Abstract
Increasing age is associated with age-related neural dedifferentiation, a reduction in the selectivity of neural representations, which has been proposed to contribute to cognitive decline in older age. Recent findings indicate that when operationalized in terms of selectivity for different perceptual categories, age-related neural dedifferentiation and the apparent age-invariant association of neural selectivity with cognitive performance are largely restricted to the cortical regions typically recruited during scene processing. It is currently unknown whether this category-level dissociation extends to metrics of neural selectivity defined at the level of individual stimulus items. Here, we examined neural selectivity at the category and item levels using multivoxel pattern similarity analysis (PSA) of fMRI data. Healthy young and older male and female adults viewed images of objects and scenes. Some items were presented singly, while others were either repeated or followed by a “similar lure.” In agreement with recent findings, category-level PSA revealed robustly lower differentiation in older than in younger adults in scene-selective, but not object-selective, cortical regions. By contrast, at the item level, robust age-related declines in neural differentiation were evident for both stimulus categories. Additionally, we identified an age-invariant association between category-level scene selectivity in the parahippocampal place area and subsequent memory performance, but no such association was evident for item-level metrics. Lastly, category- and item-level neural metrics were uncorrelated. Thus, the present findings suggest that age-related category- and item-level dedifferentiation depend on distinct neural mechanisms.
Significance Statement
Cognitive aging is associated with a decline in the selectivity of the neural responses within cortical regions that respond differentially to distinct perceptual categories (age-related neural dedifferentiation). However, prior research indicates that while scene-related selectivity is reduced in older age and is correlated with cognitive performance independently of age, selectivity for object stimuli is typically not moderated by age or memory performance. Here, we report that neural dedifferentiation is evident for both scene and object exemplars when it is defined in terms of the specificity of neural representations at the level of individual exemplars. These findings suggest that neural selectivity metrics for stimulus categories and for individual stimulus items depend on different neural mechanisms.
Introduction
Age-related neural dedifferentiation, reduction in neural selectivity (differentiation) with increasing age, has been proposed to play a role in age-related cognitive decline (Koen and Rugg, 2019; Koen et al., 2020). Evidence suggests that greater neural selectivity at encoding is predictive of better memory performance (Yassa et al., 2011; Berron et al., 2018; Bowman et al., 2019; Koen et al., 2019; Sommer et al., 2019; Srokova et al., 2020) and higher scores on psychometric tests of fluid processing (Park et al., 2010; Koen et al., 2019) and might be associated with lower levels of cortical tau deposition in cognitively healthy older adults (Maass et al., 2019). However, despite the apparent functional significance of reduced neural selectivity in older age, the underlying mechanisms remain poorly understood.
Age-related neural dedifferentiation is frequently reported when operationalized in terms of the selectivity of neural responses elicited by different categories of visual stimuli (e.g., images belonging to categories such as scenes, objects, or faces; here termed category-level differentiation). Age-related decline in selectivity has almost invariably been reported for scene images (Voss et al., 2008; Carp et al. 2011; Zheng et al., 2018; Koen et al., 2019; Srokova et al., 2020), while effects of age on selectivity for faces and objects are more variable (for null effects, see Chee et al., 2006; Voss et al., 2008; Zheng et al., 2018; Koen et al., 2019; Payer et al., 2006; Srokova et al., 2020; but see also Park et al., 2004, 2012; Voss et al., 2008; Zebrowitz et al., 2016). There are numerous possible factors that might account for these inconsistent findings. For example, significant age effects for objects and faces often arise during passive viewing tasks, raising the possibility that older adults “zone out” and pay less attention to stimulus events if the task does not require them to attend actively to each stimulus. Alternatively, during more attentionally demanding tasks, reduced selectivity for scenes but not for objects may arise due to categorical differences in the attentional control required for object and scene processing (for further discussion, see Srokova et al., 2020). These and other explanations remain unexplored; hence, it is currently unknown why age-related neural dedifferentiation is more likely to be observed for scenes than other perceptual categories.
What is the driver of age-related neural dedifferentiation at the category-level? One obvious possibility is that it reflects a decline in the specificity of the neural responses elicited by individual stimulus exemplars. There is, however, no empirical evidence in support of this assumption. Whereas category-level dedifferentiation has been extensively documented in prior work, evidence for reduced selectivity at the item level is sparse. Two prior studies employing multivoxel pattern analyses reported evidence of an age-related reduction in neural similarity between individual stimuli and their exact repetitions (Bowman et al., 2019; Trelle et al., 2019). In another study, which employed a univariate approach, Goh et al. (2010) reported that while repetition adaptation effects in younger adults were restricted to repeated presentations of the same face, older adults demonstrated repetition adaptation both for exact repeats and for faces that closely resembled a previously presented face. Repetition effects for similar items are thought to be a consequence of re-accessing a neural representation that is shared with or partially overlaps with the first presentation; thus, Goh et al.’s findings are suggestive of an age-related reduction in neural selectivity. However, contrary to the findings reported in the foregoing studies, two other studies reported null effects of age on item-level neural selectivity (St-Laurent and Buchsbaum, 2019 and after controlling for category-level effects, Zheng et al., 2018).
As discussed above, the factors that account for reduced neural selectivity in older age at the category level are largely unknown. To elucidate the neural mechanisms underlying age-related neural dedifferentiation, the present study examined whether age differences in neural selectivity at the category level extend to neural selectivity operationalized at the level of individual stimulus exemplars. Younger and older adults underwent fMRI as they completed a “mnemonic similarity task” (Stark et al., 2019). During the scanned encoding phase, participants viewed images of objects and scenes. Each novel image was followed either by an exact repetition of the image, a visually similar exemplar, or served as a test item in a subsequent unscanned memory test. Neural differentiation was quantified at both the category and item levels with multivoxel pattern similarity analysis (PSA). Following prior findings (Koen et al., 2019), we predicted that neural differentiation at the category level would be moderated by age in scene-selective, but not object-selective, cortical regions. Under the assumption that item- and category-level differentiation reflects similar neural mechanisms, we would expect to identify analogous age differences in item-level differentiation for scene but not for object images. An open question concerns the patterning of any item-related dedifferentiation effects. Given the findings of Goh et al. (2010) described above, these effects would be expected to include an increase in the neural similarity between novel items and their similar lures, suggestive of a broadening of neural representations. Alternatively, or additionally, age-related neural dedifferentiation at the item-level might manifest as a reduction in neural similarity between novel items and their exact repeats, consistent with the proposal that reduced neural selectivity reflects an age-related increase in neural noise (cf. Li et al., 2001).
