The effect of beta‐amyloid on face processing in young and old adults: A multivariate analysis of the BOLD signal

Abstract

The recent ability to measure in vivo beta‐amyloid (Aβ), a marker of Alzheimer's disease (AD), has led to an increased focus on the significance of Aβ deposition in clinically normal adults. Evidence suggests that healthy adults with elevated cortical Aβ show differences in neural activity associated with memory encoding—specifically encoding of face stimuli. Here, we examined if Aβ deposition in clinically normal adults was related to differences in neural activity in ventral visual cortex during face viewing. Our sample included 23 high‐Aβ older adults, 23 demographically matched low‐Aβ older adults, and 16 young adults. Participants underwent cognitive testing, Aβ positron emission tomography imaging with ¹⁸F‐Florbetapir, and functional magnetic resonance imaging to measure neural activity while participants passively viewed photographs of faces. Using barycentric discriminant analysis—a between‐groups classification technique—we found that patterns of neural activity in the left fusiform gyrus, a region highly responsive to faces, distinguished Aβ status of participants. Older adults with elevated Aβ were characterized by decreased activity in left fusiform compared to Aβ‐negative older adults. Further, we found that the degree to which older adults expressed decreased fusiform activity was related to worse performance on tasks of processing speed. Our results provide unique evidence that, in addition to previously studied memory and default regions, decreased neural activity in a region important for face perception was associated with elevated Aβ and may be an early manifestation of AD. Hum Brain Mapp 36:2514–2526, 2015. © 2015 Wiley Periodicals, Inc.

INTRODUCTION

Substantial beta‐amyloid (Aβ) deposition is one of the most prominent markers of Alzheimer's disease [AD; McKhann et al., 2011]. High levels of Aβ deposition in clinically normal adults may be an early symptom of a preclinical trajectory towards AD [Sperling et al., 2011] that precedes cognitive decline [Jack et al., 2013]. Somewhere between 20 and 30% of clinically normal adults have elevated cortical Aβ deposition, as shown by in vivo positron emission tomography [PET; Rodrigue et al., 2009; Rowe et al., 2010] and confirmed in comparative autopsy studies [Clark et al., 2011]. While the presence of Aβ alone is not indicative of AD, elevated Aβ deposition has been shown to be a powerful predictor of disease progression along the AD spectrum [Jack et al., 2010; Morris et al., 2009]. In fact, a recent working group suggested that Aβ deposition in the absence of major cognitive deficit may represent a stage of preclinical AD, noting that further research is needed [Sperling et al., 2011].

Sperling [2011] has argued that functional magnetic resonance imaging (fMRI) is likely sensitive to detecting neural dysfunction associated with preclinical AD, although there are few studies in the literature thus far. To date, fMRI studies of Aβ deposition in clinically normal adults report that elevated Aβ is associated with disrupted brain connectivity [Drzezga et al., 2011; Elman et al., 2014; Mormino et al., 2011; Sheline et al., 2010] and failure to suppress default mode networks [Hedden et al., 2009; Sperling et al., 2009; Vannini et al., 2012a]. Likewise, studies examining the neural correlates of memory encoding find that elevated Aβ corresponds with deficits in memory circuitry [Kennedy et al., 2012; Mormino et al., 2012a] and decreased neural adaptation to repeated trial information [Vannini et al., 2012b]. Recent work by Huijbers et al. [2014] found that older adults with elevated Aβ failed to modulate neural activity in entorhinal cortex compared to older adults without Aβ. However, both groups showed similar activation in hippocampus suggesting that entorhinal cortex activity may be selectively affected by Aβ before clinical symptoms appear.

Thus far, fMRI studies of memory and Aβ have largely been directed toward understanding the effect of Aβ on face encoding and face memory [Huijbers et al., 2014; Rentz et al., 2011; Sperling et al., 2009; Vannini et al., 2012b]. Face‐recognition performance is a sensitive measure for discriminating clinically normal from mild cognitive impairment (MCI) stages of AD [Papp et al., 2014]. Moreover, Aβ‐positivity in clinically normal adults has been associated with deficits in face‐name memory [Rentz et al., 2011]. Neuroimaging studies find strong evidence that Aβ deposition is related to disruptions in memory and default network regions during encoding and retrieval of face stimuli [Huijbers et al., 2014; Sperling et al., 2009; Vannini et al., 2012a, 2012b]. While there are multiple studies examining the effect of Aβ on the underlying memory networks involved in face memory, there are no studies focused on possible Aβ‐related differences in the neural activity underlying face perception.

