Coyette, Poitrenaud, Kalafat, Calicis and Wyns2004). FR and CE measures were computed for (1) the first recall trial, which is less influenced by systematic feedback, relearning, and ceiling effects than subsequent trials, making it potentially sensitive to AD-related memory changes (Wagner et al., Reference Wagner, Wolf, Reischies, Daerr, Wolfsgruber, Jessen and Wiltfang2012), (2) the cumulative sum of the three recall trials, which is a measure of learning, and (3) delayed recall, which is considered to be the most sensitive memory measure for detecting AD (Albert, Reference Albert1996; Chen et al., Reference Chen, Ratcliff, Belle, Cauley, DeKosky and Ganguli2000).
FR scores measured the number of correctly recalled words in all these phases, ranging from 0 to 16 for the first FR (FR1) and delayed FR trials (DFR), and from 0 to 48 for the total number of recalled words over the three FR trials (i.e., cumulative FR; CFR). Cueing efficiency was assessed using a CE index calculated as (TR1−FR1)/(16−FR1) for the first recall trial, (cumulative total recall−CFR)/(48−CFR) for cumulative recall (CE-CR), and (delayed TR−DFR)/(16−DFR) for delayed recall (CE-DR). Contrary to most previous studies, we focused on CE measures instead of TR as these latter measures are composed of both free- and cued-recall performance. CE indexes provide a purer evaluation of the cueing effect, which is particularly vulnerable in AD.
CE-DR could not be computed for four Aβ−C and one Aβ−P because they freely recalled every word (i.e., denominator=0). DFR measures, and CE-DR scores were not available for two Aβ−C and one Aβ+P due to technical problems.
Brain Imaging
Amyloid imaging
Amyloid status was assessed using a [F18]-flutemetamol PET examination. [F18]-Flutemetamol is an investigational medicinal product being studied clinically as an amyloid imaging agent (for further detail, see Ivanoiu et al., Reference Ivanoiu, Dricot, Gilis, Grandin, Lhommel, Quenon and Hanseeuw2015). Quantitative standard uptake value ratios (SUVr) were computed using PVIEW and PFUS v3.2.2 software modules (Fischl et al., Reference Fischl, Salat, van der Kouwe, Makris, Segonne, Quinn and Dale2004). Amyloid status was considered as negative (Aβ−) when SUVr was less than a threshold of 1.622, and positive (Aβ+) otherwise. This threshold is consistent with previously reported cutoff values (Jack et al., Reference Jack, Lowe, Senjem, Weigand, Kemp, Shiung and Petersen2008; Vandenberghe et al., Reference Vandenberghe, Van Laere, Ivanoiu, Salmon, Bastin, Triau and Brooks2010; Villemagne et al., Reference Villemagne, Burnham, Bourgeat, Brown, Ellis and Salvado2013, Reference Villemagne, Pike, Chetelat, Ellis, Mulligan, Bourgeat and Rowe2011).
Hippocampal volume and cortical thickness measurements
Hippocampal volume (HV) and CT were evaluated using three-dimensional (3D) T1-weighted images recorded at 3 Tesla (Achieva, Philips Healthcare, Eindhoven, The Netherlands) with a 32-channel phased array head coil. A gradient echo sequence with an inversion prepulse (Turbo Filed Echo) was acquired using the same parameters as in Ivanoiu et al. (Reference Ivanoiu, Dricot, Gilis, Grandin, Lhommel, Quenon and Hanseeuw2015). Whole-brain segmentation and CT measures were completed using the FreeSurfer image analysis pipeline (FreeSurfer, Martinos Center for Biomedical Imaging, Boston, MA). Left, right, and mean HVs normalized to total intracranial volume were computed for each participant.
Analyses
Intergroup analyses
Group differences in demographic characteristics were examined using an analysis of variance for age, and chi-squared tests for gender and education. Separate Kruskal-Wallis tests evaluated possible group differences on the Mini-Mental State Examination (MMSE) total score, each biomarker measure and FCSRT score. Non-parametric tests were used because the distributions of these latter variables were significantly non-normal, and the homoscedasticity was not systematically met.