Materials and Methods
Participants
Twenty-five young and 25 older adults participated in the study. One young and one older adult were excluded from the analyses due to incidental MR findings, resulting in a final sample of 24 young and 24 older adults. Participants were recruited from the University of Texas at Dallas and from the surrounding Dallas metropolitan area and were compensated for their time at a rate of $30/h and up to $30 for travel. Demographic information and neuropsychological test performance for the final sample are reported in Table 1. Participants were right-handed, had normal or corrected-to-normal vision, and were fluent English speakers before the age of 5. None of the participants had a history of neurological or psychiatric disease, substance abuse, diabetes, or current or recent use of prescription medication affecting the central nervous system. All participants undertook a neuropsychological test battery prior to the MRI session, and a set of inclusion and exclusion criteria were employed to minimize the likelihood of including older participants with mild cognitive impairment or early dementia (see below). All participants provided written informed consent before participation in accordance with the requirements of the Institutional Review Board of the University of Texas at Dallas.
The outcome of the neuropsychological test battery in younger and older adults
Neuropsychological testing
Participants completed our laboratory’s standard neuropsychological test battery on a separate day prior to the MRI session. The assessment battery consists of the Mini-Mental State Examination (MMSE), the California Verbal Learning Test II (CVLT; Delis et al., 2000), Wechsler Logical Memory Tests 1 and 2 (Wechsler, 2009), the Symbol Digit Modalities Test (SDMT; Smith, 1982), the Trail Making Tests A and B (Reitan and Wolfson, 1985), the F-A-S subtest of the Neurosensory Center Comprehensive Evaluation for Aphasia (Spreen and Benton, 1977), the Forward and Backward digit span subtests of the revised Wechsler Adult Intelligence Scale (WAIS; Wechsler, 1981), the Category Fluency test (Benton, 1968), Raven’s Progressive Matrices List I (Raven et al., 2000), and the Test of Premorbid Functioning (TOPF; Wechsler, 2011). Participants also completed a visual acuity test using ETDRS charts, assessed using the logMAR metric (Ferris et al., 1982; Bailey and Lovie-Kitchin, 2013). Visual acuity was tested with corrective lenses, if prescribed. Participants were excluded prior to the fMRI session if they performed >1.5 SD below age norms on two or more non-memory tests, if they performed >1.5 SD below the age norm on at least one memory-based test, or if their MMSE score was <26. Neuropsychological test scores were missing for one participant (a younger adult male).
Experimental materials
Experimental stimuli were presented using PsychoPy v2021.1.3 (Peirce et al., 2019). The study phase was completed inside the MRI scanner, while the subsequent memory test was completed after exiting the scanner. During the study phase, stimuli were projected onto a translucent screen (41 cm × 25 cm; 1,920 × 1,080 pixels resolution) placed at the rear of the scanner bore and viewed via a mirror mounted on the head coil (viewing distance ∼105 cm). The post-scan memory test was administered on a Dell laptop computer equipped with a 17 inch display and a resolution of 1,920 × 1,080 pixels. All stimuli were presented on a gray background and consisted of images of objects and scenes which were resized to fit inside a frame subtending 256 × 256 pixels. The images of objects depicted items which could be considered either “man-made” (e.g., tools, chairs, clothes) or “natural” (e.g., fruits, plants, animals). Scene exemplars illustrated a variety of indoor (e.g., living room, hallway, bar) or outdoor spaces (e.g., park, field, beach).
The critical trials in the study phase comprised 168 scene trials and 168 object trials presented across nine scanner runs. For a given image category, 48 trials (“first presentation” trials) were either re-presented (“exact repeats,” 24 trials) or “repeated” as perceptually similar lures (“lures,” 24 trials). An additional 72 stimuli belonging to each image category were presented once only, and these items were used as test items in the subsequent memory task (see below). To ensure that neural similarity between the first presentation trial and its corresponding lure or exact repeat was not driven by within-session autocorrelation of the blood oxygen level-dependent (BOLD) time series (Mumford et al., 2014), lures and exact repeat trials were always presented in the subsequent scanner run while ensuring that the lag between first and second presentations conformed to a rectangular distribution ranging between 18 and 42 trials (mean = 30). Consequently, the first study run did not contain any repetition or lure trials, and the last run did not contain any first presentation trials; an additional 24 filler trials were randomly interspersed within the first and last runs to ensure that all runs contained an equal number of trials. One hundred and twenty-six null trials, comprising a white fixation cross presented at the center of the display, were randomly interspersed among the critical trials of the study phase. During the test phase, participants viewed a total of 108 images of scenes and 108 object images. For a given image category, 36 trials were repetitions of images that the participant had viewed during the study phase (“target”), 36 trials were images that were perceptually similar to previously viewed study images (“lure”), and 36 trials were presentations of new images.
The stimulus pool described above was used to create 24 stimulus lists which were assigned to yoked pairs of younger and older adults. For all stimulus lists, the stimuli were pseudorandomized such that participants viewed no more than three consecutive trials of the same visual category or trial type, and no more than two consecutive null trials.
Study and test phase
A schematic of the study and test tasks, including examples of the experimental stimuli, is illustrated in Figure 1. Participants received instructions for the study phase and completed a short practice run prior to entering the scanner. Each of the nine scanner runs of the study phase lasted 4 min and 38 s. A given study trial began with a red fixation cross presented at the center of the screen for 500 ms, followed by an image of an object or a scene for 2 s, and then a white fixation cross for an additional 2 s. Participants had a total of 4 s following the onset of the image to make “indoor/outdoor” judgments on the presented scene or object. The task instructions specified that the indoor/outdoor judgments were entirely subjective (there was no right or wrong answer) and that participants should choose the first response that came to their mind. The indoor and outdoor judgments were mapped to the index and middle finger of the right hand with finger assignment counterbalanced across participants. Responses were made using a scanner-compatible button box.
