Age-related decrease in inter-subject similarity of cortical morphology and task and resting-state functional connectivity

Abstract

“All old people are the same” is an unfortunate characterization of the perceived homogeneity in the older age group. This study attempts to debunk this myth in the context of the structural and functional brain. Within older relative to younger age groups, individuals are hypothesized to be more dissimilar to their similar-aged peers—thus demonstrating an age-related divergence. This study analyzed functional connectivity (FC) during multiple fMRI paradigms (2 rest + 5 tasks) and cortical thickness (CT) data from two lifespan datasets (N_total = 1161). On average, between-subject FC/CT correlations became weaker in the older age groups. Further analyses ruled out the possibility that more rapid age-related changes in older brains have increased the dissimilarity in these older age groups. Brain-wide analyses revealed significant effects of age-related divergence across most of the brain. Finally, CT similarity between a dyad significantly predicted their FC similarity across multiple fMRI task paradigms—demonstrating a close relationship between brain structure and function even at the between-dyad level. Contrary to the myth that “all old people are the same,” these findings suggest young people are more similar to each other. This study presents major implications in the study of neural fingerprinting and brain-behavior associations.

Supplementary Information

The online version contains supplementary material available at 10.1007/s11357-023-01008-9.

Introduction

The outgroup homogeneity effect posits that we perceive people within our social grouping as being different from each other while those outside the group are similar to each other. This phenomenon, along with unfortunate ageist stereotypes that attempt to paint the older adult population with a broad brush stroke, has possibly led to the perception among younger age groups that “all old people are the same” [1–3]. In the physical abilities domain, it may be the case that there is indeed less variability perceived among older people than the young [3] as a result of the age-related decline in physical agility. This myth is prevalent in the behavioral and cognitive contexts as well. Some of the ageist stereotypes prey on certain impaired or atypical behavioral and cognitive characteristics in the old [4, 5], giving the impression that the older age groups are homogeneous in these areas. By extension, such behavioral and cognitive characteristics allude to the brain. To this end, the current research examines the myth that “all old people are the same” in the context of the brain.

Variability in the structural and functional aspects of the brain arises from multiple factors—some of which are modulated by age. As a result, our brains may become increasingly different (i.e. age-related divergence) or similar (i.e., age-related convergence) to our same-aged peers. Structurally, our brains are highly susceptible to experience-dependent neuroplasticity. As we undergo vastly different life experiences from our peers, these differences will be expressed in the brain as variability in cortical morphology [6]. It is conceivable that as we age, these differences in life experiences accumulate and are cumulatively encoded in the cortical structure, resulting in an age-related divergence in cortical morphology.

The variability in brain activation patterns observed during a task was suggested to be partially accounted for by the variability in brain structure [7]. Thus, the presence of age-related divergence in brain structure would correspondingly predict an age-related divergence in functional activation and consequently functional connectivity (FC) patterns. Additionally, variability in the functional activation patterns can also be explained by the fact that there are multiple ways to perform the task; as the number of possible strategies to complete the task increases, so will the variability of its task-related activation and connectivity patterns. The chosen task execution approach is subjected to one’s prior experience, practice, and proficiency, which are encoded in the brain structure as experience-dependent changes [8, 9], thus explaining the link between the experience-dependent structural variability and variability in functional activity and connectivity. Furthermore, as a result of age-related neurodegeneration, one would be inclined to perform the task in a manner that bypasses activation in the degenerated region and instead recruits compensatory activity in auxiliary regions [10]. The nature of such functional compensation differs qualitatively and quantitatively across individuals, given that age-related neurodegeneration is a highly heterogeneous process, attributing to a myriad of different genetic, epigenetic, and local contextual factors [11]. This heterogeneous neurodegeneration would thus add to the age-related divergence in functional activation and connectivity.

It was previously shown that the trajectories of changes in cortical thickness in several regions [12, 13] do not occur linearly and uniformly across the lifespan. For instance, cortical thinning in the occipital and temporal cortices occurred more rapidly with increasing age [12]. For this reason, a 1-year age difference might be associated with greater structural changes in the brain and by extension changes in functional activity as well, in the old than in the young, thereby creating the impression that there is significantly greater variability in the older age groups than in the young.

On the other hand, some common study-related factors may inadvertently constrain the heterogeneity in the older age groups. For instance, as the prevalence of dementia increases with age [14], the proportion of cognitively healthy participants would correspondingly decrease. Thus, the same cognitively normal inclusion criteria may become more selective and homogenizing to the older age groups than the young. In a similar vein, head motion is significantly higher in the old [15], and MRI contraindications (i.e., pacemakers) [16] are more prevalent in the old; the exclusion of many older subjects due to these reasons may further homogenize the older adult sample, inadvertently leading to an age-related convergence.