These statistical analyses were performed using IBM SPSS Statistics software (Version 22.0).
Analysis of the cortical substrates of the FCSRT measures
We used SBM analyses to explore the relationships between FCSRT CE and FR scores and CT across the entire brain surface by implementing the FreeSurfer analysis pipeline. We performed general linear models for the sample as a whole using each FCSRT score as a covariate and controlling for the effects of factors that may influence CT in three steps: (1) age (e.g., Hurtz et al., Reference Hurtz, Woo, Kebets, Green, Zoumalan, Wang and Apostolova2014; Lemaitre et al., Reference Lemaitre, Goldman, Sambataro, Verchinski, Meyer-Lindenberg, Weinberger and Mattay2012; Thambisetty et al., Reference Thambisetty, Wan, Carass, An, Prince and Resnick2010); (2) age and the amyloid load (e.g., Becker et al., Reference Becker, Hedden, Carmasin, Maye, Rentz, Putcha and Johnson2011; Llado-Saz, Atienza, & Cantero, Reference Llado-Saz, Atienza and Cantero2015); and (3) age, the amyloid load, and the mean normalized HV. These analyses were also performed using either the left or right normalized HV. As the results were very similar to the results obtained with the mean HV, we reported these latter results. Because the focus of this analysis was to investigate the unique relationships between FCSRT scores and CT independently of potential nuisance factors, we only reported the final analysis controlling for age, amyloid load, and mean normalized HV. The results of the two other models are reported in the Supplemental Material. Data were corrected for multiple comparisons using the false discovery rate (FDR) correction method at a q<.05 level of significance.
Predictive analyses
Using R (version 3.3.2; R Development Core Team, 2015), we performed a series of hierarchical multiple regression analyses using each FCSRT score as the dependent variable and hierarchical sets of independent variables: (1) age; (2) age and amyloid burden; (3–4) age, amyloid burden, and normalized left or right HV; (5–6) age, amyloid burden, normalized left or right HV, and each ROI defined according the Desikan-Killiany atlas (Desikan et al., Reference Desikan, Segonne, Fischl, Quinn, Dickerson, Blacker and Killiany2006) and in which SBM analyses identified clusters that were significantly related to each FCSRT score. We designed this hierarchical regression approach to study biomarkers reflecting one view of the progression of AD pathology throughout the disease course (Jack et al., Reference Jack, Knopman, Jagust, Petersen, Weiner, Aisen and Trojanowski2013). Because amyloid deposition is one of the furthest upstream processes, amyloid load was entered as the first predictor of FCSRT scores. HV and CT were subsequently entered, as HV is thought to be affected downstream of amyloid deposition, followed later by cortical thinning (Braak & Braak, Reference Braak and Braak1991; Jack et al., Reference Jack, Knopman, Jagust, Petersen, Weiner, Aisen and Trojanowski2013). We used the left or right HV separately in steps 3–6 as we expected stronger relationships with the left rather than the right HV because the FCSRT is a verbal memory test. Squared semi-partial correlation coefficients were computed to identify the unique proportion of variance in each FCSRT index accounted for by each predictor. The number of analyses was limited to the number of significant ROIs identified in the SBM analysis. Rather than evaluating statistical significance, we here focused on identifying predictors that could uniquely account for at least 4% of the FCSRT score variance, which represents a threshold for a “practically” meaningful effect size (Ferguson, Reference Ferguson2009). All predictors meeting this criterion were nevertheless statistically significant (p<.05).