A, An illustration of representative trials from the study and test phases and B a schematic illustrating the trial types. Study trials were categorized into five trial types: tested trials (later employed in the retrieval task), first presentation trials (to be followed by an exact repeat or by a perceptually similar lure in the subsequent scanner run), and second presentation trials (their exact repeats and similar lures). At test, trials were binned into three trial types: target items (previously studied exemplars), lure items (similar to a studied exemplar), or new items. At study, participants made “indoor/outdoor” judgments, and at test they made “old/similar/new” judgments. C, Examples of similar lure exemplars for each of the two image categories (scenes and objects).
The instructions and a practice run for the test phase were administered immediately following the MRI session. Retrieval data for one younger adult participant was missing due to equipment malfunction. The test phase consisted of two blocks lasting ∼8 min each. Each trial of the test phase began with a red fixation cross for 500 ms, followed by the test image for 2 s and a white fixation cross for an additional 2 s, thus providing a 4 s response window. Participants were instructed to indicate whether the test image was either the same as one they had viewed at study (“old”), similar to an image they had viewed at study (“similar”), or a completely new exemplar (“new”). The three response alternatives were mapped onto the index, middle, and ring fingers of the right hand with finger assignment counterbalanced across participants.
Online similarity rating task
As discussed in the introduction, one of the aims of the present study was to determine whether metrics of item-level neural differentiation for scene and object exemplars are differentially impacted by age. To ensure that any potential category effects did not arise merely because of categorical differences in visual similarity between items and their similar lures, we collected similarity ratings for pairs of perceptually similar images from 210 online raters. The similarity rating task was programmed in JavaScript using PsychoJS and hosted on Pavlovia.org. Participants were recruited through Prolific.co, compensated $15/h, and provided informed consent in accordance with the requirements of the Institutional Review Board at the University of Texas at Dallas.
We collected ratings on a total of 86 scene and 86 object perceptually similar triplets. The triplets were presented as three pairs, such that for a given triplet “ABC,” the similar images were presented in three trials “AB,” “BC,” and “AC” randomly interspersed within the stimulus list. The stimulus list included an additional 68 perceptually dissimilar image pairs which served as “catch trials” to ensure that participants were paying attention to the stimuli. As a result, the full stimulus list included a total of 516 lure pairs and 68 catch pairs. The object and scene pairs were presented in a pseudorandomized order such that participants viewed no more than three consecutive trials of the same image category. Given the length of the list, we split it into two lists of 258 lure pairs each while ensuring that all of the similar image pairs belonging to a given triplet were contained within one list. Each participant was administered one list only; thus, each image pair was assessed by 105 raters.
Prior to the similarity judgment task, participants completed an instruction phase and a small number of practice trials to familiarize themselves with the procedures and the rating scale. Each critical trial began with a 500 ms duration red fixation cross followed by a 5 s presentation of the image pair and a continuous “slider” rating scale. During this 5 s, participants used a computer mouse to place a response marker anywhere along the scale (Fig. 2A). For the purposes of the analysis of similarity ratings, the rating scale ranged from 0 (not at all similar) to 10 (very similar). The response marker was re-set to the center position (middle of the scale) at the onset of each trial. The task lasted ∼30 min with a 30 s break provided halfway through.
A, An illustration of a single trial in the online similarity rating task. B, Average across-subject similarity ratings for object and scene lures, demonstrating that pairs of scene lures were on average rated as less similar than object lures. Object and scenes did not differ in their similarities following the similarity matching procedure (see main text).
Figure 2B illustrates across-subject similarity ratings for all object and scene image pairs as well as object and scene catch trials. The similarity ratings for catch trials were robustly lower than those for lure trials, indicating that participants complied with the task instructions. Across all trials, object lure pairs were on average rated as more similar (M = 6.85) than scene lure pairs (M = 6.51; p < 0.001). To ensure that the scene and object lures employed in the fMRI experiment were matched in perceptual similarity, we followed these steps: First, we sorted the lure pairs belonging to each triplet as the “most similar pair,” “middle pair,” and “least similar pair,” according to their across-participant ratings. Next, we selected all pairs from the “most similar” scene group, and all “middle pairs” from the object group, resulting in a total of 86 scene and 86 object pairs. Lastly, we iteratively and randomly selected 60 scene and 60 object pairs until the average across-participant similarity ratings for scenes and objects were equal. The outcome of the matching procedure is illustrated in Figure 2B (see “Matched”). The average across-participant ratings for the final 60 lure pairs employed in the fMRI experiment was 6.74 for both objects and scenes.
MRI data acquisition and preprocessing
Functional and structural MRI data were acquired using a Siemens Prisma 3 T scanner at the Sammons BrainHealth Imaging Center at the University of Texas at Dallas. The data were acquired with a 32-channel head coil. A whole-brain anatomical scan was acquired with a T1-weighted 3D MPRAGE pulse sequence (FOV = 256 × 256 mm; voxel size = 1 × 1 × 1 mm; 160 slices; sagittal acquisition). Functional data were acquired with a T2*-weighted BOLD echoplanar imaging (EPI) sequence with a multiband factor of 3 (flip angle = 70° FOV = 220 × 220 mm; voxel size = 2 × 2 × 2 mm; TR = 1.52 ms; TE = 30 ms; 66 slices). A dual-echo fieldmap sequence which matched the 3D characteristics of the EPI sequence was acquired at TEs of 4.92 and 7.38 ms immediately after the last run of the study phase, resulting in two magnitude images (one per echo), and a pre-subtracted phase image (the difference between the phases acquired at each echo).