The current study set out to study the age-related divergence in the structural and functional brain. Using cortical thickness (CT), resting-state, and task-based FC data from two lifespan datasets, the age-related divergence hypothesis was tested—brains are hypothesized to be less similar to each other in the older than in the younger age groups. Then, the possibility that this age-related divergence can be accounted for by the increased rate of change in older brains was investigated. Next, the specific regions and connectomes that exhibited significant age-related divergence/convergence effects were identified via brain-wide analyses. Finally, given the previously theorized link between structural variability and FC variability, as a secondary goal, it was examined if CT similarity between a dyad can predict their FC similarity.

Methods

Participants

The present study utilized data from two publicly accessible datasets. The first was obtained from the HCP-Aging Release 2.0 of the Lifespan Human Connectome Project in Aging. This dataset included 725 participants. Information regarding the recruitment procedures and exclusion criteria has been described in detail elsewhere [17, 18]. Among them, 95 with missing or invalid scans, such as those with excessive head motion in one or more fMRI paradigms, were excluded from the study. Six additional subjects who were aged 100 years and above were excluded because there was a significant age gap between these participants and the next younger participant (90 years); the inclusion of these centenarians would otherwise distort the partitioning of the age groups. As a result, the final sample comprised 624 participants (344 females), with a mean age of 59.2 (SD = 14.9) and an average of 17.6 years of education (SD = 2.2). Ethical approval for this study was obtained by the HCP coordinating center.

The second dataset was obtained from the CAM-CAN repository (available at http://www.mrc-cbu.cam.ac.uk/datasets/camcan/). Specifically, data acquired from stage 2 of the CAM-CAN study was used. Information regarding the recruitment procedures and exclusion criteria has been described in detail elsewhere (Shafto et al., 2014; Taylor et al., 2017). After excluding 117 subjects with missing data or invalid scans in at least one of three fMRI paradigms, the final sample consisted of 539 subjects (274 females), with a mean age of 52.2 (SD = 18.3). The CAM-CAN study has been approved by the local ethics committee, Cambridgeshire 2 Research Ethics Committee (reference: 10/H0308/50).

In-scanner tasks

Both the HCP-Aging and CAM-CAN datasets contain resting-state (REST) and task-based fMRI scans. Three task-based scans—VISMOTOR, FACENAME, and CARIT—were included in the HCP-Aging dataset.

The VISMOTOR task [19] required participants to view black and white circular checkerboards with red flickering square targets, presented in pairs on either side of a central fixation point. The participants were instructed to quickly press a left or right button in response to the red squares appearing on the left or right, respectively. This task was designed to engage the visual and motor networks.

The CARIT task [20] was a Go/NoGo task designed to evaluate inhibitory control. Participants were asked to rapidly press a button in response to seeing all shapes except for two target shapes.

The FACENAME task [21] was designed to recruit encoding and retrieval-related neural correlates. Participants were instructed to memorize the names of a series of faces during the “encoding” blocks and attempt to recall their names in separate “recall” blocks. The encoding and recall blocks were intermixed with a distractor task (20 s of Go/NoGo). During the encoding blocks, participants were instructed to press a left button when they saw each face/name pair appear on the screen. During the recall blocks, the same faces would appear one at a time without their accompanying names, and participants were instructed to press the left button whenever they believed they could correctly recall the name of the presented face. The face stimuli were presented in the same order in the encoding and recall blocks.

Two task-based scans—MOVIE and sensorimotor (SMT)—were included in the CAM-CAN study. During the MOVIE task, the participants watched a portion of an old film, Alfred Hitchcock’s “Bang! You’re Dead,” which was a black-and-white television drama. The film’s original duration was 30 min, but it was edited down to 8 min while keeping the main plot intact [22].

In the SMT task, participants pressed a button with their index finger if they detected the simultaneous appearance (duration = 34 ms) of two circular checkerboards on both sides of a central fixation cross, and/or a binaural tone (300, 600, or 1200 Hz, duration = 300 ms). The visual and auditory stimuli were presented either simultaneously in 120 trials, or separately, in eight trials (four visual-only, and four auditory-only). The inter-trial interval varied between 2 and 26 s.