Given our sample size, we were unable to perform separate training and test samples to determine the generalizability of our prediction equations to external samples. Therefore, in addition to presenting R 2 adjusted values, we computed a predicted R 2 using the PRESS (Predicted Residual Sum of Squares) statistic (Stevens, Reference Stevens2002) for each of the regressions involving cortical thickness. This predicted R 2 is considered to be a good index of generalizability to external samples (Myers, Reference Myers1990; Stevens, Reference Stevens2002). As it is essentially equivalent to “jackknifing,” or LOO (leave-one-out) cross-validation approach, the prediction error for each participant is determined on the basis of the equation using the remaining participants. The PRESS statistic is computed by summing these squared prediction errors for each subject. A predicted R 2 statistic, which reflects the expected performance of the regression equation in a new sample, can then be computed using the following formula (Stevens, Reference Stevens2002):
$$Predicted\,R^{{\rm 2}} \,{\equals}\,1\,{\minus}\,{{PRESS} \over {\mathop{\sum}{\left( {y_{i} {\minus}\bar{y}} \right)} ^{2} }}$$
RESULTS
Demographic Characteristics and Biomarker Values
There was no significant group effect for age, F(2,85)=1.6, p=.208, ω=.12, gender, χ2(2)=.61, p=.739, nor education, χ2(4)=1.7, p=.423 (Table 1). The MMSE total score significantly differed between groups, H(2)=33.98, p<.001. Pairwise comparisons with adjusted p-values indicated that the MMSE was significantly lower in Aβ+P than in Aβ−P (p=.04; r=.28) and Aβ−C (p<.001; r=.77), and in Aβ−P compared to Aβ−C (p<.001; r=.49). Biomarker values significantly differed between groups (all p-values<.001; Table 1). As expected, the amyloid load was higher in Aβ+P than in Aβ−P (p<.001; r=−.85) and Aβ−C (p<.001; r=−.87), but did not differ significantly between amyloid negative groups (p=1.00; r=−0.02). The same pattern was found for the mean, left and right HV (i.e., mean VH: Aβ+P vs. Aβ−P, p<.001, r=.56, Aβ+P vs. Aβ−C, p<.001, r=.65, Aβ−P vs. Aβ−C, p=1.00, r=.08; left HV: Aβ+P vs. Aβ−P, p<.001, r=.52, Aβ+P vs. Aβ−C, p<.001, r=.68, Aβ−P vs. Aβ−C, p=.669, r=.16; right HV: Aβ+P vs. Aβ−P, p<.001, r=.55, Aβ+P vs. Aβ−C, p<.001, r=.58, Aβ−P vs. Aβ−C, p=1.00, r=.02).
Demographic characteristics and biomarker values
a Level 1=less than 12 years of education; Level 2=more than 12 years of education.
b The amyloid load represents the quantitative standard uptake value ratios (SUVr).
HV=hippocampal volume normalized to the ICV.
Group Differences for Each FCSRT Measure
Kruskal-Wallis tests showed significant group differences for each FCSRT measure (all p-values<.001; Table 2). Pairwise comparisons with adjusted p-values highlighted that FR1 was significantly lower in the two patient groups than in Aβ−C (i.e., Aβ+P vs. Aβ−C, p<.001, r=.77; Aβ−P vs. Aβ−C, p<.001, r=.54). However, FR1 did not differ significantly between patient groups, p=.242, r=.23. CFR was significantly lower in Aβ+P than in the two amyloid negative groups (i.e., Aβ+P vs. Aβ−C, p<.001, r=.80; Aβ+P vs. Aβ−P, p=.006, r=.40), and in Aβ−P relative to Aβ−C, p=.006, r=.40. The same pattern was observed for DFR (i.e., Aβ+P vs. Aβ−C, p<.001, r=.77; Aβ+P vs. Aβ−P, p=.003, r=.43; Aβ−P vs. Aβ−C, p=.031, r=.34).
Descriptive statistics (i.e., median, quartiles 25 and 75) and group differences for each FCSRT score
FR1=free recall 1; CFR=cumulative free recall; DFR=delayed free recall; CE-R1; cue efficiency in recall 1; CE-CR=cue efficiency for cumulative recall; CE-DR=cue efficiency in delayed recall; TR-R1=total recall in recall 1; TR-CR=total recall for cumulative recall; TR-DR=total recall in delayed recall.