The MRI data were preprocessed using Statistical Parametric Mapping (SPM12, Wellcome Department of Cognitive Neurology) and custom MATLAB code (MathWorks). The functional data were preprocessed in six steps. First, we employed the FieldMap toolbox in SPM to calculate voxel displacement maps prior to the fieldmap correction. These maps were calculated using the magnitude and phase difference images acquired in the aforementioned dual-echo fieldmap sequence. Second, SPM’s realign and unwarp procedure was applied, operating in two steps: spatial realignment of the time series registered to the mean EPI image and a dynamic correction of the deformation field using the voxel displacement maps. Third, the functional images were reoriented along the anterior and posterior commissures, then spatially normalized to SPM’s EPI template, and renormalized to an age-unbiased sample-specific EPI template according to procedures standardly employed in our laboratory (de Chastelaine et al., 2016). Lastly, the functional data were smoothed with a 5 mm FWHM Gaussian kernel.
Region of interest selection
The scene-selective parahippocampal place area (PPA) and the object-selective lateral occipital complex (LOC; Fig. 3) were selected as our two a priori regions of interest (ROIs). The LOC and PPA were defined as the intersection between category-selective univariate activity (see below) and anatomical labels provided by the Neuromorphometrics atlas (available in SPM12), following the approach described in Koen et al. (2019; see also Srokova et al., 2020). Prior to ROI selection, the fMRI data were analyzed with a two-stage univariate GLM approach. At the subject level, all study trials were binned into 10 events of interest (five events separately for scene and object trials): (1) trials which went on to be presented in the memory test, (2) first presentation of a to-be-repeated image, (3) first presentation of images which were followed by similar lures, (4) exact repeats, and (5) similar lures. Neural activity elicited by the events of interest was modeled with a boxcar function extending over a 2 s period coincident with image presentation. The boxcar functions were convolved with a canonical hemodynamic response function (HRF) to estimate the predicted BOLD responses. Additional regressors in the design matrix were trials of no interest (filler trials and trials with missing responses or with responses occurring within 500 ms post-stimulus onset), six motion regressors reflecting rigid-body translation and rotation, spike covariates regressing out volumes with displacement >1 mm or 1° in any direction, and the mean signal of each run. Prior to model estimation, the fMRI time series from each scanner run was concatenated into a single session using the spm_fmri_concatenate function.
A priori regions of interest (PPA and LOC) for one representative younger/older adult pair illustrated on a T1-weighted ICBM 152 MNI brain.
At the second level, the parameter estimates from the subject-wise GLMs were entered into a group-level mixed factorial ANOVA with factors of age group (2) and events of interest (10). To ensure that the ROIs were derived independently of the to-be-analyzed data and were unbiased with respect to age group, we employed a leave-one-pair-out approach where group-level GLMs were generated by iteratively leaving out a randomly yoked pair of a younger and an older adult (cf. Hill et al., 2021). As a result, ROIs for each held out pair were derived from the data belonging to the remainder of the sample. The ROIs were defined using category-selective contrasts and comprised all voxels that fell within a 10 mm radius of the cluster’s peak and were included in anatomical masks of the Neuromorphometrics atlas. The PPA was defined using a scene > object group-level contrast inclusively masked by the anatomical labels defining the parahippocampal and fusiform gyri (left PPA, M = 147 voxels; SD = 4.5; right PPA, M = 156 voxels; SD = 1.7). The LOC was delineated with an object > scene contrast, masked by the labels defining the inferior and middle occipital gyri (left LOC, M = 286 voxels; SD = 7.5; right LOC, M = 258 voxels; SD = 6.4).
Multivoxel pattern similarity analyses
For the purposes of the PSAs (Kriegeskorte et al., 2008), the data from the study phase were subjected to a “least-squares-all” GLM (Rissman et al., 2004; Mumford et al., 2014) in which each trial was modeled by a separate 2 s boxcar regressor tracking the image presentation. The six motion regressors reflecting rigid-body translation and rotation were included as covariates of no interest. Item- and category-level PSAs were conducted using approaches similar to those employed in prior work from our laboratory (Koen et al., 2019; Srokova et al., 2020; Hill et al., 2021) and are described in detail below. As noted above, to ensure that category- and item-level similarity metrics were not confounded by within-run autocorrelation effects (Mumford et al., 2014), we computed all correlations between trials belonging to different scanner runs. We also note that the outcomes of all analyses described below were unchanged when visual acuity was employed as a covariate of no interest.
Category-level PSA
Category-level PSA was operationalized for each ROI (PPA/LOC) and image category (scene/object) by computing the difference between within-category and between-category similarity metrics. Within-category similarity was computed as the average voxel-wise Fisher’s z-transformed correlation between a given trial and all trials of the same image category. Between-category similarity was calculated in an analogous fashion—as the average Fisher’s z-transformed correlation between a given trial and all trials of the alternate image category. As illustrated in Figure 4, the primary category-level PSA was performed on novel trials only (i.e., excluding second presentation trials). Restricting the PSA to these trials ensured that age differences in neural selectivity at the category level were not confounded by age differences in repetition effects. However, as described below, a secondary category-level PSA was performed on second presentation trials.
PSA was employed to quantify neural similarity at the level of individual items and stimulus categories. Item-level similarity was computed as the similarity between a given trial and its repeat or lure minus the average similarity with all other repeats or lures belonging to the same image category. Category-level similarity was computed as the difference between within-category similarity (e.g., average correlation between all scene trials with all other scene trials) and between-category similarity (e.g., average correlation between all scenes with all objects).
Item-level PSA
Item-level PSA was computed separately for each ROI, image category, and trial type (i.e., exact repeats and similar lures). The item-level similarity metric was computed as the difference between the within-trial and between-trial similarities. Within-trial similarity was computed as Fisher’s z-transformed correlation between a given item and its second presentation (the item’s exact repeat or its similar lure). Between-trial similarity was computed as the average Fisher’s z-transformed correlation between a given item and the exact repeat (or lure) trials for every other item belonging to the same image category (and belonging to a different scanner run; Fig. 4).