MRI acquisition

HCP-Aging participants were scanned using a Siemens 3 Tesla Prisma system with a 32-channel head coil. T1-weighted images were acquired using an MPRAGE protocol (TR = 2500 ms; TE = 1.8 ms; flip angle = 8, 208 sagittal slices; voxel size = 0.8 mm isotropic). The fMRI volumes for all 4 fMRI paradigms were acquired using a 2D multiband gradient-recalled echo echo-planar imaging (EPI) sequence (TR = 800 ms; TE = 37 ms; flip angle = 52°, 72 axial slices; voxel size = 2.0 mm isotropic). Two REST runs with opposite phase-encoding directions (i.e., anterior–posterior and posterior-anterior), each lasting 6 min 30 s, were carried out for each participant and were concatenated into a single fMRI volume for the analyses. The CARIT (4 min 11 s), VISMOTOR (2 min 46 s), and FACENAME (4 min 47 s) tasks were acquired in the posterior-anterior phase-encoding direction. Additionally, opposite phase-encoding spin-echo field map pairs were acquired separately for all tasks (TR = 400 ms; TE1 = 5.19 ms; TE2 = 7.65 ms) to correct for signal distortions.

CAM-CAN participants were scanned using a 3 T Siemens TIM Trio scanner equipped with a 32-channel head coil. T1-weighted images were acquired using an MPRAGE protocol (TR = 2250 ms; TE = 2.99 ms; FOV = 256 × 240 × 192 mm; voxel size = 1 mm isotropic). EPI (for REST and SMT) and multi-echo EPI (for MOVIE) volumes (TE = 30 ms; 32 axial slices; FOV = 192 × 192; voxel size = 3 × 3 × 4.44 mm) were acquired during the fMRI scans. A total of 261 volumes were acquired each for REST and SMT and 193 volumes for MOVIE. The TR for REST and SMT were the same (1970 ms), while that of MOVIE was longer (2470 s). Additionally, gradient echo field maps (TR = 400 ms; TE1 = 5.19 ms; TE2 = 7.65 ms) were acquired for correcting the fMRI distortions.

Image processing

T1-weighted structural scans were preprocessed using Freesurfer 7.2.0 in the CAM-CAN and Freesurfer 6.0.0 in the HCP-Aging datasets. The latter was preprocessed by the HCP team and the preprocessed files from the “PreprocStrucFreesurfer” directory in the HCP-Aging Release 2.0 were downloaded for use in the current study. The preprocessed images in both datasets were subsequently resampled onto the fsaverage5 space. Since smoothing would inflate the between-subject similarity, smoothing was not carried out, unless otherwise specified.

For the HCP-Aging sample, the previously preprocessed [23, 24] fMRI scans, in particular the *Atlas_MSMAll_hp0_clean.dtseries.nii files from the HCP-Aging Release 2.0, were downloaded. Briefly, these preprocessed fMRI volumes had undergone the GenericfMRIVolumeProcessingPipeline, GenericfMRISurfaceProcessingPipeline, hcp_fix_multi_run, and MSMAllPipeline steps. These volumes were aligned across subjects using multi-modal surface registration. Subjects with excessive head motion were identified as those with a root mean squared displacements greater than 0.25 were excluded. Using the R package “ciftiTools” [25], the preprocessed time series data were parcellated into 100 cortical nodes using the Schaefer-100 atlas [26] and 19 subcortical nodes using the Freesurfer subcortical segmentations [27]. Then, for each subject, a 119 × 119 FC matrix was generated from the bivariate Fisher-Z transformed correlations between nodes in the parcellated time series.

Resting-state and task-based scans in the CAM-CAN dataset were preprocessed using fMRIPrep 22.0.2 [28]. Functional data were slice time corrected using 3dTshift from AFNI [29] and motion-corrected using MCFLIRT [30]. This process was followed by co-registration to the corresponding T1w using boundary-based registration [31] with 9 degrees of freedom, using bbregister from freesurfer. Motion correcting transformations, BOLD-to-T1w transformation, and T1w-to-template (MNI) warp were concatenated and applied in a single step using antsApplyTransforms employing Lanczos interpolation. Subsequently, these preprocessed volumes were denoised by regressing out six motion parameters, the average signal of white matter and cerebrospinal fluid masks, global signal, and their derivatives, as well as cosines covering slow time drift frequency band using the “nilearn” library in Python. The volumes were subjected to a 0.1-Hz low-pass filter. Just like the HCP-Aging data, the same excessive head motion criterion was used to exclude subjects, and the same 119-parcel atlas was used to construct the FC matrices.