Each CE measure was significantly lower in Aβ+P than in the amyloid negative groups but did not differ significantly between the amyloid negative groups (i.e., CE-R1: Aβ+P vs. Aβ−C, p<.001, r=.59, Aβ+P vs. Aβ−P, p=.011, r=.38, Aβ−P vs. Aβ−C, p=.319, r=.21; CE-CR: Aβ+P vs. Aβ−C, p<.001, r=.65, Aβ+P vs. Aβ−P, p=.005, r=.40, Aβ−P vs. Aβ−C, p=.164, r=.25; CE-DR: Aβ+P vs. Aβ−C, p<.001, r=.70, Aβ+P vs. Aβ−P, p<.001, r=.50, Aβ−P vs. Aβ−C, p=.439, r=.20).
As a comparison, the pattern of results for TR scores was relatively similar to that obtained for CE measures (i.e., TR-R1: Aβ+P vs. Aβ−C, p<.001, r=.65, Aβ+P vs. Aβ−P, p=.014, r=.37, Aβ−P vs. Aβ−C, p=.089, r=.28; TR-CR: Aβ+P vs. Aβ−C, p<.001, r=.70, Aβ+P vs. Aβ−P, p=.008, r=.39, Aβ−P vs. Aβ−C, p=.050, r=.31; TR-DR: Aβ+P vs. Aβ−C, p<.001, r=.71, Aβ+P vs. Aβ−P, p<.001, r=.50, Aβ−P vs. Aβ−C, p=.268, r=.22).
Analysis of the Cortical Substrates of the FCSRT Scores
SBM analyses controlling only for the effects of age showed that CE measures correlated with the CT of a much wider left-lateralized set of temporal, frontal, and parietal regions than FR measures. When controlling for the effects of the amyloid burden in addition to age, there were fewer correlations between FCSRT scores and CT, suggesting that a proportion of them were partly linked to the effects of amyloid (see Figures 2 and 3 in Supplemental Material).
The correlations highlighted in the SBM analyses controlling for the effects of the mean HV in addition to age and the amyloid burden were very similar to the correlations observed in the latter analyses controlling for age and the amyloid load (Figure 1). FR1 did not correlate significantly with the CT of any region. CFR was significantly related to the CT of two clusters localized in the left inferior temporal gyrus and the left fusiform gyrus. DFR was significantly related to the CT of a wider set of brain regions, mostly localized in the left and right temporal lobes, than those seen for CFR (Figure 1). Of interest, CE measures were globally correlated with a greater number of temporal regions, including the left entorhinal gyrus, and parietal regions than FR measures. These latter relationships were lateralized in the left hemisphere.
Correlational maps showing significant relationships between FCSRT scores and cortical thickness, when controlling for the effects of age, the amyloid load, and the mean normalized hippocampal volume. From left to right, correlational maps represent a lateral, medial and ventral view of the cortical brain surface. Maps are represented on an inflated cortical surface of an average brain. Dark gray areas correspond to sulci, while light gray areas correspond to gyri. Ant.=anterior; front.=frontal; G.=gyrus; inf.=inferior; LH=left hemisphere; mid.=middle; occipit.=occipital; pariet.=parietal; post.=posterior; RH=right hemisphere; S.=sulcus; sup.=superior; temp.=temporal.
Predictive Analyses
In the first step of the regression analyses, age alone explained between 5.50% and 6.12% of the variance in CFR and DFR measures, respectively (Table 3). However, age was not a significant predictor of the FCSRT scores in any of the remaining regression analyses.
Hierarchical multiple regressions models for FCSRT scores
* >.04 (minimum effect size representing a “practically” significant effect, Ferguson, Reference Ferguson2009).
+ =inferior to the defined significance criterion of .04 when unrounded.
SE=standard error; r 2 sp =squared semi-partial correlation coefficients; HV=hippocampal volume normalized to the ICV.
The second step of the regression analyses using age and amyloid burden as predictors revealed that amyloid load uniquely accounted for between 20.73% and 27.17% of the variance of each FCSRT measure. Amyloid burden did not account for a greater proportion of variance in CE relative to FR measures.
