Functional Changes in the Language Network in Response to Increased Amyloid β Deposition in Cognitively Intact Older Adults

Abstract

Word finding symptoms are frequent early in the course of Alzheimer's disease and relate principally to functional changes in left posterior temporal cortex. In cognitively intact older adults, we examined whether amyloid load affects the network for language and associative-semantic processing. Fifty-six community-recruited subjects (52–74 years), stratified for apolipoprotein E and brain-derived neurotrophic factor genotype, received a neurolinguistic assessment, ¹⁸F-flutemetamol positron emission tomography, and a functional MRI of the associative-semantic system. The primary measure of amyloid load was the cerebral-to-cerebellar gray matter standardized uptake value ratio in a composite cortical volume of interest (SUVR_comp). The primary outcome analysis consisted of a whole-brain voxelwise linear regression between SUVR_comp and fMRI response during associative-semantic versus visuoperceptual processing. Higher activity in one region, the posterior left middle temporal gyrus, correlated positively with increased amyloid load. The correlation remained significant when only the word conditions were contrasted but not for pictures. According to a stepwise linear regression analysis, offline naming reaction times correlated positively with SUVR_comp. A binary classification into amyloid-positive and amyloid-negative cases confirmed our findings. The left posterior temporal activity increase may reflect higher demands for semantic control in the presence of a higher amyloid burden.

Introduction

Modern techniques such as amyloid positron emission tomography (PET) allow one to detect hallmark lesions related to Alzheimer's disease (AD) directly in vivo (Clark et al. 2011; Herholz and Ebmeier 2011; Clark et al. 2012; Vandenberghe, Adamczuk, Dupont et al. 2013; Vandenberghe, Adamczuk, Van Laere et al. 2013). Depending on mainly age and apolipoprotein E (APOE) genotype, 10–30% of cognitively intact older adults have a positive amyloid scan, which can be indistinguishable from what is seen in clinically probable AD (for review, see Chételat et al. 2013). Longitudinally, increased Aβ load is associated with greater risk of cognitive decline (Morris et al. 2009; Resnick et al. 2010; Villemagne et al. 2011; Doraiswamy et al. 2012) and gray matter volume loss (Chételat et al. 2012). Amyloid PET has become one of the principal ways to define the “preclinical” stage of AD, a term that refers to the AD-related pathogenetic processes that happen before clinical symptoms become apparent (Sperling et al. 2011). In this study, we used ¹⁸F-flutemetamol (Vandenberghe et al. 2010) as our ligand. Previous studies have revealed a high correlation between the cortical retention levels obtained with this ligand and those obtained with ¹¹C-Pittsburgh Compound B (Vandenberghe et al. 2010; Hatashita et al. 2014) as well as with neuritic plaque density based on Bielschowsky silver staining (Thurfjell et al. 2014).

Word finding difficulties are frequent in clinically probable AD, even at a pre-dementia stage (Bayles and Tomoeda 1983; Huff et al. 1986; Chertkow and Bub 1990; Vandenbulcke et al. 2007; Apostolova et al. 2008; Clark et al. 2009; Sugarman et al. 2012). The first language area to become dysfunctional in early-stage AD and amnestic mild cognitive impairment (MCI) is the left posterior superior temporal sulcus (STS) (Nelissen et al. 2007; Vandenbulcke et al. 2007). Functional magnetic resonance imaging (fMRI) activity is lower in this region during associative-semantic compared with visuoperceptual processing in MCI patients (Vandenbulcke et al. 2007) and in clinically probable AD (Nelissen et al. 2007) compared with controls. In these populations, fMRI activity levels positively correlate with Boston Naming test scores (Nelissen et al. 2007) and with word identification speed (Vandenbulcke et al. 2007). Furthermore, in AD patients in whom naming is preserved, fMRI activity in the homotopical right-sided STS is increased compared with controls (Nelissen et al. 2007). Accordingly, we hypothesized that posterior temporal cortex may show adaptive changes in the presence of increased amyloid burden also in cognitively intact individuals. The study of functional changes related to amyloid burden in cognitively intact subjects is important because it could explain why some brains appear to be more resilient against Aβ-related injury than others. This factor may determine which individuals show clinical manifestations of underlying Alzheimer pathology and who remain cognitively intact despite the presence of Alzheimer pathology in the brain (Crystal et al. 1988; Katzman et al. 1988; Troncoso et al. 1996; Davis et al. 1999; Price and Morris 1999; Driscoll et al. 2006; Aizenstein et al. 2008). Even during the initial stages of neurodegenerative disease, the brain retains a potential for plasticity (Hyman et al. 1987; Nathan et al. 1994; Becker et al. 1996; Arendt et al. 1997; Mesulam 1999; Saykin et al. 1999; Grady et al. 2001; Grossman, Koenig, DeVita et al. 2003).

One of the genes that have been implicated in functional plasticity (Gorski et al. 2003; Webster et al. 2006) is brain-derived neurotrophic factor (BDNF), both in humans (Erickson et al. 2011) and in animal models (Okuno et al. 1999; Li et al. 2008; Osada et al. 2008). The presence of 1 or 2 met alleles on codon 66 is often considered to reduce the capacity for functional reorganisation. As our second hypothesis, we examined whether adaptive changes occurring in the language network in response to amyloid load differ between BDNF met carriers and non-carriers and how this interacts with APOE ϵ4 genotype (Adamczuk et al. 2013).