Statistical analysis

Sliding window approach

A sliding window approach was used to partition the age groups. The first age group included participants within the age range of youngest age to (youngest age + 10 years). Then, the upper and lower age limits in each successive group were moved upwards by 5 years from the previous group. Using this approach, the HCP and CAMCAN datasets were partitioned into 10 and 13 age groups, respectively (see Fig. 1).

Fig. 1.

Distribution of age group partitions

Between subject FC/CT similarity

Next, the CT values at each vertex in the fsaverage5 space and the edge values in the top half of the FC matrices were vectorized. The empty columns in the CT vector (i.e., vertices where their values were 0 for all participants) were removed. Each subject in the HCP-Aging study thus had 5 vectors (i.e., CT + REST + 3 fMRI tasks) and each subject in the CAM-CAN dataset had 4 vectors (i.e., CT + REST + 2 fMRI tasks). Then, within each age group, for every unique pair of subjects, the similarity for each vector between both subjects in the dyad [32, 33] was calculated using Pearson’s correlation coefficients.

Whole-brain age-related divergence/convergence

Then, to investigate the connectomes and cortical regions that exhibited patterns of age-related divergence/convergence, the standard deviation (SD) at each edge and vertex, was calculated across participants in their respective age groups. Each age group thus has its own CT SD maps (shown in the supplementary materials) and FC SD matrices. In relation to the latter, the associations between these edge-wise SDs with the rank of the age group (youngest to oldest) were examined using network-based statistics (NBS) as implemented in the R package “NBR” [34]. Edges were selected if they were significant at the p < 0.001 level in the edge-level univariate analyses. Following which, these selected edges were thresholded at the network level. That is, edges, each representing the coefficient of the relationship between a functional connection and the rank of the age group, and whole-brain age-related divergence/convergence were summed within their respective networks to compute the network strengths. Then, these network strengths were tested, at the p < 0.05 level, against a null distribution of network strengths generated via 1000 nonparametric permutations; networks with strengths larger than 950 (i.e., p < 0.05) of these permuted network strengths are considered to be statistically significant. For the CT modality, the associations between the vertex-wise SDs and age groups were examined using the “standard linear regression” module in the “BrainStat” Python library [35]. The random field theory (RFT) correction, which corrects the probability of ever reporting a false positive result, was applied with a cluster-defining threshold of p < 0.05. RFT, unlike the classical multiple comparison correction approaches like the Bonferroni correction, is not susceptible to the spatial correlation of the data [36]. Only for this analysis, the CT images were smoothed (10 mm kernel) prior to the analysis since cluster thresholding procedures were carried out.

Hierarchical linear model analyses

To address the possibility that the age-related divergence may be linked to the increased rate of change in older brains, the influence of between-dyad age differences on the dyad’s FC/CT similarity in each age group was investigated. Specifically, a hierarchical linear model (HLM) was fitted within each age group to predict FC similarity with a random intercept and the fixed effects of sex difference, age difference, and its quadratic term.

To investigate the relationship between CT similarity and FC similarity, two other HLMs were fitted to predict FC similarity using CT similarity. The first included a random intercept and the fixed effect of CT similarity. The second (adjusted) model was similar to the first but additionally included sex difference, age difference, and its quadratic term as fixed effects to control for their influences on FC similarity. These HLM analyses were carried out as implemented in the R package “lme4” [37].

All analyses were carried out in R 4.2.2 and Python 3.10. Statistical significance was set at p < 0.05. The code for running these analyses and generating their related visualizations is available at https://osf.io/mf6ru/.

Results

Between dyad similarity across age groups

The distributions of FC/CT similarity for each age group were plotted out in Fig. 2 to illustrate their age-related trends. The means of these distributions were fitted to a regression slope. These regression slopes were all statistically significant (− 0.99 ≤ β ≤  − 0.82; ps ≤ 0.004). Some key findings were noted. First, there was a gentle decline in the mean between-dyad FC/CT correlations across age groups, suggesting that FC/CT profiles were more dissimilar between dyads in the older age groups. This decline was the steepest for the CARIT task in the HCP-Aging dataset and the MOVIE task in the CAM-CAN dataset.

Fig. 2.

Distribution of between-dyad FC/CT correlations across age groups in both datasets. The horizontal blue lines represent the means for each distribution, and the red lines represent regression lines fitted across these means

Does a 1-year age difference predict greater brain differences in the old than in the young?