An extensive literature has yielded evidence for a highly specialized network of regions in ventral visual cortex involved in face perception [e.g., bilateral fusiform, inferior occipital; Grill‐Spector and Malach, 2004; Kanwisher et al., 1997]. In the present study, we investigated whether patterns of neural activity associated with face perception in ventral visual cortex were affected by Aβ deposition in a sample of clinically normal adults. We build on evidence that the specialized neural signal evoked by viewing faces becomes less selective in old age [Carp et al., 2011; Lee et al., 2011; Park et al., 2004; Voss et al., 2008] and shows a broader response to more stimulus categories [Park et al., 2012]. Importantly, this age‐related difference in neural response to faces corresponds to poorer performance on cognitive measures of fluid processing [Park et al., 2010] and face‐matching [Burianová et al., 2013], suggesting that differences in neural activity underlying face perception may be associated with normal cognitive aging processes [Park and Reuter‐Lorenz, 2009].

Here, we examined the possibility that elevated Aβ deposition in clinically normal adults affects neural activity underlying face perception in ventral visual cortex. We utilized data from the Dallas Lifespan Brain Study [Kennedy et al., 2012; Rodrigue et al., 2012], which provided the unique ability to draw three highly‐selective groups of participants from a large lifespan sample of healthy adults: (1) a group with significantly elevated Aβ deposition, (2) a group with low levels of Aβ that was carefully matched to high‐Aβ participants for age, sex, and education, and (3) a group of young adults. Because the current literature suggests that the effect of Aβ on cognition is relatively subtle [Hedden et al., 2013], we utilized a data‐driven multivariate analysis [Abdi et al., 2012a; Westman et al., 2011], which enabled us to detect potentially small differences in patterns of neural activity among our three groups of participants. Specifically, we used a between‐group principal components analysis (PCA) technique called Barycentric Discriminant Analysis (BADA; Abdi et al., 2012a, 2012b] to identify neural patterns in ventral visual cortex specific to the groups of participants while they underwent a passive face‐viewing task in‐scanner.

The current study had three main hypotheses: (1) Aβ positivity in older adults would be related to differences in neural patterns in regions of the ventral visual cortex highly selective to viewing faces (e.g., fusiform) compared to young and Aβ negative older adults. (2) Congruent with past findings, we hypothesized that young and older adults would show robust differences in neural patterns associated with viewing faces. (3) Lastly, we explored whether neural activity in ventral visual regions affected by Aβ would be sensitive to cognitive performance, as such a finding would support evidence suggesting that neural activity may be sensitive to cognitive deficits in preclinical stages of AD [Sperling, 2011].

MATERIALS AND METHODS

Participants

This study included a total of 62 healthy adults, ages 30–88, in three groups: high‐Aβ adults, demographically matched low‐Aβ adults, and young adults. These participants were drawn from 142 healthy adults in the Dallas Lifespan Brain Study who underwent cognitive testing, Aβ‐PET imaging, and fMRI [for details on recruitment see Rodrigue et al., 2012]. All participants were right‐handed and fluent English speakers with normal or corrected‐to‐normal vision (at least 20/30), and if necessary, vision was corrected using MR‐compatible corrective lenses during fMRI. Participants were cognitively normal (Mini Mental Status Exam (MMSE) ≥26; Folstein et al., 1975] with no history of neurological or psychiatric conditions, head trauma, drug or alcohol problems, or significant cardiovascular disease. Informed consent was obtained in accordance with protocol approved by the University of Texas at Dallas and the University of Texas Southwestern Medical Center.

Forming participant groups

We computed a single mean Aβ score from eight cortical regions and used this value to isolate our high‐ and low‐Aβ adults (calculation of Aβ levels described in detail in “PET Processing”). First we identified 23 high‐Aβ adults using an iterative outlier approach, which identified and then removed those participants with elevated Aβ (greater than the upper quartile value plus the interquartile range) until no more outliers remained in our original sample of 142 [Mormino et al., 2012b; Aizenstein et al., 2008]. We then selected 23 low‐Aβ adults from the remaining participants who were age‐, sex‐, and education‐matched to the 23 high‐Aβ adults. Specifically, all high‐ and low‐Aβ participants were matched within approximately 2 years of each other, with a range of .49–3.64 years. Finally, all participants under the age of 40 years old in the sample of 142 (n = 16) were included as a young adult control group (Table 1 provides sample demographics; see also Supporting Information Fig. 1).

Table 1.