Materials and Methods

The protocol was approved by the Ethics Committee University Hospitals Leuven (EudraCT: 2009-014475-45), and written informed consent was obtained from all subjects in accordance with the latest version of the Declaration of Helsinki.

Participants

Subjects were recruited from the community via advertisement in local newspapers and via a website for seniors, asking for healthy volunteers between 50 and 75 years of age for participation in a scientific study at the University Hospital Leuven, Belgium, involving brain imaging. The relationship between genotype (APOE vs. BDNF) and amyloid levels in the present cohort has been described by Adamczuk et al. (2013).

At screening, subjects underwent blood sampling for genotyping, Mini Mental State Examination (MMSE), Clinical Dementia Rating (CDR) score, and a structured interview about medical history. Inclusion was stratified per age bin (50–59, 60–64, 65–69, and 70–75) for 2 genetic factors: BDNF (met allele present or absent) and APOE (ϵ4 allele present or absent). The cells of this 2 × 2 factorial design were prospectively matched for number of cases, gender, age, education, and handedness (Edinburgh Handedness Inventory) (Adamczuk et al. 2013). BDNF and APOE variants were genotyped by sequencing at the Genetic Service Facility (GSF, www.vibgeneticservicefacility.be) of the VIB Department of Molecular Genetics. The study exclusion criteria were an MMSE score below 27, a CDR score above 0, neurological or psychiatric history, brain lesions on structural MRI, left-handedness, non-native Dutch speaker, and below-normal test scores on conventional neuropsychological assessment (< 1.9 SD on published norms adapted for age, gender, and education) (Table 1). The conventional neuropsychological test protocol consisted of the Rey Auditory Verbal Learning Test, Boston Naming Test, Letter Verbal Fluency and Animal Verbal Fluency, Raven's Standard Progressive Matrices, and the Trail Making Test (Table 1).

Fifty-six healthy, right-handed adults between 50 and 75 years of age (mean age = 65, SD = 5.5, range 52–74) who fulfilled all criteria were included in the study.

Experimental Language Tests

Given our a priori hypothesis of early involvement of left posterior STS and given its possible role in lexical-semantic retrieval (Vandenbulcke et al. 2007), the experimental language tests conducted outside the fMRI scanner consisted of confrontation naming, lexical decision, and speeded word identification (Table 2). Each of these tests was presented by Presentation 14.8 (NeuroBehavioural Systems) and was displayed on a 19-inch cathode ray tube monitor (resolution 1024 × 768 pixels, refresh rate 75 Hz) 60 cm from subjects' eyes.

Confrontation Naming Task

In a computerized version of the picture naming task from Laiacona and Capitani (2001), 60 white line drawings of concrete entities were presented on a black background (picture size 9.68 deg × 7.74 deg; Snodgrass and Vanderwart 1980). The 60 items comprised 3 living (10 animals, 10 fruits, and 10 vegetables) and 3 non-living (10 tools, 10 pieces of furniture, and 10 vehicles) categories. Item order was randomized for each individual. A trial started with the appearance of a fixation point displayed for 2 s before stimulus onset. A warning sound (177 ms duration) was presented 500 ms before stimulus onset. The stimulus was on the screen until the subject provided a response, for a maximum duration of 1 min.

Reaction times (RT) were measured for the correct responses from the onset of the stimulus to the onset of the naming response. Voice recordings were manually analyzed in the WavePad Sound Editor version 4.57 (http://www.nch.com.au/wavepad). Accuracy was measured as percentage correct responses. Responses were considered correct if they were the picture's dominant name, a synonym, the name of a subordinate to the entity designated by the dominant name, or else if it occurred in at least 3 out of 30 other healthy controls viewing the picture for 2 s. Spontaneous, immediate auto-corrections were allowed.

Lexical Decision Task

In a computerized version of the visual lexical decision test from the Dutch version of the Psycholinguistic Assessment of Language Processing in Aphasia (PALPA 24, Bastiaanse et al. 1995), words and non-words were presented as white letters (letter height: 1 deg) on a black background. Stimuli consisted of 40 words with high imageability (20 high- and 20 low-frequency words), 40 words with low imageability (20 high- and 20 low- frequency words) and 80 non-words, randomly divided into 4 blocks (literal transcription of PALPA 24 paper version). Each stimulus was preceded by a fixation point for 1 s. Subjects were instructed to use their dominant hand and respond by key press whether the stimulus was a word or non-word. The stimulus was on the screen until the subject responded, for a maximum duration of stimulus presentation of 30 s.

RTs were measured for correct responses from the onset of the stimulus to the time of the button press. A’ was used as our accuracy measure (Pallier 2002).

Speeded Word and Picture Identification

The purpose of the speeded word and picture identification task was to analyze written word and picture identification under varying time constraints (Vandenbulcke et al. 2007). We derived a time-accuracy curve for stimulus presentation durations varying between 30, 60, 90, 150, 200, 500, 800, and 2000 ms. Subjects were instructed to read the word or name the picture. A trial consisted of a warning sound, a forward mask (200 ms duration, 9.68 × 7.74 deg), followed by either a word (letter height 1 deg) or a picture (9.68 × 7.74 deg; Snodgrass and Vanderwart 1980), which was immediately followed by a backward mask (200 ms duration, 9.68 × 7.74 deg), and a fixation point for 3 s. For each individual subject, the onset (a), steepness of the curve (b), and asymptote (c) of the time-accuracy function for words and pictures were calculated by means of the equation: $accuracy=c∗(1−e(a−Δt)/b)$ for Δt ≥ a (Verhaeghen et al. 1998; Vandenbulcke et al. 2007). Goodness of fit was estimated as the sum of squared differences between the measured and calculated values (sum of the squared errors).