Regression analyses including the left or right HV, in addition to age and amyloid load, as predictors of FCSRT scores revealed that the unique proportion of FCSRT score variance accounted for by amyloid load was smaller than in the previous analyses, dropping from 20.73–27.17% to 7.03–15.15%. The left HV uniquely explained between 4.04 and 7.69% of the variance in each FCSRT score, except the CE-DR (i.e., 3.96%), while the right HV did not account for at least 4% of the variance in any FCSRT measure. The AD-biomarkers (i.e., amyloid burden, left HV) did not appear to account for a greater proportion of the variance in CE measures than in FR measures.
Finally, the mean CT of SBM-identified ROIs that were significantly related to each FCSRT score (i.e., ROIs listed in Figure 1) were individually entered, along with the three previous predictors. FR1 and CFR were not significantly predicted by the mean CT of any ROI. The mean CT of the right temporal pole accounted for at least 4% of the variance in DFR, when either the left or right HV was included as one of the three additional covariates (Tables 4a–b). The left entorhinal gyrus CT uniquely accounted for DFR variance when the right HV (but not left HV) was a predictor. The mean CT of the left entorhinal and middle temporal gyri uniquely predicted variance in all CE measures, regardless of which HV was included. The mean CT of the left posterior cingulate gyrus also uniquely predicted variance in CE-DR.
Hierarchical multiple regression models, including age, the amyloid load, the left hippocampal volume and the cortical thickness of specific ROIs as predictors of the FSCRT measures
* >.04 (minimum effect size representing a “practically” significant effect, Ferguson, Reference Ferguson2009).
+ =inferior to the defined significance criterion of .04 when unrounded.
R 2 adj =R 2 adjusted values; Pred. R 2=predicted R 2 values computed using the PRESS (predicted residual sum of squares) statistic (Stevens, Reference Stevens2002); HV=hippocampal volume normalized to the ICV; SE=standard error; r 2 sp =squared semi-partial correlation coefficients; R 2 change =Predicted R 2 changes associated with the addition of each of the CTs of ROIs to each hierarchical regression model; temp.=temporal; G=gyrus; mid.=middle; post.=posterior.
Hierarchical multiple regression models, including age, the amyloid load, the right hippocampal volume and the cortical thickness of specific ROIs as predictors of the FSCRT measures
* >.04 (minimum effect size representing a “practically” significant effect, Ferguson, Reference Ferguson2009).
R 2 adj =R 2 adjusted values; Pred. R 2=Predicted R 2 values computed using the PRESS (predicted residual sum of squares) statistic (Stevens, Reference Stevens2002); HV=hippocampal volume normalized to the ICV; SE=standard error; r 2 sp =squared semi-partial correlation coefficients; R 2 change =Predicted R 2 changes associated with the addition of each of the CTs of ROIs to each hierarchical regression model; temp.=temporal; G=gyrus; mid.=middle; post.=posterior.
Amyloid burden still accounted for a unique proportion of variance in FCSRT indexes (i.e., DFR, CE-R1, CE-CR, CE-DR) in each model, except for CE-R1 and CE-CR when the CT of the left middle temporal gyrus and the left HV were introduced as predictors. The left HV uniquely accounted for variance in DFR, CE-CR, and CE-DR only when the CT of the right temporal pole, left middle temporal, or posterior cingulate gyrus were introduced as independent variables, respectively. As in the previous analyses, the right HV did not contribute significantly to the variance of any FCSRT measure.
Predicted R 2 values for each of the regressions in Tables 4a and 4b revealed that AD biomarkers accounted for between one-fifth and one-third of the variance in the three CE measures and in DFR. These predicted R 2 values are those that would be expected if the regression equations were applied to a new sample. Predicted R 2 changes associated with the addition of each of the CTs of ROIs from Tables 4a and 4b were generally consistent with the r 2 semi-partial values, suggesting that the unique contribution of CT to the prediction of CE measures and DFR would likely hold up well under cross-validation.