Sample demographic information by group

	Young adults	Low amyloid	High amyloid
N (n male)	16 (7)	23 (10)	23 (10)
Age	33 ± 2.53	77 ± 8.99	77 ± 9.02
Age range	30–38	57–89	58–88
MMSE	28.75 ± 1.18	27.74 ± 1.32	27.65 ± 1.15
Education	17.69 ± 2.75	15.87 ±3.04	16.39 ± 2.54
Cortical SUVR	1.08 ± 0.05	1.13 ± .06	1.46 ± 0.17
SUVR range	1.03–1.19	1.00–1.22	1.26–1.77

Note. Mean ± SD; MMSE – Mini Mental State Exam; SUVR – Standardized Uptake Value Ratio

An analysis of variance (ANOVA) revealed that the three groups did not differ significantly on education level, F(2,59) = 2.06, P = 0.14, but did show a small difference on MMSE score, F(2,59) = 4.42, P = 0.02. Post hoc t‐tests with Šidák corrections showed that the MMSE difference was due to higher scores in younger adults compared to both older adult groups, Ps < 0.05, with no significant difference between high‐Aβ and low‐Aβ older adults, P = 0.99. Moreover, by design there were significant differences in mean Aβ burden, F(2,59) = 71.05, P < 0.001. Post hoc t‐tests with Šidák corrections showed both the young adult and low‐Aβ group had significantly less Aβ than the high‐Aβ older adult group, Ps < 0.001. Notably, there was no difference in Aβ burden between young and low‐Aβ older adults, P = 0.38 (Table 1).

Measurements of Cognition

Participants completed two tasks each in five different domains of cognition (described below). To make individual scores across tasks comparable, scores on each task were normalized via z‐score transformation for the current sample (N = 62). Next, the two tasks per each domain were averaged in order to form composite scores for each of five cognitive domains. Cronbach's alpha coefficients were computed to verify that the tasks selected for each domain showed high internal reliability.

Processing speed

Processing Speed (Cronbach's α = 0.87) was an index of how rapidly an individual could process information and was measured by number of items completed on the Digit Comparison [Hedden et al., 2002] and Wechsler Adult Intelligence Scale (WAIS) Digit Symbol [Wechsler, 1997].

Working memory

Working Memory (Cronbach's α = 0.69) measured ability to simultaneously manipulate and store information and was measured by number of items correct on the WAIS Letter Number Sequencing [Wechsler, 1997] and absolute span (sum of the number of words in each set on perfectly recalled trials) from the Operation Span Task [Turner and Engle, 1989].

Verbal recall

Verbal Recall (Cronbach's α = 0.73) assessed ability to remember verbal information over time and was measured by number of items recalled from two 12‐word‐list immediate recall tasks: the Hopkins Verbal Learning Task [Brandt, 1991] and the Verbal Recognition Memory Task from the Cambridge Neuropsychological Test Automated Battery [Robbins et al., 1994].

Reasoning

Reasoning (Cronbach's α = 0.74) described an individual's ability to recognize and apply novel rules to solve problems and was measured with number of correct answers on the Raven's Progressive Matrices [Raven et al., 1998] and Educational Testing (ETS) Services Letter Sets [Ekstrom et al., 1976].

Crystallized Intelligence

Crystallized intelligence (Cronbach's α = 0.92) provided an estimate of an individual's world knowledge and was measured by number of items correct on two vocabulary tasks: ETS Advanced Vocabulary Sets [Ekstrom et al., 1976] and Shipley Vocabulary [Zachary, 1986].

Measurement of Aβ with PET Imaging

Participants underwent PET imaging with the ¹⁸F‐Florbetapir ligand to measure in vivo Aβ deposition which was used as the basis for forming our high‐ and low‐Aβ older adult groups.

PET acquisition

Participants were injected with a 10mCi bolus of ¹⁸F‐Florbetapir and scanned exactly 50 min postinjection on a Siemens ECAT HR PET scanner. A 10‐min dynamic emission scan was acquired in two 5‐min frames, immediately followed by a 7‐min internal rod source transmission scan. The transmission image was reconstructed with back projection and a 6 mm³ full‐width‐at‐half‐maximum (FWHM) Gaussian filter. Emission images were processed by iterative reconstruction, four iterations and 16 subsets with a 3 mm³ FWHM ramp filter.

PET processing

Each participant's PET scan was spatially normalized to a Florbetapir uptake template (2 mm³ isotropic voxels; Clark et al., 2011] positioned in Monteal Neurological Institute (MNI) space using Statistical Parametric Mapping 8 (SPM8; Wellcome Department of Cognitive Neurology, London, UK) and in‐house MATLAB scripts (Mathworks, Natick, MA). In order to assess Aβ burden across a wide sampling of cortical areas, standardized uptake value ratios (SUVRs) were computed by normalizing counts from eight bilateral cortical regions (precuneus, posterior and anterior cingulate, dorsolateral prefrontal, orbital frontal, temporal, parietal, and occipital cortices) to cerebellar hemisphere (excluding peduncles). These eight cortical regions were selected because they are regions in which in vivo Aβ deposition has been shown to discriminate Alzheimer's patients from healthy controls [Clark et al., 2011], and these regions are commonly used in Aβ‐imaging studies of preclinical AD [Clark et al., 2011; Rodrigue et al., 2012]. All regions were defined using Automated Anatomical Labeling (AAL; Tzourio‐Mazoyer et al., 2002] masks modified to minimize nonspecific white matter binding [see Rodrigue et al., 2012 for more detail and an illustration of these regions]. Total mean cortical SUVR was calculated by averaging SUVR across the eight cortical regions of interest.