Functional MRI

Stimuli and Tasks

Stimuli were projected onto a screen (resolution of 1024 × 768 pixels, refresh rate 60 Hz) using Presentation 14.8 (NeuroBehavioural Systems). The fMRI paradigm has been described in detail before (Vandenberghe et al. 1996; Vandenbulcke et al. 2005, 2006; Nelissen et al. 2007; Vandenbulcke et al. 2007; Nelissen et al. 2011; Vandenberghe, Wang et al. 2013). In summary, the experimental design was factorial (Vandenberghe et al. 1996). The first factor, task, had 2 levels: associative-semantic (Fig. 1, blue and purple) versus visuoperceptual judgment (Fig. 1, cyan and yellow). The second factor, input modality, also had 2 levels: printed words (Fig. 1, blue and cyan) versus pictures (Fig. 1, purple and yellow). The associative-semantic condition was derived from the Pyramids and Palm Trees test (Howard and Patterson 1992), a classical neuropsychological test of associative-semantic processing for words and pictures. During a trial, a triplet of stimuli was presented for 5250 ms, 1 stimulus on top (the sample stimulus) and 1 in each lower quadrant (the test stimuli) at 4.6 deg eccentricity (mean picture size was 3.7 deg and mean letter size 1.2 deg), followed by a 1500 ms of interstimulus interval. Subjects were asked to press a left- or right-hand key depending on which of the 2 test stimuli matched the sample stimulus more closely in meaning. A given triplet was presented in either the picture or the word format, and this was counterbalanced across subjects. In the visuoperceptual control condition, a picture or word stimulus was presented in 3 different sizes (mean picture size was 3.7 deg and mean letter size 1.2 deg). Subjects had to press a left- or right-hand key depending on which of the 2 test stimuli matched the sample stimulus more closely in size on the screen. An epoch, that is, a block of trials belonging to the same condition, consisted of 4 trials (total duration 27 s). The fifth condition consisted of a resting baseline condition during which a fixation point was presented in the center of the screen (Fig. 1, red). During each fMRI run (5 runs in total), a series of the 5 epoch types was replicated 3 times (Fig. 1, timeline). The order of conditions was pseudorandom and different across runs of the same subject.

Prior to the fMRI session, visual acuity was tested in each participant. Subjects were asked to read aloud a text written in font 12 at 40 cm distance from their eyes. In case a correction to normal vision was necessary, subjects received MR compatible glasses with lenses matched to the subjects’ sight defect. Following this, subjects performed an offline practice session of fMRI task. In this session, we determined which size difference (9%, 6%, 3%, or 1%) for the visuoperceptual conditions was needed for each individual subject to obtain comparable accuracies as for the associative-semantic conditions.

Image Acquisition

Twenty-eight subjects were scanned on a 3T Philips Intera system equipped with an 8-channel receive-only head coil (Philips SENSitivity Encoding head coil). Twenty-eight subjects could not undergo the fMRI in the Intera system because their body in the scanner lumen obstructed the beam from the projector to the screen. These subjects were scanned on a 3T Philips Achieva system equipped with a 32-channel receive-only head coil (Philips SENSitivity Encoding head coil), which used a screen placed behind the individual's head for the projection. There were no statistically significant differences of sex (P = 0.79) (chi-square test), genetic groups (P = 0.17), age (P = 0.49), or MMSE (P = 0.92) (Kolmogorov–Smirnov comparison of 2 datasets) between subjects scanned on the Intera versus the Achieva system. Scanner type was included as a covariate of no interest for all analyses.

Sequence parameters were the same for both scanners. A high-resolution T1-weighted structural scan was obtained using a 3D turbo field echo sequence (coronal inversion recovery prepared 3D gradient-echo images, inversion time 900 ms, TR = 9.6 ms, TE = 4.6 ms, flip angle 8°, field of view = 250 × 250 mm, 182 slices; voxel size 0.98 × 0.98 × 1.2 mm³). Functional MRIs were acquired using T2* echo-planar images (50 transverse slices, voxel size 2.5 × 2.5 × 2.5 mm³; TR = 3000 ms, TE = 30 ms, flip angle 90°, field of view 200 × 200 mm).

Image Analysis

All analyses were performed using Statistical Parametric Mapping 8 (SPM8, http://www.fil.ion.ucl.ac.uk/spm/). Functional MR scans of each subject were realigned to correct for potential head motion. The structural MR image was coregistered to the average of the realigned fMRI images. The structural MR image was then normalized to the SPM8 T1 template in Montreal Neurological Institute space. The same normalization matrix was applied to the coregistered fMRI scans. The normalized fMRI images (voxel size 3 × 3 × 3 mm³) were smoothed using a 6 × 6 × 6 mm³ Gaussian kernel. A high-pass filter with a Full Width at Half Maximum of 270 s and a low-pass filter consisting of a canonical hemodynamic response function (HRF) were applied. The epoch-related response was modeled by a canonical HRF convolved with a boxcar.