DISCUSSION
The main finding of this study is that performance on both FCSRT FR and CE measures appear to be related to the degree of underlying AD-related neuropathology, including amyloid burden and left HV. However, compared to FR measures, CE scores appeared to be related to the CT of a greater number of brain regions meeting our effect size threshold. These regions are known to be susceptible to atrophy in early AD. As such, performance on CE indices would be expected to be more effective than FR scores for assessing the risk of evolving to AD.
Our first set of analyses revealed that CE performance can be specifically influenced by the presence of brain amyloid, while FR indexes appear to be influenced not only by the presence of amyloid but also by the presence of memory complaints. In the current study, participants with and without memory complaints were classified according to their brain amyloid status, as abnormal amyloid deposition in the brain is a central event in the pathogenesis of AD (Blennow, Mattsson, Scholl, Hansson, & Zetterberg, Reference Blennow, Mattsson, Scholl, Hansson and Zetterberg2015; Sperling et al., Reference Sperling, Aisen, Beckett, Bennett, Craft, Fagan and Phelps2011). FR measures (CFR, DFR) discriminated between the three groups (i.e., Aβ+P<Aβ−P<Aβ−C) with only one exception for FR1 where Aβ+P ≈ Aβ− P.
However, CE measures significantly differed only between Aβ+P and the two amyloid negative groups. Therefore, CE measures may be more specific than FR to the presence of an underlying AD-related pathology. These results are in accordance with converging evidence suggesting that CE measures are superior to FR for identifying individuals who are at-risk of AD (Dierckx et al., Reference Dierckx, Engelborghs, De Raedt, Van Buggenhout, De Deyn, Verte and Ponjaert-Kristoffersen2009; Ivanoiu et al., Reference Ivanoiu, Adam, Van der Linden, Salmon, Juillerat, Mulligan and Seron2005; Tounsi et al., Reference Tounsi, Deweer, Ergis, Van der Linden, Pillon, Michon and Dubois1999). These findings are also in line with the Wagner et al. (Reference Wagner, Wolf, Reischies, Daerr, Wolfsgruber, Jessen and Wiltfang2012) study showing that FCSRT TR, which indirectly assesses CE, distinguished between patients with a normal versus abnormal Aβ/tau ratio more accurately than did FR. The pattern of results for TR performance was similar to the pattern obtained for CE measures (i.e., Aβ−C ≈ Aβ−P>Aβ+P). While not statistically significant, the difference between Aβ−C and Aβ−P for the TR-CR score reflected a medium effect size. This pattern supports our assumption that some CE measures may be more specific than TR to the presence of AD-related pathology, given that TR measures do not constitute an unambiguous measure of the cueing effect, which is particularly affected early in AD.
The finding that FR measures were influenced by the presence of memory complaints, in addition to amyloid, may be linked to the fact that the Aβ−P group may have included individuals with undiagnosed pathological conditions other than AD that affect FR performance without disrupting CE. The underlying pathology of Aβ−P from similar studies is debated and probably heterogeneous (see data from ADNI in Wisse et al., Reference Wisse, Butala, Das, Davatzikos, Dickerson and Vaishnavi2015). Only the availability of longitudinal data could shed more light on the etiology of cognitive impairment in this group. Therefore, FR measures appear to be non-specific indicators of an underlying pathology, as these measures distinguished both Aβ−P and Aβ+P from Aβ−C, while CE measures appear to be more specific early cognitive markers of an underlying AD pathology, despite the presence of ceiling effects. Future studies could further reinforce this idea by implementing Receiver Operating Characteristic Curve Analysis in larger groups identified with AD biomarkers.