Measurement of Face Processing with fMRI

fMRI task design

The face processing task was a passive viewing task in which participants were instructed to look at grayscale photographs of faces in‐scanner. The face task occurred during the first 14 min of a longer neuroimaging session. While in‐scanner, participants viewed 64 different faces (including images of males and females of a variety of ages and races; Minear and Park, 2004] in a block design across two runs. Within each run there were four blocks of face stimuli, with eight faces randomly presented in each block. Individual images were presented for 2 s each with no interstimulus interval. We note that blocks of other image categories were included in the scanning session (see Park et al. 2012 for more detail); however, the current analysis was limited to neural response to viewing images of faces as previous studies have reported Aβ‐related disruptions in neural encoding of face stimuli [Huijbers et al., 2014; Sperling et al., 2009; Vannini et al., 2012b].

fMRI acquisition

Functional brain images were obtained with Philips Achieva 3T whole body scanner equipped with an 8‐channel head coil, using the Philips SENSE parallel acquisition technique. High‐resolution anatomical images were collected with a T1‐weighted MP‐RAGE sequence with 160 sagittal slices, 1 × 1 × 1 mm³ voxels; 256 × 256 × 160 matrix, field of view (FOV) = 220 mm, TE = 3.76 ms, TR = 8.18 ms, FA = 12°. Blood oxygen level dependent (BOLD) fMRI data were acquired using a T2*‐weighted echo‐planar imaging sequence with 43×64×64 interleaved axial slices per volume acquired parallel to the AC‐PC line, 3.4 × 3.4 × 3.5 mm³ voxels, FOV = 220 mm, TE = 25 ms, TR = 2 s, FA = 80°.

fMRI preprocessing

Individual participant's time series data were preprocessed with SPM8 according to a standard pipeline of procedures. First, images were corrected for differences in slice time acquisition and within‐run participant movement. Participants were not considered for the present study if they showed head movement greater than 3.4 mm translation in any direction (equivalent to 1 voxel) between scans. Next, images were normalized by segmentation to standard MNI space and resampled into 3 mm³ isotropic voxels. Finally, images were smoothed with an isotropic 8 mm³ FWHM Gaussian kernel. For each individual, voxels in ventral visual cortex were isolated using an AAL mask of four combined bilateral regions: fusiform, lingual, inferior temporal, and parahippocampal gyri.

The BOLD images were further processed following a standard series of preprocessing steps described by Abdi et al. [2012b]. First, for each participant, a scan × voxel matrix was created for BOLD data acquired during scanning. Using built‐in functions in MATLAB, time‐series data for each individual were detrended to remove any linear time correlations in the volumes. Next, each individual's volumes were centered (by subtracting the mean scan value from each scan) and normalized by the first singular value to ensure the analysis was not dominated by outlier individuals with highly variable time‐series [Abdi et al., 2013]. The average activation pattern associated with human faces was calculated for each participant by averaging scans associated with viewing faces (taking into account a 4 s lag in hemodynamic response) to create a final data matrix in which each row corresponds to one participant and each column to one voxel (i.e., Participants × Voxels matrix).

Multivariate Analysis of fMRI Data

For the present analysis we conducted Barycentric Discriminant Analysis (BADA) and inference tests with the TInPosition [Beaton et al., 2013; see also Beaton et al., 2014] package in R (R Core Team, Vienna, Austria). BADA was particularly suited to identify subtle differences between groups of participants through the use of a between‐class (i.e., group‐level) PCA [also equivalent to mean‐centered partial least squares; Krishnan et al., 2011; McIntosh et al., 2004]. With BADA we were able to identify orthogonal components (referred to as factors) of neural activity that characterized different groups. Specifically, BADA allowed us to identify group‐based differences in neural activity during the in‐scanner passive viewing face task.

To perform the analysis, we started with two data matrices: (1) Participants × Voxels (of BOLD activity during passive face viewing) and (2) Participants × Group Membership. With these two matrices, a Group × Voxels matrix (also referred to as the barycenter matrix) was computed that equally weighted the means for each group to ensure that the analysis was not influenced by differently sized groups. The Group × Voxels matrix was then subjected to the generalized singular value decomposition, which isolated two orthogonal patterns of neural activity that maximally discriminated among the three participant groups. For each of the two orthogonal factors, BADA produced three descriptive measures: (1) group factor scores (termed barycenters) which were a measure of how each group's pattern of neural activity contributed to the factor, (2) participant factors scores which described how each individual participant's pattern of neural activity was reflected on the factor, and (3) voxel factor scores which described how individual voxels contributed to each factor. For more detail on how the matrices resulting from the generalized singular value decomposition were used to compute these three descriptive measures see Supporting Information Appendix A and the work of Abdi [2007] and cowokers [Abdi et al., 2012a; Beaton et al., 2014].