Whole-brain exploratory PSA
ROIs defined using univariate category-selective contrasts have frequently been employed in prior studies which examined age differences in neural differentiation at the category level (Koen et al., 2019). However, item- and category-level PSA effects are unlikely to be restricted to regions which exhibit univariate category selectivity. To address this limitation, and following Hill et al. (2021), we supplemented the ROI-based analyses with an exploratory whole-brain PSA conducted across 384 functionally defined cortical parcels comprising the Atlas of Intrinsic Connectivity of Homotopic Areas (AICHA; Joliot et al., 2015). Note that, unlike the a priori ROIs described above, the AICHA parcels were not constrained by univariate effects. PSA was computed at the item and category levels using the β-parameters extracted from each AICHA parcel, following the approach described in the preceding paragraphs. We focused these analyses on regions exhibiting positive similarity effects (similarity metrics that were significantly greater than zero) that survived an FDR-adjusted significance threshold of q < 0.05.
Behavioral data analysis
Item recognition performance was assessed separately for object and scene stimuli by computing the difference in the proportion of target trials correctly endorsed “old” (item hits) and the proportion of new trials incorrectly endorsed “old” (false alarms). To quantify discriminability between old targets and lure items, we employed a “target–lure discriminability” metric (TLD)—computed as the proportion of old trials correctly endorsed “old” minus the proportion of lure trials which were incorrectly endorsed “old.” This metric differs from the more widely employed “lure discrimination index” (Stark et al., 2013) which instead indexes discriminability between similar lures and new trials. We consider TLD to be the preferable metric because it is a more direct behavioral correlate of pattern separation—the putative hippocampally mediated process that supports the ability to distinguish between similar inputs (Yassa et al., 2011).
Statistical analyses
Statistical analyses were performed in R studio (R Core Team, 2020). Analyses of variance were performed using the afex package (Singmann et al., 2016), with degrees of freedom corrected for nonsphericity with the Greenhouse–Geisser procedure (Greenhouse and Geisser, 1959). The t tests and multiple regressions were performed using the t test and lm functions in base R. Partial correlations were conducted using pcor.test in the ppcor package (Kim, 2015).
Results
Neuropsychological performance
Neuropsychological test performance is illustrated in Table 1. Young adults outperformed older adults on the following measures: CVLT free recall (short delay), Raven’s Progressive Matrices I, category fluency test, SDMT, Trails A, and visual acuity. In contrast, older adults performed better than their younger counterparts on the TOPF.
Study reaction times
A 2 (age group) × 2 (category) ANOVA was performed to examine possible categorical and age differences in reaction times (RTs) during the study phase. The ANOVA revealed a significant main effect of category (F(1, 46) = 69.419; p < 0.001; partial η2 = 0.601), reflective of faster RTs during scene [M (SD) = 1.185 s (0.24)] than object trials [M (SD) = 1.293 s (0.27)]. The effect of age and the interaction between age group and category were not significant (age group, F(1, 46) = 0.130; p = 0.720; partial η2 = 0.003; age × category, F(1, 46) = 0.196; p = 0.660; partial η2 = 0.004), indicating that younger and older adults were equally fast at categorizing scenes and objects as “indoor” versus “outdoor.”
Memory performance
Trial proportions binned according to category, trial type, and response type are illustrated in Table 2. The estimates of item memory performance (hits–false alarms) were entered into a 2 (age) × 2 (image category) mixed-effects ANOVA which revealed a significant effect of category (F(1, 45) = 73.027; p < 0.001; partial η2 = 0.619), reflective of better item memory for objects than scenes. The main effect of category was accompanied by a null effect of age (F(1, 45) = 0.852; p = 0.361; partial η2 = 0.019); the age-by-category interaction was also not significant (F(1, 45) = 2.482; p = 0.122; partial η2 = 0.052). Thus, there were no reliable age differences between young and older adults in item memory performance.
Memory performance in younger and older adults
The TLD metrics were entered into an analogous 2 (age group) × 2 (image category) mixed-effects ANOVA. The ANOVA resulted in a null effect of age group (F(1, 45) = 2.343; p = 0.133; partial η2 = 0.049), while the main effect of category and the age × category interaction were both significant (category, F(1, 45) = 4.741; p = 0.035; partial η2 = 0.095; interaction, F(1, 45) = 4.353; p = 0.043; partial η2 = 0.088). The category effect arose because of better discrimination performance for objects than scenes. The age × category interaction reflected significantly lower discrimination performance in older relative to younger adults for objects (t(44.44) = 2.406; p = 0.020) but a null effect of age for scenes (t(43.10) = 0.509; p = 0.613; Fig. 5).
Item recognition and TLD in younger and older adults. Error bars reflect 95% confidence intervals.
Category-level results
ROI analyses
Similarity indices in the LOC and PPA ROIs revealed reliable category-level effects in younger and older adults for both stimulus categories (Fig. 6A). The category-level similarity indices were entered into a 2 (age group) × 2 (ROI) × 2 (hemisphere) × 2 (image category) mixed-effects ANOVA. The ANOVA revealed a main effect of age group (F(1, 46) = 8.206; p = 0.006; partial η2 = 0.151), an age × ROI interaction (F(1, 46) = 15.412; p < 0.001; partial η2 = 0.251), and an age × category interaction (F(1, 46) = 9.947; p = 0.003; partial η2 = 0.178). Significant effects which did not include the factor of age group, and thus are not discussed further, were a significant main effect of ROI (F(1, 46) = 106.438; p < 0.001; partial η2 = 0.698), a two-way interaction between ROI and category (F(1, 46) = 139.133; p < 0.001; partial η2 = 0.752), and a tree-way interaction between ROI, hemisphere, and category (F(1, 46) = 9.196; p = 0.004; partial η2 = 0.167). No other effects were significant (ps > 0.092). Follow-up analyses of the age × ROI interaction revealed significantly lower category-level similarity in older adults in the PPA (t(43.60) = 3.490; p = 0.001), but no age differences in the LOC (t(45.99) = 0.269; p = 0.789). The age × category interaction was reflective of significant age differences for scenes (t(42.14) = 4.313; p < 0.001), but null effects of age for objects (t(43.60) = 0.416; p = 0.680). Therefore, as is evident in Figure 6A, age-related neural dedifferentiation at the category level was observed only in the PPA and only during scene trials.