Our second set of analyses revealed that CE measures correlated with the CT of a wider set of largely left-lateralized parietal and temporal regions, including the left entorhinal gyrus, than FR measures. The substantial left lateralization is not surprising given the verbal memory demands required by the FCSRT. Our findings are also consistent with previously identified relationships between episodic memory scores and the CT or gray matter volume of temporal and parietal regions in healthy older participants (Dickerson et al., Reference Dickerson, Fenstermacher, Salat, Wolk, Maguire, Desikan and Fischl2008), asymptomatic amyloid positive individuals (Dore et al., Reference Dore, Villemagne, Bourgeat, Fripp, Acosta, Chetelat and Rowe2013), prodromal AD patients (Rami et al., Reference Rami, Sole-Padulles, Fortea, Bosch, Llado, Antonell and Molinuevo2012), and AD patients (Dickerson, Feczko, et al., Reference Dickerson, Feczko, Augustinack, Pacheco, Morris, Fischl and Buckner2009; Dickerson et al., Reference Dickerson, Fenstermacher, Salat, Wolk, Maguire, Desikan and Fischl2008).
Compared to FR, we anticipated that CE scores would rely on more cortical substrates outside the hippocampus but situated in the regions most affected by early AD pathology, making them particularly well-suited for detecting incipient AD. Cortical thinning in AD is known to occur in a specific set of regions, including temporal and parietal regions, in the earliest stages of the disease (Dickerson, Bakkour, et al., Reference Dickerson, Bakkour, Salat, Feczko, Pacheco, Greve and Buckner2009; Lerch et al., Reference Lerch, Pruessner, Zijdenbos, Hampel, Teipel and Evans2005). It is notable that the overlap between these AD cortical “signature” regions and the regions that were significantly related to FCSRT measures is larger for CE than for FR measures. This overlap may explain why CE measures appeared to identify patients with an underlying AD-related pathology better than FR scores.
Our last set of regression analyses demonstrated a substantial relationship between amyloid load and FCSRT performance, with the former accounting for one-fifth to one-quarter of the test-related variance. The presence of significant relationships between amyloid load and episodic memory performance has been equivocal (Rabinovici & Jagust, Reference Rabinovici and Jagust2009), depending notably on the disease stage. Our findings support a significant relationship between amyloid load and FCSRT scores in our sample that included patients in preclinical or prodromal stages of the disease. The unique contribution of amyloid burden to FCSRT score variance decreased but was still significant when adding left or right HV as a predictor of FCSRT measures. This pattern suggests that considerable disease-related variance is shared between amyloid burden and HV, but each process also accounts for unique aspects of the disease. This idea is consistent with findings from Mormino et al. (Reference Mormino, Kluth, Madison, Rabinovici, Baker and Miller2009), who used a recursive regression procedure to show that the relationship between episodic memory performance and amyloid deposition is mediated by hippocampal atrophy. The left HV accounted for a unique and significant proportion of variance in each FCSRT score, except CE-DR, while the right HV did not significantly predict any FCSRT score. This finding is consistent with prior studies investigating relationships between verbal episodic memory measures and left HV (e.g., Bonner-Jackson, Mahmoud, Miller, & Banks, Reference Bonner-Jackson, Mahmoud, Miller and Banks2015; Nho et al., Reference Nho, Risacher, Crane, DeCarli, Glymour and Habeck2012).
A possible explanation of the apparent independent contribution of amyloid load and the left HV to FCSRT score variance is that amyloid dysmetabolism may express itself by impairing synaptic function via overproduction of amyloid oligomers (Jongbloed et al., Reference Jongbloed, Bruggink, Kester, Visser, Scheltens, Blankenstein and Veerhuis2015). Moreover, the effect of the left HV on memory independently of amyloid load could be the result of the presence of the tau-pathology. Hippocampal tau-pathology has been reported to be the principal contributor to impaired memory performance in early AD (Markesbery, Reference Markesbery2010). In addition, tau protein pathology without the presence of amyloid deposits in the mesio-temporal lobe was also described and recently labeled as PART (Primary age-related tauopathy) by Crary et al. (Reference Crary, Trojanowski, Schneider, Abisambra, Abner, Alafuzoff and Nelson2014). Whether this condition is an age-related phenomenon, a specific form of taupathy or a “pre-fibrillar amyloid” part of the AD spectrum is not known (for a discussion, see Jack, Reference Jack2014).