Inference Testing

To assess the stability and reliability of the model, we conducted several nonparametric inference tests via permutation [Berry et al., 2011; McIntosh et al., 1996; Nichols and Holmes 2002], the bootstrap (Efron and Tibshirani, 1986; Hesterberg, 2011], and leave‐one‐out cross‐validation. Each of these procedures is described in detail below.

Permutation resampling

Permutation resampling allowed us to test the significance of the model's factor structure by providing a nonparametric approach to test the null hypothesis. In permutation, participant group association was randomly assigned, and the analysis was rerun 1,000 times to create a nonparametric distribution on which to test the original model. We used the permutation procedure to test significance of the model's omnibus variance and R ² to determine if the model could reliably separate patterns of neural activity associated with each group. Permutation was also used to test if the correlations between participant factor scores and cognitive domains were reliable.

Bootstrap resampling

The bootstrap procedure assessed whether there were significantly different patterns of neural activity across the three participant groups. The bootstrap procedure was conducted by resampling participant factor scores 1,000 times with replacement in order to create 95% confidence intervals (CIs) around the group means (i.e., barycenters). If CIs for two groups did not overlap, then those groups were considered significantly different at P < 0.05 [Abdi et al., 2009]. Bootstrap resampling also allowed us to determine which voxels significantly contributed to differences between the participant groups through the computation of a t‐like statistic called the Bootstrap Ratio (BSR; Krishnan et al., 2011; McIntosh and Lobaugh, 2004; McIntosh and Mišic, 2013]. BSRs were computed as the ratio of each participant's voxel factor score to an estimate of standard error generated through the resampling procedure. In the current analysis, we utilized a BSR cutoff of three (analogous to approximately P < 0.0013) to determine which voxels significantly contributed to the factor structure.

Leave‐one‐out cross‐validation

Two classification values assessed how accurately individual factor scores predicted an individual's group‐membership. The first estimate was a fixed‐effects (i.e., observed) model, and the second estimate was a random‐effects model. The fixed‐effects model provided an accuracy estimate specific to the current dataset, whereas the random‐effects analysis extrapolated the findings to the inclusion of “left‐out” participants (via leave‐one‐out cross‐validation). In leave‐one‐out cross validation, each participant was removed in turn and a new BADA performed on the data, excluding the left‐out observation. The group membership of the left‐out participant was then predicted based on a model from which they were excluded, thus providing classification accuracy for how well the model generalized to predicting the group status of “new” participants.

RESULTS

Barycentric Discriminant Analysis Results

Because we had three groups of participants, BADA produced two factors: Factor 1 (79.4% of total variance) and Factor 2 (20.5% of total variance), which are described below. First, we note that the omnibus multivariate factor structure was reliable and systematic, which we assessed using a 1,000 sample permutation test, P < 0.001. Additionally a permutation test of the within‐group variance versus the between‐group variance showed a significant R ² value of .39, P = 0.02, which indicated that the factor structure describing different patterns of neural activity during face processing was stable, and the a priori group assignment of individuals was reliable.

Factor 1: Effect of age

Not surprisingly Factor 1, which explained 79.4% of total variance, was driven by differences in patterns of neural activity between young and both older adult groups (Fig. 1a). Young adults showed significantly different patterns from both older adult groups at P < 0.05, as evidenced by the finding of no overlap of the 95% CIs. The two older adults groups did not differ from one another on this factor. The BSR test of the neural activity in each voxel illustrated that the difference between young and older adults was characterized by widespread differences in posterior bilateral lingual, right fusiform, bilateral inferior temporal gyrus, and anterior regions of bilateral parahippocampal gyrus (Fig. 1b). Table 2 lists significant clusters going from posterior to anterior brain slices for clusters of at least 10 voxels displaying a BSR with a magnitude (plus or minus) greater than three—the nonparametric equivalent a t‐score greater than three (P ≈ 0.0013). For Factor 1, positive BSRs indicated patterns of neural activation more similar to the mean pattern of young adults, whereas negative BSRs indicated patterns more similar to both older adult groups.

Figure 1.

Factor 1 – Effect of Age. The factor scores for each group are plotted for Factor 1 (Effect of Age) and the error bars represent 95% confidence intervals created by bootstrapping the factor scores 1000 times. Voxels displaying BSR magnitude greater than 3 have been mapped, with the heatmap indicating the magnitude of the voxel's BSR. On Factor 1, the difference between young adults and both groups of older adults accounted for the most variance in the data (A) characterized by widespread differences in neural activity (B).

Table 2.