Flutemetamol PET

Image Acquisition

As described before (Koole et al. 2009; Nelissen et al. 2009; Vandenberghe et al. 2010; Vandenberghe, Nelissen et al. 2013), images were acquired on a 16-slice Siemens Biograph PET/CT scanner (Siemens). The PET tracer was injected intravenously as a bolus (mean activity 151.2 MBq, SD 8.3, range 137.9–192.5 MBq) in an antecubital vein. Image acquisition started 90 min after tracer injection and lasted for 30 min. Prior to the PET scan, a low-dose computed tomography scan was performed for attenuation correction. Random and scatter corrections were also applied. Images were reconstructed using Ordered Subsets Expectation Maximization (4 iterations × 16 subsets).

Image Analysis

The PET data were reconstructed as 6 frames of 5 min and realigned to the first frame to correct for potential head motion. Subsequently, the 6 frames were summed to create 1 summed image. The individual's T1-weighted structural image was then coregistered to this PET summed image. This MR image was subsequently normalized to the SPM8 T1 template. The same normalization matrix was then applied to the individual's coregistered PET summed image. From the spatially normalized PET images (voxel size 2 × 2 × 2 mm³), standardized uptake value ratios (SUVR) were calculated in a voxelwise manner with cerebellar gray matter (GM) as reference region. The cerebellar GM reference region was defined as areas 91–108 of the Automated Anatomical Labelling atlas (AAL) (Tzourio-Mazoyer et al. 2002). The cerebellar reference region was resliced to each individual's normalized PET summed image. In order to exclude most of the white matter (WM) content, it was masked by the normalized and modulated subject-specific GM map, with the threshold for masking set at > 0.3.

Our primary PET outcome measure was the mean SUVR in a composite cortical volume of interest (VOI) (SUVR_comp). This composite VOI consisted of 5 bilateral cortical regions: frontal (AAL areas 3–10, 13–16, and 23–28), parietal (AAL 57–70), anterior cingulate (AAL 31–32), posterior cingulate (AAL 35–36), and lateral temporal (AAL 81–82 and 85–90). The composite cortical VOI was resliced to each individual's normalized PET summed image. In order to exclude most of the WM content, it was masked by the normalized and modulated subject-specific GM map, with the threshold for masking set at > 0.3.

While we used SUVR_comp as a continuous variable in our primary analysis, we also conducted a secondary analysis where amyloid load was treated as a binary variable and cases were classified as amyloid-positive versus amyloid-negative based on an SUVR_comp cutoff. Such a binary approach is closer to the way in which Sperling et al. (2011) conceptualized preclinical AD. The SUVR_comp cutoff for binary classification was derived from an independent dataset (Vandenberghe et al. 2010), which contained 27 scans from AD patients (mean age 70, SD 7.0) and 15 scans from healthy older controls (HC) (mean age 69, SD 7.6). ¹⁸F-flutemetamol scans from the Vandenberghe et al. (2010) study were re-analyzed using the MRI-informed PET analysis method described earlier. The cutoff was defined based on the statistical distance between the AD group and the HC as described in Vandenberghe et al. (2010). This gave an SUVR_comp cutoff equal to 1.38. Note that this cutoff is lower than the cutoff defined by Vandenberghe et al. (2010) or Thurfjell et al. (2014) for a purely PET-based approach, probably due to exclusion of more WM signal in the MRI-informed method in the amyloid-negative cases. Because of this difference, we also verified our binary case classification using the PET-based method and cutoff from Thurfjell et al. (2014) (cutoff equal to 1.57).

We verified our findings using partial volume corrected (PVC) data. PVC was based on the MRI using the modified Müller-Gärtner method (Müller-Gärtner et al. 1992; Adamczuk et al. 2013). This method makes use of probabilistic segmentation and determines tracer concentration per unit volume of GM. The normalized unmodulated GM and WM segmentations were used to estimate different tissue fractions per voxel. PVC was applied to the normalized PET summed images. The remaining procedures were identical to those outlined earlier.

Statistical Analysis

Analysis of Behavioral Data Obtained During fMRI

RTs and accuracies (% correct responses) were analyzed by means of a four-factor repeated-measures ANOVA, with stimulus modality (2 levels: pictures vs. words) and task (2 levels: associative-semantic vs. visuoperceptual) as within-subject factors and, as between-subject factors, BDNF (2 levels: codon 66 met carriers vs. non-carriers) and APOE (2 levels: ϵ4 carriers vs. non-carriers) genotype (Table 3). Pairwise comparisons were performed using Bonferroni post hoc tests.

Whole-Brain Voxelwise Analysis

All voxelwise analyses were performed using SPM8. For each subject, parameter estimates were generated modeling each of the 5 conditions. We then created the following contrast images, averaging across runs: The first-level contrast images were then used for second-level whole-brain analysis.

1. (Associative-semantic task with words + associative-semantic task with pictures) − (visuoperceptual task with words + visuoperceptual task with pictures)

2. Associative-semantic task with words − visuoperceptual task with words

3. Associative-semantic task with pictures − visuoperceptual task with pictures

4. (Associative-semantic task with words − visuoperceptual task with words) − (associative-semantic task with pictures − visuoperceptual task with pictures) and inversely

5. (Associative-semantic task with words + associative-semantic task with pictures) − baseline

6. (Visuoperceptual task with words + visuoperceptual task with pictures) − baseline.

7. Visuoperceptual task with words − baseline

8. Visuoperceptual task with pictures − baseline

Our primary outcome analysis consisted of a whole-brain voxelwise linear regression analysis with SUVR_comp as independent variable, and fMRI response in Contrast 1 (main effect of task) as dependent variable. The statistical map was thresholded at a significance threshold of voxel-level P_uncorrected of <0.001 combined with a cluster-level P_corrected < 0.05, family-wise error (FWE) corrected for the whole-brain volume.