A, Category-level similarity indices in the LOC and PPA ROIs for novel trials only, demonstrating category-level age-related neural dedifferentiation for scenes in the PPA. B, Complementary category-level analyses performed on second presentation trials, demonstrating that older adults exhibited lower selectivity in the PPA but not in the LOC. Error bars in panels A and B reflect 95% confidence intervals. C, Category-level similarity effects for objects and scenes across age groups. D, Regions exhibiting reliable age differences in neural similarity for scenes and objects (red, greater similarity for younger adults; blue, greater similarity for older adults).
In a secondary analysis, we examined whether the key findings reported above are replicated when the category-level analyses were restricted to second presentation trials only (repeats and lures; Fig. 6B). Given the considerably lower trial numbers available for these analyses relative to the analysis of the novel trials, the selectivity metrics were averaged across repeats and lures while focusing on the regions “preferred” stimulus category. A 2 (age group) × 2 (ROI) mixed-effects ANOVA revealed a significant main effect of age group (F(1, 46) = 5.082; p = 0.029; partial η2 = 0.099), a significant main effect of ROI (F(1, 46) = 10.310; p = 0.002; partial η2 = 0.183), and an age × ROI interaction (F(1, 46) = 4.771; p = 0.034; partial η2 = 0.094). Follow-up analysis of the interaction confirmed that age-related neural dedifferentiation at the category-level was present in the PPA (t(43.157) = 2.742; p = 0.009), but not in the LOC (t(43.157) = 0.345; p = 0.732).
Exploratory whole-brain analyses
Figure 6C depicts those AICHA parcels which exhibited reliable positive scene and object main effects across the whole sample. Scene-selective effects were primarily evident across the occipital and posterior temporal cortex, the parahippocampal and fusiform gyri, and the retrosplenial complex. By contrast, object-selective effects were more widely distributed and were prominent in frontal, parietal and anterior temporal cortex. Regions demonstrating reliable age differences in category-level similarity are depicted in Figure 6D. For scene trials, older adults demonstrated lower category-level similarity in the occipital cortex, as well as in many of the regions typically implicated in scene processing, such as the parahippocampal cortex and the retrosplenial complex. For object trials, age effects were relatively sparse, with younger adults showing greater similarity than older adults in two clusters in the anterior fusiform gyrus and in the medial prefrontal cortex. However, greater similarity for objects in older adults was evident in a number of frontal regions, along with the posterior hippocampus and posterior cingulate cortex. Therefore, consistent with the ROI analyses, age-related neural dedifferentiation at the category level was most reliably observed during scene processing in scene-selective cortical regions.
Item-level results
ROI analyses
Analogously to the analyses of category-level PSA, item-level similarity metrics were subjected to a 2 (age group) × 2 (ROI) × 2 (hemisphere) × 2 (image category) × 2 (trial type, repeat/lure) mixed-effects ANOVA. The data are illustrated in Figure 7A. The ANOVA revealed a significant main effect of age group (F(1, 46) = 20.915; p < 0.001; partial η2 = 0.313), which was driven by greater similarity metrics in younger relative to older adults. The age × category interaction was statistically equivocal (F(1, 46) = 4.043; p = 0.050; partial η2 = 0.081). Follow-up pairwise comparisons indicated that older adults exhibited lower similarity for both objects (t(45.68) = 4.053; p < 0.001) and scenes (t(45.35) = 2.273; p = 0.028), although the age effect was seemingly greater for objects. However, given that the category × age interaction only approached significance, we urge caution in interpreting the differences in the magnitude of the age effects for the two classes of category exemplars. We also note that any potential category × age interaction is confounded by differential scene-related “baseline” (between-item) similarity in younger and older adults. Because older adults exhibited relatively lower similarity across all scene stimuli (see category-level results), age differences in scene selectivity might have been harder to detect at the item-level. Lastly, effects which did not interact with the factor of age group, and thus are not discussed further, included a main effect of ROI (F(1, 46) = 21.647; p < 0.001; partial η2 = 0.320), a significant two-way interaction between ROI and category (F(1, 46) = 4.890; p = 0.032; partial η2 = 0.096), and a four-way interaction between ROI, hemisphere, category, and trial type (F(1, 46) = 9.052; p = 0.004; partial η2 = 0.164). No other effects were significant (ps > 0.071). In summary, item-level analyses revealed that, in contrast to the category-level PSA, age-related neural dedifferentiation was evident for both scene and object stimuli. Moreover, neural similarity was lower in older adults regardless of trial type (repeat vs lure). Indeed, except for the four-way interaction, effects of trial type were uniformly absent.
A, Item-level similarity indices in the LOC and PPA ROIs, revealing reduced similarity in older adults for objects and scenes across both repeats and lures. Error bars reflect 95% confidence intervals. B, Reliable item-level main effects for objects and scenes across both age groups.
Exploratory whole-brain analyses
Reliable across-participant item-level effects were observed across much of the occipital cortex (Fig. 7B). Whereas effects for scene lures were largely restricted to these occipital regions, effects for scene repeats and for objects (repeats and lures alike) extended into parietal, frontal, and posterior temporal areas (e.g., parahippocampal and fusiform gyri). The outcomes of age group contrasts were sparse and not particularly informative. Age differences were evident in only a few AICHA parcels, all of which were scattered across the cortex with little clustering across neighboring parcels. These few parcels exhibited small effect sizes which were not systematically greater in one age group relative to the other (the four age contrasts are available from the first author upon request). One exception was evidence for greater item-level similarity in younger adults for object repeats and lures in small parcels within the posterior parahippocampal and lateral occipital cortices bilaterally (echoing Fig. 7A). Thus, the whole-brain item-level analyses essentially mirrored the outcomes of the ROI analysis.