Significant voxels clusters for each factor

MNI coordinates (mm)			Region	BA	Peak BSR	Cluster size
X	Y	Z
Factor 1 (Effect of age)
−24	−96	−18	L Lingual	18	5.44	100
18	−90	−15	R Lingual	18	6.16	104
−15	−72	0	L Lingual	18	5.58	27
30	−63	−9	R Fusiform	19/37	−7.05	42
54	−51	−15	R Inferior temporal	20/37	6.91	113
30	−36	−21	R Fusiform	37	5.78	23
−18	−24	−24	L Parahippocampal	30	9.45	117
54	−24	−24	R Inferior temporal	20	6.50	73
21	−21	−24	R Parahippocampal	30	5.91	29
36	−9	−33	R Fusiform	20	5.31	23
−54	−6	−30	L Inferior Temporal	20	4.85	65
Factor 2 (Effect of Aβ beyond age)
−24	−72	−7	L Fusiform/Lingual	18/19	−5.58	14
−30	−39	−11	L Fusiform	20	−5.28	59

Note. Minimum cluster size = 10; Minimum BSR = magnitude of 3; BSR – Bootstrap Ratio; Region labels were derived based on comparison to a standardized AAL template

Factor 2: Effect of Aβ beyond age

Importantly, Factor 2, which explained 20.5% of the variance, was driven by differences between low‐ and high‐Aβ older adult groups, confirming our hypothesis that there were differences in the neural signature in ventral visual cortex for high‐ and low‐Aβ adults. The two older adult groups showed significantly different neural patterns associated with viewing faces as revealed by the nonoverlapping 95% CIs (Fig. 2a). As predicted, high‐ and low‐Aβ older adults showed activation differences in a region of ventral visual cortex involved in face perception, specifically the left fusiform gyrus. An additional smaller cluster of neural activity was found on the border of left lingual and fusiform gyrus (Table 2). For Factor 2, positive BSRs indicated patterns of neural activation more similar to the mean pattern of low‐Aβ adults, whereas negative BSRs indicated patterns more similar to high‐Aβ adults. Further analysis via a comparison of mean activity in high‐ and low‐Aβ adults indicated that, on average, high‐Aβ adults showed less activity in left fusiform compared to low‐Aβ adults. Because this factor was orthogonal to the first factor, these neural patterns can be considered to be specific to the two older adult groups and independent of general age effects.

Figure 2.

Factor 2 – Effect of Aβ beyond Age. Factor 2 represented the differences in neural activity for older adults with elevated Aβ compared to those without (A). On average, older adults with high levels of Aβ deposition showed decreased activity in left fusiform gyrus compared to older adults with low Aβ (B).

Figure 3.

Correlation between participant factor scores on Factor 2 and processing speed. Participant Factor 2 scores (which described how much each participant expressed fusiform activity) for each older adult were transformed to z‐scores and correlated with the composite measure of performance on processing speed. Lower Factor 2 scores corresponded with decreased activity in fusiform. [Color figure can be viewed in the online issue, which is available at http://wileyonlinelibrary.com.]

Testing model classification accuracy

Both fixed‐ and random‐effects classification accuracies were tested with a χ ² test of independence. Fixed‐effects accuracy was 64.5% and was significantly above chance (34.2%), χ ² (4) = 32.90, P < 0.001. Leave‐one‐out cross‐validation test produced a random‐effects classification accuracy of 48.4% that was significantly above chance, χ ² (4) = 14.43, P < 0.01. In general, our model could systematically predict group membership for left‐out participants, even though the model was based on a relatively small sample.

Behavioral Results

First we assessed whether the three groups differed in cognitive performance by conducting a one‐way ANOVA on each composite domain of cognition (Supporting Information Table I). There was a significant main effect of group for each domain, and post hoc comparisons with Šidák corrections revealed that this difference was due to younger adults performing significantly better than both older adult groups on tasks of Processing Speed (Ps < 0.001), Reasoning (ps< 0.01), Episodic Memory (ps< 0.001), and Working Memory (ps< 0.01). Both older adult groups showed significantly better performance on tasks of Crystallized Intelligence (ps < 0.05) when compared to the young adult group which was consistent with theories of cognitive aging for typically aging adults [Park et al., 2002]. There were no significant group differences on any domain of cognition between older adults with high‐ and low‐Aβ, (ps > 0.4).

The Relationship of Neural Factor Patterns to Cognition

Next, we explored the relationship between cognitive performance and the patterns of neural activity identified by the multivariate analysis. Using an individual differences approach, each participant's factor score (which measured the extent to which each participant expressed the factor's neural pattern) was examined in relation to cognitive performance. We found that participant factor scores for the first factor (which represented an age effect) were significantly correlated with all domains of cognition (except crystallized intelligence) as tested via a 1,000 sample permutation of the correlation (Supporting Information Table II). Participants with positive factor scores (i.e., those who showed neural patterns more like the average young adults) performed better on Processing Speed (r= 0.34, P < 0.001), Reasoning (r = 0.26, P < 0.05), Episodic Memory (r= .40, P < 0.01), and Working Memory (r = 0.31, P < 0.05). This finding confirmed that the neural patterns on Factor 1 were also related to the general effects of cognitive aging.