The final regression analyses including selected CT ROIs as additional independent variables revealed important differences between CE and FR in terms of neuropathological predictors of each index. Relatively few CT ROIs accounted for unique variance in FR. No ROI predicted FR1 or CFR, and only the mean CT of the right temporal pole or the left entorhinal gyrus accounted for unique variance in DFR beyond that accounted for by amyloid burden and HV. In contrast, the mean CT of the left entorhinal and middle temporal gyri accounted for significant unique variance in each CE measure, and the mean CT of the left posterior cingulate gyrus uniquely predicted CE-DR performance beyond that of amyloid burden and HV. Again, it is notable that all the ROIs uniquely predicting CE performance are part of the cortical AD signature (Dickerson, Bakkour, et al., Reference Dickerson, Bakkour, Salat, Feczko, Pacheco, Greve and Buckner2009; Lerch et al., Reference Lerch, Pruessner, Zijdenbos, Hampel, Teipel and Evans2005).
As predicted, our findings suggest that the CT of regions affected most in early AD account for unique variance in more CE than FR measures. In contrast, both amyloid burden and left HV accounted for relatively similar unique proportions of variance in CE and FR. Therefore, these results suggest that the apparent superiority of CE scores in discriminating patients with an underlying AD-related pathology is not linked to the fact that CE measures are more sensitive to the presence of these AD biomarkers than FR measures but rather to the fact that these measures tend to rely on more cortical substrates that are the most susceptible to atrophy early in AD.
Furthermore, these regions might also be more affected by AD pathophysiological processes other than amyloidosis. Amyloid burden predicted a similar but unique proportion of variance in CE measures as the cortical substrates identified, including the entorhinal gyrus in particular. One may for instance speculate that this parallel, independent contribution may be due to the fact that these cortical regions are affected more by tau-pathology than by amyloidosis in our sample, which included patients in preclinical or prodromal AD stages. The entorhinal gyrus in particular is known to be affected by neurofibrillary pathology and cell loss at very early stages of AD (Braak & Braak, Reference Braak and Braak1991).
It should be noted that the mean CT of the regions significantly related to FCSRT scores did not systematically predict the corresponding FCSRT scores in the hierarchical multiple regression analyses. These discrepancies may be linked to slight differences between the two statistical methods, to different patterns of shared variance depending on whether a variable is treated as a dependent or an independent variable, and particularly to the fact that SBM analyses may highlight significant clusters in specific regions in relation to FCSRT scores while the mean CT of an entire region does not necessarily predict these scores.
A strength of this study is its investigation of relationships between performance indices from a widely used clinical memory test and cortical and biological correlates of AD risk. However, an important caveat relates to the ceiling effect found for some FCSRT scores, especially for CE measures in Aβ−C. Replication of the current findings using a comparable memory test that would be more challenging for older controls while still feasible for patients would be helpful. Furthermore, we performed hierarchical multiple regressions to investigate whether AD biomarkers and the CT of selected ROIs may predict FCSRT scores. Other interesting approaches may consist of performing mediation analyses to examine more precisely the relationships between these or other variables.
Despite its limitations, the current study suggests that FCSRT CE measures may identify patients with an underlying AD-related pathology better than FR measures. This finding does not appear to relate to a greater sensitivity of CE scores to the presence of specific AD biomarkers than FR measures, as both amyloid load and the left HV uniquely accounted for similar proportions of variance in both sets of measures. However, the apparent superiority of CE scores in identifying patients with a positive amyloid marker may be linked to the fact that these measures tend to rely on more cortical substrates that are the most susceptible to atrophy in AD than FR measures.
ACKNOWLEDGMENTS
This research was supported by the Belgian Foundation for Scientific Research (LQ, FNRS grant number A 2/5 -MCF/DM) and Saint-Luc Foundation (BH: Bourse Ordre de Malte Oeuvre du Calvaire). We are grateful to the firm GE Healthcare Ltd. for having supplied with [18F]flutemetamol for the PET scan imaging according to a convention with our Clinic. The authors do not have any potential conflict of interest to declare.
SUPPLEMENTARY MATERIAL
To view supplementary material for this article, please visit http://dx.doi.org/10.1017/S1355617716000813