Relationship between neural differentiation and memory performance
Motivated by prior reports of a positive, age-invariant relationship between neural differentiation in the PPA and memory performance (Koen et al., 2019; Srokova et al., 2020), we performed multiple regression analyses to examine the relationship between TLD and neural similarity. The TLD scores were highly correlated between objects and scenes (rpartial = 0.670; p < 0.001; controlling for age group). Therefore, to reduce the number of multiple comparisons, we averaged the scores across categories to generate a single mean TLD metric. The analysis approach consisted of two steps. First, a regression model was defined using TLD as the dependent variable and the variables of age group, within-between similarity, and their interaction term as the predictors. A statistically significant age group × similarity interaction would indicate that the relationship between TLD and similarity is moderated by age group, motivating follow-up analyses in the form of the computation of zero-order correlations separately for each age group. A nonsignificant interaction term would indicate that any relationship between the neural and behavioral variables was age-invariant. In these cases, the relationship was quantified by the partial correlation between neural similarity and TLD, controlling for age group. We performed a total of four regression analyses, separately for object-related similarity in the LOC and scene-related similarity in the PPA, operationalized either at the item or category level.
Turning first to category-level metrics of scene stimuli, the age group × scene similarity interaction in the PPA was not significant (p = 0.166), indicating that any potential relationship between scene-related similarity and TLD was not moderated by age group. A partial correlation (controlling for age group) between scene-related similarity in the PPA and TLD was statistically significant (rpartial = 0.390; p = 0.007; Fig. 8), revealing evidence for an age-invariant relationship (when the TLD was examined separately for the two stimulus categories, the TLD for scenes was reliably correlated with scene selectivity, rpartial = 0.418; p = 0.004; and the analogous correlation for object TLD approached significance, rpartial = 0.286; p = 0.054; the two correlations did not significantly differ from each other). Turning to category-level differentiation for objects, the interaction between age group and object similarity in the LOC was again nonsignificant (p = 0.393), as was the partial correlation between LOC similarity and TLD (rpartial = 0.277; p = 0.062). In conclusion, our analyses demonstrate that the relationship between TLD and category-level similarity was restricted to scene-related similarity in the PPA. The regression and correlation analyses examining the association between TLD and neural similarity at the item-level (either for repeats or lures) did not yield any statistically significant relationships.
Age-invariant relationship between mean TLD and scene-related category-level neural differentiation in the PPA. The relationship remains significant (p = 0.035) following exclusion of the highlighted outlier.
Lastly, we performed partial correlations (controlling for age) to examine the relationships between the category-level and item-level metrics of neural differentiation (for the item-level metrics, we collapsed similarity across exact repeats and lures). Category- and item-level metrics of both scene and object differentiation were uncorrelated in either ROI (p min = 0.148).
Discussion
In the present study, we examined the effects of age on object- and scene-related neural selectivity at the levels of image category and individual stimulus exemplars. Consistent with prior reports (Koen et al., 2019), multivoxel PSAs at the category level identified lower selectivity in the older relative to the younger age group in scene-selective, but not object-selective, cortical regions. In contrast, item-level analyses revealed reduced neural selectivity in older adults for exemplars of both stimulus categories and regardless of trial type (repeat vs similar lure). Also consistent with prior findings, regression analyses revealed that greater category-level scene selectivity in the PPA was predictive of higher target–lure discrimination regardless of age group or image category.
The present manuscript describes the first direct comparison of two separate metrics of neural selectivity derived from the same samples of younger and older adults. As noted in the introduction, most prior work has focused on examining age-related reductions in neural selectivity at the level of stimulus categories. In contrast, studies examining neural selectivity at the level of individual stimulus exemplars are less common, and the evidence for age differences is inconsistent. Moreover, theories of age-related neural dedifferentiation have largely assumed that category- and item-level selectivity metrics reflect a common neural mechanism driven by a general age-related decline in neural selectivity. The present findings challenge this assumption, providing novel evidence that null age effects in category-level selectivity do not necessarily extend to item-level selectivity metrics. Thus, the findings suggest that category- and item-level measures reflect distinct neural mechanisms, providing a foundation for further studies aimed at elucidating their mechanistic underpinnings.
Behavioral results
Participants completed the retrieval phase of the MST outside of the scanner immediately following the scanned encoding phase. We found no age differences in item recognition, consistent with numerous prior findings of relatively preserved familiarity-based recognition memory in older age (for review, see Koen and Yonelinas, 2014). In contrast, object, but not scene, TLD was lower in older adults, indicative of an age-related decline in the ability to discriminate between previously viewed images of objects and their similar lures. These findings are consistent with prior reports, which indicate that age differences in lure discriminability are frequently observed in MST variants utilizing object stimuli (Stark et al., 2019), but are less common or weaker in variants of the task employing spatial or scene stimuli (Reagh et al., 2016; Stark and Stark, 2017; Berron et al., 2018; note that, in most prior studies, lure discriminability was measured using an alternate metric to that adopted here). The object-specific decline in putative behavioral metrics of pattern separation has been attributed to age-related decline in the functional integrity of perirhinal and lateral entorhinal cortex, in recognition of the fact that these regions have been implicated in object processing and, additionally, are a major source of object information flowing into the hippocampus (Reagh et al., 2016). However, the notion that scene-specific input to the hippocampus is relatively preserved with increasing age is difficult to reconcile with the present and prior findings of robust neural dedifferentiation in scene-selective cortical regions, including the posterior parahippocampal cortex (which forms part of the PPA). We conjecture that the age differences in TLD for objects, but not scenes, are reflective of greater conceptual confusability between similar objects. Relative to scenes, object stimuli may contain less perceptual information that can support the discrimination between conceptually identical but physically similar items. Given prior evidence for an age-related decrease in mnemonic precision (Nilakantan et al., 2018; Korkki et al., 2020), and hence a tendency to retrieve gist-based (Koutstaal and Schacter, 1997) or conceptual information (Deng et al., 2021; Srokova et al., 2022), older adults may have greater difficulty discriminating between studied objects and their lures because of a lack of diagnostic perceptual information.