As a secondary outcome analysis, we examined for each of the 10 regions that constitute the composite cortical VOI, the correlation between regional SUVR and the main effect of task (Contrast 1) across the whole brain.

Furthermore, we examined whether any significant correlations with amyloid load were found for contrasts 2–8.

As a further secondary outcome analysis, we performed a whole-brain voxel-by-voxel linear regression between SUVR images and fMRI images representing associative-semantic minus visuoperceptual activity (Contrast 1) using Biological Parametric Mapping (BPM) (Casanova et al. 2007). The BPM is a toolbox for multimodal image analysis, which is based on a voxelwise use of the SPM's general linear model. This allows comparison of different imaging modalities within voxels.

As a further secondary analysis, we categorized the cases into amyloid-positive versus amyloid-negative and compared the main effect of task between the 2 groups (Contrast 1) using a two-sample t test.

We also tested whether there was any difference in fMRI response between the 4 genetic groups by means of a factorial ANOVA with BDNF (2 levels: met allele present vs. absent) and APOE (2 levels: ϵ4 allele present vs. absent) as between-subject factors and the main effect of task (Contrast 1) as dependent variable.

All whole-brain voxelwise analyses were thresholded at a significance threshold of voxel-level P_uncorrected < 0.001 combined with a cluster-level P_corrected < 0.05, FEW corrected for the whole brain volume. In the BPM analysis, we used threshold of voxel-level P_uncorrected < 0.001 combined with cluster size of at least 10 voxels.

Relationship to Offline Measures of Linguistic Performance

When our primary analysis revealed clusters of significant correlation between SUVR_comp and fMRI response during associative-semantic versus visuoperceptual processing (Contrast 1), we examined in further detail whether mean fMRI response in these clusters correlated with a pre-specified set of offline measures of linguistic performance. Clusters of voxels exhibiting a significant correlation between SUVR_comp and fMRI response (Contrast 1) were extracted using the MarsBaR 0.43 toolbox (http://marsbar.sourceforge.net/). Given the proposed role of the posterior STS in lexical-semantic retrieval (Vandenbulcke et al. 2007), the principal measures that we selected a priori were 1) RT during the confrontation naming task, 2) RT during the lexical decision task, 3) The b parameter from the speeded word identification task. For each of these parameters, we performed a stepwise linear regression analysis with this parameter as dependent variable and the independent variables: fMRI response during the associative-semantic minus the visuoperceptual condition (Contrast 1), SUVR_comp, age, education level, BDNF genotype and APOE genotype. Probability to enter the model was set at P < 0.05 with probability to remove set to P > 0.1. This analysis was performed outside SPM in a VOI-based manner using STATISTICA 11 (http://www.statsoft.com/) as SPM software does not include stepwise linear regression.

In an additional, binary approach, we examined which of these neurolinguistic measures differed between the amyloid-positive and the amyloid-negative class (two-sample t test).

Results

Analysis of Behavioral Scores During fMRI

The main effect of task was significant: Subjects responded more accurately during the associative-semantic conditions than during the visuoperceptual conditions (F_1,52 = 17.4, P = 0.0001), albeit with longer RTs (F_1,52 = 24.9, P = 0.000007) (Table 3). The main effect of modality was also significant: Subjects responded more slowly (F_1,52 = 21.3, P = 0.00003) and less accurately (F_1,52 = 5.7, P = 0.02) for pictures than for words (Table 3). The interaction between task and modality was significant (F_1,52 = 11, P = 0.002): The associative-semantic task was performed more accurately with words than with pictures (P = 0.00008), whereas there was no difference between words and pictures for the visuoperceptual task (P = 1). There was no main effect of BDNF and APOE genotypes on accuracy or RT (P > 0.2) (Table 3). The three-way interaction between task, modality and APOE genotype was significant for accuracies (F_1,52 = 8.2, P = 0.006) (Table 3). According to a post hoc analysis, APOE ϵ4 non-carriers performed the associative-semantic task more accurately with words than with pictures (P = 0.0006), whereas there was no difference between the 2 input-modalities for the visuoperceptual task (P = 0.35). No such difference was seen in the APOE ϵ4 carriers (P = 0.23). We did not find any significant interactions for RT (P > 0.05).

Whole-Brain Voxelwise Analysis

Univariate Contrast Between fMRI Conditions

The contrast between the associative-semantic minus the visuoperceptual conditions (main effect of task, Contrast 1) revealed a distributed semantic network consistent with previous findings (Vandenberghe et al. 1996; Vandenbulcke et al. 2007; Nelissen et al. 2007) (Fig. 2A). The interaction between task and input modality (Contrast 4) revealed word-specific activation during the semantic compared with the visuoperceptual task in left STS, extending from posterior to more anterior portions of the STS (cluster peak coordinates −66, −36, 9, extent (ext) = 113 voxels, cluster-level P_corrected = 0.0001) (Fig. 2B). Picture-specific semantic activation (inverse of Contrast 4) occurred bilaterally in ventral occipitotemporal cortex extending to superior occipital gyrus (right cluster peak coordinates 33, −45, −21, ext = 1756 voxels, cluster-level P_corrected < 0.0001 and left cluster peak coordinates −39, −51, −15, ext = 1286 voxels, cluster-level P_corrected < 0.0001) and in right inferior frontal gyrus (cluster peak coordinates 48, 12, 27, ext = 141 voxels, cluster-level P_corrected < 0.0001) (Fig. 2C).