Next, we computed the correlations between cognitive scores and participant factor scores on Factor 2, a factor that described neural differences between the two older adult groups with high or low Aβ. The correlations were restricted to only participants in the two older adult groups (n = 46). The results indicated that the extent to which an individual expressed a pattern of neural activity that was similar to that of high‐Aβ adults (i.e., decreased activity in fusiform) was related to worse performance on measures of Processing Speed (r = 0.42, P < 0.003), even after controlling for age (r = 0.30, P = 0.048) and mean cortical SUVR (r = 0.29, P = 0.049; Fig. 3; Supporting Information Table II). This activation‐cognition association appeared to be selective to processing speed, as Factor 2 participant scores were not significantly related to the other cognitive domains.

DISCUSSION

We report three main findings in this study. First, we confirmed our hypothesis that Aβ deposition in clinically normal adults was associated with differences in neural activity underlying face perception. Specifically, we found decreased neural activity for high‐Aβ compared to demographically matched low‐Aβ adults in the left fusiform gyrus, an important cortical region associated with face perception. Second, in line with prior work [Burianová et al., 2013; Carp et al., 2011], we report that, regardless of Aβ status, older adults showed significantly different patterns of neural activity across the ventral visual pathway when compared to younger adults. Finally, we found that the degree to which an older adult expressed a pattern of decreased activity in left fusiform gyrus was significantly correlated with poorer performance on tasks of processing speed. Each finding is discussed in turn below.

High‐ and Low‐Aβ Older Adults Differed in Activation While Viewing Faces

The present findings provide evidence that the effects of Aβ on neural activity may be more widespread than have been previously demonstrated. Thus far, prior studies examining the effects of Aβ on face‐memory in clinically normal adults have shown that elevated Aβ is related to decreased connectivity in default mode regions, particularly those related to episodic memory [Hedden et al., 2009; Huijbers et al., 2014; Mormino et al., 2011; Sheline et al., 2010]. Additionally, Aβ has been implicated in: (1) poor modulation of default mode activity [Sperling et al., 2009], (2) failure to show repetition suppression to repeated trials [Vannini et al., 2012b], and (3) decreased modulation of entorhinal cortex [Huijbers et al., 2014]. All of these findings have been interpreted as a breakdown in neural activity associated with face encoding, which may be related to behavioral deficits in face‐name memory associated with high Aβ burden [Rentz et al., 2011].

Our results showed that Aβ deposition may affect neural activity in a region of ventral visual cortex involved in face perception. Specifically, we report that during face viewing, the left fusiform gyrus showed less activation in older adults with elevated Aβ compared to demographically matched adults without Aβ elevation. In young adults, successful face recognition is correlated with heightened neural activity in both perceptual (e.g., fusiform) and memory regions of the cortex [i.e., hippocampus; Golby et al., 2001]. Therefore, Aβ‐related decreases in fusiform activity, in conjunction with prior findings of neural disruptions during face encoding, may jointly contribute to face memory deficits in preclinical AD. An important focus for a future study would be to relate decreased Aβ‐related activity in the fusiform to poorer face memory performance in‐scanner.

The multivariate analysis we conducted isolated Aβ‐related neural patterns that were lateralized in left fusiform gyrus. It is clear from research in young adults that neural representations underlying faces are widely distributed across higher‐order visual cortex [Haxby et al., 2001], but engagement of particular core‐regions, including left fusiform, is necessary for face perception [Ishai, 2008]. Here, we report that older adults with elevated Aβ deposition show left‐lateralized fusiform decreases in neural activity when compared to older adults with low Aβ. We have reported in previous work greater age‐related decreases in neural selectivity for faces in left fusiform activity compared to right fusiform during a face viewing task. Together, these findings suggest that left fusiform activity may be susceptible to brain changes associated with aging, including increased Aβ deposition.

The notion that neural activity underlying face perception is affected by Aβ is congruent with the little evidence available in clinical populations. For example, compared to healthy controls, MCI patients failed to show face‐selective activation in fusiform gyrus during a face‐matching versus location‐matching task [Bokde et al., 2008] and showed weaker connectivity between left and right fusiform during face‐matching [Bokde et al., 2006]. These findings provide some evidence that neural activity in specialized cortical regions involved in face perception is disrupted in early stages of AD. Similarly, patients with AD showed a trend for greater right fusiform activity during a passive face‐viewing task compared to an older adult control group [Sauer et al., 2006]. Although this contrasts with our finding of decreased activation in left fusiform, we note that without in vivo measures of Aβ in these previous studies, it is not possible to determine whether the normal adult controls were truly free of AD neuropathology. In sum, the present findings and clinical observations suggest that neural activity during face‐processing could be affected in early stages of AD, but additional research including healthy adults (both with and without Aβ) in addition to individuals in early stages of AD is needed to examine differences in face perception as disease severity progresses.