Effects of age on category-level neural differentiation
Whereas age-related neural dedifferentiation has consistently been reported for scene images in scene-selective cortical regions, dedifferentiation for objects, words, or faces in their respective category-selective regions has been identified much less consistently (for review, see Koen and Rugg, 2019; Koen et al., 2020). The mechanisms responsible for category-level neural dedifferentiation and the reasons for its seeming regional specificity are the subjects of debate. One candidate mechanism, proposed by Li et al. (2001), posits that age-related neural dedifferentiation reflects a decline with age in the integrity of the ascending dopaminergic neuromodulatory system. From this perspective, reductions in selectivity are caused by reduced dopaminergic availability, which compromises neural signal-to-noise ratio, and hence the fidelity of neural representations. Somewhat analogously, neural dedifferentiation has also been hypothesized to reflect age-related decline in γ-aminobutyric acid (GABA) inhibitory neurotransmission (Lalwani et al. 2019; Cassady et al., 2019, 2020; Chamberlain et al., 2021) and a consequent broadening of neural tuning. However, the current and prior evidence for regional specificity of neural dedifferentiation (at least at the category level) arguably challenges the notion that the phenomenon is attributable such factors, at least if they are considered to operate cortex-wide.
A possible explanation for the ubiquity of category-level age-related dedifferentiation of scene stimuli stems from findings that the magnitude of fMRI BOLD responses in scene-selective cortical areas, such as the PPA, are strongly modulated by perceptual complexity and attentional factors (Aminoff et al., 2013). As we have proposed previously (Srokova et al., 2020), age-related decline in neural selectivity for scenes might be a consequence of decline in the ability to process complex visual inputs and, especially, difficulty in differentiating the multiple, spatially distributed elements that constitute a scene. Alternately, or additionally, scene dedifferentiation might be related to reduced availability of domain-general attentional resources (cf. Serences et al., 2004; Gazzaley et al., 2005), such that older adults fail to exert sufficient attentional control over scene processing (Bouhassoun et al., 2022). These potential mechanisms assume that the null effects of age for faces and object selectivity that have been reported reflect their relatively lower visual complexity and lower attentional demands. As we discuss below, on its face, this account is compromised by the present finding that age effects on item-level selectivity extend beyond scenes to include objects.
Effects of age on item-level neural differentiation
As noted in the Introduction, it might be assumed that age-related category-level neural dedifferentiation is driven by a decline in the selectivity of neural patterns elicited by individual category exemplars. However, no prior study has directly examined whether age differences in category-level differentiation are associated with age differences at the item-level. As outlined in the introduction, Goh et al. (2010) reported that while younger adults exhibited robust univariate repetition suppression effects for exact repeats of faces, older adults exhibited repetition effects for both repeats and perceptually similar lures, a finding indicative of item-level dedifferentiation. Additionally, in two studies that employed MVPA, it was reported that pattern similarity for exact repetitions was lower in older adults (Bowman et al., 2019; Trelle et al., 2019). By contrast, two other studies reported null effects of age on the neural similarity between successive stimulus presentations (Zheng et al., 2018; St-Laurent and Buchsbaum, 2019). The present findings are consistent with those prior studies that reported a decline with age in item-level neural selectivity.
Of importance, the present study offers novel evidence opposing the idea that category- and item-level selectivity depend on the same neural mechanisms. We demonstrate that absent age effects on object selectivity at the category level do not extend to neural selectivity at the level of individual objects, at least as this is indexed by multivoxel PSA. Rather, age-related dedifferentiation at the item level was evident for both scenes and objects.
In attempting to understand the implications of this dissociation between category- and item-level metrics, we note that metrics of item-level selectivity quantify the extent to which a given item elicits neural patterns that are distinct from those elicited by other items belonging to the same category (i.e., item-specific information). In contrast, category-level selectivity indexes the extent to which neural patterns are shared between category exemplars, potentially reflecting processes that are engaged by all (or most) members of a category. There is no necessary reason why these two expressions of neural selectivity should be related to one another. Indeed, consistent with this point, and regardless of age group or stimulus category, item- and category-level similarity metrics were uncorrelated across participants in the present study.
Lastly, we note that in both age groups, we were unable to identify any effect on item-level selectivity of the “repeat” versus “similar lure” manipulation. We have no ready explanation for this null finding, not least since our participants demonstrated the ability to discriminate between repeats and lures on the post-scan memory test; our expectation was that, at the least, we would find evidence of lower similarity between items and their lures in the younger age group (cf. Goh et al., 2010). One possibility is that the failure to detect differences in neural selectivity between the two classes of items reflects a mismatch between the spatial resolution at which these differences were expressed in the cortex and the resolution of our imaging method.
Relationship between category-level neural differentiation and memory performance
Category-level neural selectivity has consistently been reported to correlate positively across participants with memory performance in an age-invariant manner (Koen et al., 2019; Srokova et al., 2020). Based on the limited available data (the majority of prior studies where a relationship between neural selectivity and memory performance was reported collapsed across different category-selective regions and exemplars), this relationship appears to be selective for metrics of scene selectivity derived from the PPA. The present findings are fully consistent with these prior results and extend them to a memory metric (TLD) not hitherto examined in this context. Why PPA selectivity, but seemingly not selectivity metrics derived from other regions, should be sensitive to memory performance is unclear (for further discussion, see Srokova et al., 2022).
Conclusions
In summary, our data replicate and extend prior studies of age-related dedifferentiation. We replicated earlier findings indicating that, at the category level, evidence for neural dedifferentiation is limited to scene-selective cortical regions. Crucially, however, the present findings indicate that this dissociation does not extend to neural differentiation at the item level, when age-related dedifferentiation was evident for both scene and object images. Thus, item- and category-level metrics of neural differentiation are differentially sensitive to increasing age and likely reflect distinct neural mechanisms.
Footnotes
This work was supported by the National Institute on Aging Grants R56AG068149 and RF1AG039103 and an award from BvB Dallas. The authors would like to acknowledge Joshua Olivier, Nehal Shahanawaz, and Eduardo Hernandez for their assistance with recruitment and neuropsychological assessments.
The authors declare no competing financial interests.
S.S.’s present address: Department of Psychology, University of Arizona, Tucson, Arizona 85721
- Correspondence should be addressed to Sabina Srokova at sabinasrokova{at}arizona.edu.