Linear Regression Between fMRI Response and Amyloid Load

In a whole-brain analysis, the posterior third of the left middle temporal gyrus (MTG) (Fig. 3A) exhibited a significant positive correlation between fMRI response during the associative-semantic versus the visuoperceptual task (Contrast 1) and SUVR_comp: Activity levels were higher with a higher amyloid load (−57, −45, 9, ext = 64 voxels, cluster-level P_corrected = 0.006) (Fig. 3A–E). No other regions showed a correlation, even when we lowered the significance threshold to cluster-level P_corrected < 0.25. The MTG belonged to the amodal network, as evidenced by the conjunction analysis of Contrasts 2 and 3 (Fig. 4A–C).

When we restricted the contrast between the associative-semantic and the visuoperceptual task to the words (Contrast 2) and examined the correlation with amyloid load in the whole-brain analysis, SUVR_comp correlated positively with fMRI response during the associative-semantic minus visuoperceptual control condition for words in the same region (−60, −48, 9, ext = 49 voxels, cluster-level P_corrected = 0.02). For pictures, there was no correlation (Contrast 3) (cluster-level P_corrected > 0.6). No other regions showed a correlation, even when we lowered the significance threshold to cluster-level P_corrected < 0.10. Neither were any correlations with SUVR_comp found for Contrasts 4–8.

Analysis of PVC data confirmed these results. PVC SUVR_comp positively correlated with fMRI response during the associative-semantic minus the visuoperceptual condition (Contrast 1) in the posterior third of the left MTG (−54, −39, 12, ext = 87 voxels, cluster-level P_corrected = 0.001). Analysis restricted only to the word conditions showed that SUVR_comp correlated positively with fMRI response during the associative-semantic minus visuoperceptual control condition for words (Contrast 2) in the same region (−54, −36, 9, ext = 45 voxels, cluster-level P_corrected = 0.028). No correlation was found between SUVR_comp and the fMRI response during the associative-semantic minus visuoperceptual condition when only pictures were used (Contrast 3) (cluster-level P_corrected > 0.07).

Among the regions which constituted the composite cortical VOI, average SUVR in each of the regions besides the left and right anterior cingulate and left lateral frontal region contributed to the correlation of fMRI response during Contrast 1 and SUVR_comp (cluster-level P_corrected< 0.038).

BPM indicated a significant correlation within-voxels between fMRI activity and SUVR in the posterior left MTG (cluster peak coordinates −60, −54, 12, ext = 18 voxels, voxel-level P_uncorrected = 0.0001, Z = 3.65, r = 0.50).

Binary Classification

We evaluated whether similar results would be obtained had we used a binary approach: amyloid-positive versus amyloid-negative group (Fig. 5). Eight subjects (14%) were classified as positive (Fig. 5). fMRI performance parameters did not differ between the amyloid-positive and the amyloid-negative group (Table 4). In a whole-brain voxelwise analysis, the amyloid-positive group exhibited a higher fMRI response compared with the amyloid-negative group during the associative-semantic minus visuoperceptual conditions (Contrast 1) in the posterior third of the MTG (−54, −42, 9, ext = 55 voxels, cluster-level P_corrected = 0.013) (Fig. 5B, red cluster). This was also true for the contrast between the associative-semantic minus visuoperceptual task presented as words (Contrast 2) (−57, −45, 6, ext = 59 voxels, cluster-level P_corrected = 0.008) (Fig. 5B, green cluster). Overlap between clusters is shown in dark orange (Fig. 5B). We did not find any significant differences elsewhere and neither did we find any significant between-group differences for other contrasts. When we applied the Thurfjell et al. (2014) method and cutoff for binary classification, 4 cases were positive. The between-group differences remained essentially the same: The amyloid-positive group had higher fMRI response compared with the amyloid-negative group during the associative-semantic versus visuoperceptual condition (−60, −48, 9, ext = 54 voxels, cluster-level P_corrected = 0.014).

Genotype Effect on fMRI Response and Relationship Between fMRI Response and SUVR

APOE and BDNF genotype did not affect the activity patterns during the associative-semantic versus visuoperceptual conditions (cluster-level P_corrected > 0.8). Nor was there an effect of BDNF or APOE genotype on the correlation between mean fMRI response in the left posterior MTG (Contrast 1) and SUVR_comp (P > 0.1) (BDNF met carriers r = 0.64, P = 0.0002; BDNF non-carriers r = 0.61, P = 0.0005; APOE ϵ4 carriers r = 0.71, P = 0.00006; APOE ϵ4 non-carriers r = 0.40, P = 0.03) (Fig. 6).

Relationship with Offline Language Measures

According to a stepwise regression analysis, variance in RT during the confrontation naming task was partly explained by SUVR_comp (r = 0.27, P = 0.04), rather than fMRI response (Contrast 1; P = 0.62), age (P = 0.34), education (P = 0.57), BDNF (P = 0.21), or APOE status (P = 0.42) (Fig. 7).

Response latencies during confrontation naming were longer in the amyloid-positive compared with the amyloid-negative group (P = 0.047) (Table 4). There were no differences for any of the other experimental language tests and neither did the conventional neuropsychological test scores differ between the amyloid-positive and amyloid-negative class (Table 4).