Young and Older Adults Differed in Activation While Viewing Faces

Next, the present findings demonstrate clearly that age plays a critical role in neural activity underlying face perception. We report that the greatest amount of variance in the fMRI data was due to differences in neural patterns between young and older adults in multiple cortical regions including fusiform, lingual, inferior temporal, and parahippocampal gyri. These findings are congruent with previous studies using both univariate and multivariate techniques to examine age‐differences in neural activity underlying face perception. Studies using univariate techniques have reported that compared to young, older adults had less selective activation for faces [Voss et al., 2008; Park et al., 2010, 2012], less neural adaptation to identical faces [Lee et al., 2011] and greater neural adaptation to nonidentical faces [Goh et al., 2010] in face‐responsive regions of fusiform. Likewise, studies that use data‐driven analyses to examine neural patterns across large regions of cortex have reported that during face perception, young, and older adults showed robust differences in occipitotemporal cortices, including fusiform, inferior temporal and parahippocampal gyri [Burianová et al., 2013; Carp et al., 2011]. These data‐driven analyses also found differences in neural patterns that extended beyond ventral visual cortex to frontal [Burianová et al., 2013] and prefrontal regions [Carp et al., 2011]. In sum, our current study, in conjunction with previously mentioned work, suggests that age‐related differences in face processing are widely distributed across ventral visual cortex, whereas Aβ deposition may be selectively related to differences in left fusiform cortex.

Decreased Left Fusiform Activation Predicts Slowed Processing Speed

Finally, our third major finding was that participants who exhibited a pattern of neural activity characteristic of high‐Aβ adults showed poorer cognitive performance. This is a particularly interesting finding because the effects of Aβ on cognition in clinically normal adults have been demonstrated to be quite subtle [Hedden et al., 2013]. In the current study, even though adults with high‐ and low‐Aβ did not show group‐level differences on processing speed performance, individual differences in neural activity were sensitive to differences in processing speed, even when accounting for age and mean Aβ burden. Specifically, a lower Factor 2 score (which indicated a neural pattern more characteristic of high‐Aβ adults) was selectively related to slower speed of processing. Our findings contribute to limited existing knowledge of the relationship between functional activation and Aβ deposition and suggest that continuous measures may be more sensitive to cognitive differences in clinically normal adults than group comparisons.

Limitations

We note some limitations of the current study. First, the equipment at our scanning center did not allow us to measure eye‐gaze during the passive viewing task; therefore we cannot determine which portion of each face‐stimulus participants focused on during the scan. Nonetheless, cameras located inside the scanner did allow us to monitor participants' faces to ensure they remained attentive to the presentation screen. Second, the passive‐viewing design of our fMRI task did not provide us with in‐scanner behavioral measures with which to relate group‐differences in neural patterns. Future work incorporating a simple measure of reaction time (e.g., responding to gender of the face stimulus) would make it possible to examine the relationship between in‐scanner measures of processing speed and neural patterns. Finally, the cross‐sectional design of the current study limited the interpretation of meaning of elevated Aβ deposition in clinically normal adults. Longitudinal research including healthy adults (both with and without Aβ) is needed to examine how longitudinal change in Aβ deposition influences neural activity underlying face perception in preclinical stages of AD.

Conclusions

FMRI has been proposed as a potentially powerful biomarker for detecting neural dysfunction that may be associated with preclinical AD [Sperling, 2011] because fMRI enables us to examine networks of neural activity. Our present results suggest clinically normal older adults with elevated Aβ show significantly different patterns of neural activity in a region of ventral visual cortex involved in face perception during a passive face‐viewing task. Similar fMRI paradigms may be advantageous for tracking changes in neural dysfunction in a preclinical AD population—unlike memory tasks, which may become too difficult as a participant advances in the disease process, passive viewing tasks require little training or cognitive effort. Furthermore, the present study demonstrated the advantages of using multivariate techniques to assess potential markers of AD neuropathology. With the present analysis we were able to generate a model of neural activation based on known group memberships (i.e., high‐Aβ old, low‐Aβ old, and young), and then use that model to test group‐membership of new participants. The current work offers a platform from which to find new neural markers of preclinical AD, and a next step would be to include patients with MCI and AD to assess what neural patterns associated with face perception characterize different stages of AD. In sum, the present findings suggest that even highly selective brain regions in the ventral visual pathway are affected by Aβ deposition and that this may be an important domain for understanding the impact of Aβ on cognitive aging. Longitudinal follow‐up is underway to determine whether or not this reduction in neural activity is a reliable biomarker of progression of Aβ deposition.

Supporting information

Supplementary Information

Supplementary Information

Supplementary Information

Supplementary Information