Discussion

In cognitively intact older adults, a higher amyloid burden was associated with subclinical alterations of the network for language and associative-semantic processing. Activity in left posterior MTG was higher with a higher amyloid load (Fig. 3A,B). Higher amyloid levels were correlated with slower confrontation naming (Fig. 7). Our posterior temporal findings are based on a whole-brain search without prior restriction of the search volume. They were in agreement with our a priori hypothesis about posterior temporal cortex, although the exact location was in the amodal posterior MTG (Fig. 4A) adjacent to the word-specific posterior STS found before in MCI (Vandenbulcke et al. 2007) and AD (Nelissen et al. 2007) patients.

While we used the SUVR_comp as a continuous variable for the primary outcome analysis, we also conducted a secondary analysis where amyloid load was treated as a binary variable and cases were classified as amyloid-positive versus amyloid-negative based on an SUVR_comp cutoff. Such a binary approach is closer to the way in which Sperling et al. (2011) conceptualize preclinical AD. The findings based on a binary approach were entirely in line with the findings obtained with the linear regression approach (Fig. 5).

During the visuoperceptual control conditions, subjects engaged in an active comparison of the size-on-the-screen of a picture or a word. It is highly plausible that the meaning of this word or picture is automatically activated to some degree during the visuoperceptual condition too. A lower-level control condition with consonant letter strings or scrambled pictures (Ricci et al. 1999) would have been necessary had we wanted to isolate the regions activated during word and picture processing in the visuoperceptual condition. In any case, we did not find a correlation between amyloid load and the activity pattern during the visuoperceptual control conditions minus fixation baseline. This could suggest that the network underlying activation of word and picture meaning during the visuoperceptual condition is relatively intact in preclinical AD and that the principal changes are at the level of explicit associative-semantic processing.

Our findings are based on observational cross-sectional data, analyzed by means of correlational analysis and between-group comparisons. One therefore has to be careful in drawing conclusions about a causal link between the increase in amyloid burden, the increase in left posterior temporal activity, and the decrease in confrontation naming latencies. To adequately resolve this fundamental limitation, one would need to conduct interventional studies. This is difficult since there are no proven amyloid-lowering interventions available. Transcranial magnetic stimulation targeting the left posterior temporal cortex could be an option to examine how altering activity within a subject affects naming latencies and whether this depends on amyloid load.

We also observed right-hemispheric inferior frontal activation during the associative-semantic versus the visuoperceptual task (Fig. 2A), and this was particularly pronounced for the pictures (Fig. 2C). The right inferior frontal activation could be related to the older age range of our individuals, in line with the Hemispheric Asymmetry Reduction in Older Adults model (Cabeza 2002; Cabeza et al. 2005). However, one would need fMRI data over a wider age range including also young adults to confirm this. In any case, no increase in right-hemispheric activation was found as a consequence of increasing amyloid load in our dataset, contrary to our original hypothesis (Nelissen et al. 2007).

Previous studies have investigated changes in task-related fMRI in AD principally within the episodic memory domain. AD patients consistently show lower hippocampal activation in episodic memory encoding tasks in comparison with controls and/or MCI subjects (Small et al. 1999; Machulda et al. 2003; Sperling et al. 2003; Golby et al. 2005; Celone et al. 2006; Sperling 2007). Subjects with late MCI also have decreased hippocampal activity during episodic memory encoding (Machulda et al. 2003; Johnson, Schmitz, Moritz et al. 2006) whereas subjects with early MCI compared with controls show an increase in hippocampal activity during memory encoding (Dickerson et al. 2005; Johnson, Schmitz, Trivedi et al. 2006). Young healthy presenilin 1 mutation carriers destined for early-onset AD exhibit higher activity in the hippocampal formation in comparison with the non-carrier controls (Mondadori et al. 2006; Reiman et al. 2012). Subjects with a higher risk for AD due to family history and APOE ϵ4 carrier status also have higher hippocampal activation during encoding compared with non-carrier controls (Fleisher et al. 2005). These studies led to a model where the direction of functional changes in medial temporal cortex is stage-dependent: In the preclinical AD stage and the early MCI stage, activity during memory encoding in the hippocampus is increased compared with controls, whereas in the late MCI and the clinically probable AD stage, activity is decreased compared with controls. Our findings indicate that a similar sequence may occur in the language domain in left posterior temporal cortex. The current data show increased activity during associative-semantic processing in preclinical AD according to the National Institute on Aging and Alzheimer's Association criteria (Sperling et al. 2011). In a previous study in amnestic MCI, activity in left posterior STS was decreased and correlated with the speed of written word identification (Vandenbulcke et al. 2007). In clinically probable AD, the same region also showed lower activity levels (Nelissen et al. 2007). Our study is the first to report an increase in left posterior temporal cortex in a stage that has been referred to as preclinical AD (Fig. 3), similarly to what has been described in the hippocampal formation for episodic memory (Sperling 2007). Taken together, this series of studies may suggest a similar sequence of increased activity in posterior temporal cortex followed by activity decreases as Alzheimer's disease progresses to the clinical stages.

Initial functional imaging studies of language and semantic memory in clinically probable AD have emphasized prefrontal increases which correlated positively with task performance (Becker et al. 1996; Saykin et al. 1999; Grady et al. 2003). These AD-related prefrontal increases generalized across episodic and semantic memory tasks (Grady et al. 2003) and presumably reflect general adaptive strategic processes (Becker et al. 1996; Grady et al. 2003). A series of studies revealed functional changes also in temporal cortex, most notably left inferior temporal cortex (Grossman, Koenig, Glosser et al. 2003; Grossman, Koenig, DeVita et al. 2003), left and right MTG (Grossman, Koenig, Glosser et al. 2003; Grossman, Koenig, DeVita et al. 2003; Seidenberg et al. 2009), and left posterior STS (Nelissen et al. 2007; Vandenbulcke et al. 2007). Functional differentiation exists within left temporal cortex even within nearby areas. For instance, the posterior third of the left STS is activated during semantic processing specifically for words (Vandenberghe et al. 1996; Vandenbulcke et al. 2007). It has been principally implicated in lexical-semantic (Vandenbulcke et al. 2007) or lexical-phonological retrieval (Binder et al. 2000; Price and Mechelli 2005). In contrast, an adjacent more inferior region, the posterior third of the left MTG, is activated during semantic processing for both words and pictures (Fig. 4A,B,C, Fig. 3D,E). The left posterior MTG is 1 of the most consistent hubs in the associative-semantic network (Buckner et al. 2005; Vandenberghe, Wang et al. 2013). It has been implicated in amodal semantic processing (Vandenberghe et al. 1996; Vandenbulcke et al. 2007) as well as in semantic control (Whitney et al. 2011, 2012). In the current study, the correlation principally occurred within the amodal posterior MTG region rather than the word-specific STS (Fig. 3A and Fig. 4A). Our findings can be readily integrated in current hypotheses that attribute to posterior MTG a role in cognitive control: Regions involved in semantic control may be the prime candidates for compensatory processes in response to increases in amyloid load. Increased amyloid burden in cortical areas may hamper normal neuronal functioning due to its neurotoxic effects. In order to cope with such functional changes at the neuronal level, demands for cognitive control may increase and this may account for the increase in MTG activity levels. Cognitively normal older persons may have increased Aβ levels yet intact neuropsychological performance (Driscoll et al. 2006; Aizenstein et al. 2008), possibly due to ongoing compensatory processes. As the disease advances, mechanisms responsible for maintaining constant level of increased activation may become exhausted, resulting in the first cognitive symptoms. Thus, early word finding difficulties in the course of AD might arise due to failure of semantic control processes, followed by a gradual activity decrease in other language areas, for example, areas directly involved in word processing like posterior STS.

We did not find any effect of genetic polymorphisms of APOE or BDNF on the language network in our cohort. In AD, APOE ϵ4 status has been associated more closely with episodic memory deficits than with language symptoms (Lehtovirta et al. 1996; Rasmusson et al. 1996; Mendez 2012; Mez et al. 2013). HC with a family history of Alzheimer's disease and that at least 1 APOE ϵ4 allele have increased fMRI activation during a semantic memory task (famous vs. unfamiliar names) in bilateral temporoparietal areas, posterior cingulate and precuneus, posterior middle and superior temporal regions, and left hippocampal complex (Woodard et al. 2009). It is not that we did not find any effect of APOE. As reported before, APOE genotype exerted an effect on the amyloid burden in our cohort: A higher amyloid load in APOE ϵ4 carriers was present in posterior cingulate; a region outside the network that our paradigm is activating (Adamczuk et al. 2013).

As of yet, the relationship between BDNF and language has been principally studied during development and early adulthood (Freundlieb et al. 2012; Simmons et al. 2010; Li and Bartlett 2012) and in schizophrenia (Kebir et al. 2009). Contrary to our hypothesis, the regression between fMRI response and amyloid load was not influenced by BDNF status. In the same cohort, we previously reported that BDNF exerted a direct effect on amyloid load in interaction with APOE: BDNF met carriers had increased levels of Aβ in typical regions of predilection in comparison with the BDNF met non-carriers with APOE ϵ4 (Adamczuk et al. 2013). In summary, contrary to our prediction, the effect of BDNF in our cohort is situated at the level of amyloid aggregation rather than at the level of fMRI response.

Our findings highlight the critical role of left posterior temporal cortex in AD-related processes. Changes in left posterior STS lead to lexical-semantic retrieval deficits, which may explain the word finding difficulties in clinical AD (Nelissen et al. 2007) and the subclinical slowing in word identification speed in MCI (Vandenbulcke et al. 2007). The changes we observed in this study in the posterior MTG may reflect higher demands for semantic control in those subjects who are cognitively intact despite a high amyloid burden.

To conclude, our cross-sectional data indicate that a higher amyloid load in cognitively intact individuals has functional consequences for the network mediating language and associative-semantic processing. The converging evidence obtained in cognitively intact older adults, amnestic MCI and AD may suggest a sequence of events similar to that proposed for the hippocampal formation in episodic memory. The initial compensatory role of increased neuronal activity may precede later deterioration.

Funding

This work was supported by the Foundation for Alzheimer Research SAO-FRMA (09013, 11020, 13007); Research Foundation Flanders (G.0660.09); KU Leuven (OT/08/056, OT/12/097); IWT VIND; IWT TGO BioAdapt AD; Belspo IAP (P7/11); Research Foundation Flanders senior clinical investigator grant to R.V. and K.V.L.; and Research Foundation Flanders doctoral fellowship to K.A. ¹⁸F-flutemetamol was provided by GE Healthcare free of charge for this academic investigator-driven trial.

Notes

We thank the staff of Nuclear Medicine, Radiology, and the Memory Clinic at the University Hospitals Leuven, and the staff of the Genetic Service Facility at the University of Antwerp—VIB Department of Molecular Genetics. Conflict of Interest: R.V. has been the PI of the phase 1 and the phase 2 ₁₈F-flutemetamol studies and his institution had a consultancy agreement with GEHC.

References