As shown in Fig. 3a, the HLM analyses revealed the between-dyad age difference to be a significant and negative predictor of their FC/CT similarity. The earlier results showed that the between-dyad FC/CT similarity decreased in the older age groups. Given that each of the age groups spans across 10 years, the between-dyad FC/CT variability within the age groups may be at least partially attributable to the between-dyad age difference, which could be as large as 10 years. This age difference could be associated with a larger FC/CT difference in the older age groups than in the young due to the increased rate of changes in older brains, thus inadvertently giving the impression of an age-related divergence.

Fig. 3.

Predictors of between-dyad FC/CT correlations in the a full sample and the b overlapping age groups. The error bars represent 95% confidence intervals. All coefficients are statistically significant in a

To investigate and rule out this possibility, the same HLM analyses were carried out within each of the overlapping age groups. If it were the case that a 1-year age difference meant greater differences in the brains among the older age groups than the young, one would expect the coefficient for the age difference and its quadratic term to become more negative in the older age groups. According to the results (see Fig. 3b), while it was observed that this coefficient did become more negative across two consecutive age groups (“71 to 81” and “76 to 86” in the HCP Aging), there was not a consistent pattern where these coefficients become increasingly negative across the lifespan. Furthermore, these age coefficients were positive in some of these older age groups, suggesting that larger between-dyad age differences predicted more similar FC/CT. Overall, these results do not suggest that the age-related divergence can be explained by the possibility that a 1-year age difference could be associated with greater changes in the older age groups than in the young.

Relationship between edge-wise/vertex-wise variability and age groups

Next, the cortical regions and connectomes that exhibited age-related divergence or convergence effects were examined. The thresholded NBS results (see Fig. 4) revealed increased variability, across most of the connectome, in the older age groups. Additionally, some bundles of connections showed a significant age-related convergence effect, such as those connecting to the visual nodes in the HCP-Aging VISMOTOR and CARIT connectomes. To assess the similarity of the REST FC results illustrating the areas of significant age-related divergence in both datasets, the edge level coefficients in both connectomes were correlated; edges that were not located within the significant networks took on the value of 0 for this correlation analysis. The results suggest that both sets of REST FC results were not significantly correlated (r = 0.02, p = 0.083).

Fig. 4.

Connectograms illustrating the relationship between edge-wise variability and age, thresholded at the p_network < .05. The coefficients of this relationship are color-coded using the same hot–cold color scale limits across all connectograms to facilitate comparisons. Darker shades of red suggest that the edge is more variable in the older age groups; darker shades of cyan suggest that the edge is less variable in the older age groups. The Common REST connectogram illustrates the REST FC edges that emerge significant in both datasets

In terms of cortical thickness, the thresholded vertex-wise results revealed significant and widespread clusters that exhibited age-related divergence effects. These clusters were commonly located in the Broca’s areas, cingulate, medial posterior, and temporal regions in both datasets (see Fig. 5). To assess the similarity of these cortical surface maps illustrating the areas of significant age-related divergence in both datasets, the vertex-wise t-statistics in both cortical surface maps were correlated; vertices that were not located within the significant clusters took on the value of 0 for this correlation analysis. The results suggest that both cortical surface maps were significantly but weakly correlated (r = 0.22, p < 0.001).

Fig. 5.

a Cortical surface maps illustrating the relationship between cortical thickness variability and age groups, thresholded at p_cluster < .05. The coefficients of this relationship are color-coded using the same hot–cold color scale limits across both datasets. Darker shades of red suggest that the vertex is more variable in the older age groups; darker shades of cyan suggest that the vertex is less variable in the older age groups. b Common vertices identified to be significant in both datasets

Is FC similarity predicted by CT similarity?

In relation to the secondary goal, the relationship between CT and FC similarity was investigated using HLMs. In the previous analyses, the FC matrices included nodes from the subcortical regions, which could not be measured in terms of CT. Thus, to ensure that both FC similarity and CT similarity covered the same regions in the brain, the 19 subcortical nodes from the FC matrices were excluded in the current set of analyses.

According to the results (see Fig. 6), CT similarity significantly predicted FC similarity significantly on its own and in a demographics-adjusted model across all fMRI paradigms in both datasets. The inclusion of the age difference, age difference², and sex difference covariates into the model resulted in smaller coefficients for CT similarity. These standardized coefficients were mostly small in magnitude (standardized coefficients ≤ 0.30), except for the MOVIE and SMT tasks in the CAM-CAN dataset where CT similarity had a moderate effect (standardized coefficients > 0.30) on FC similarity.

Fig. 6.

Predicting FC similarity with CT similarity in adjusted and unadjusted models. the absolute magnitudes of the coefficients are reported to facilitate comparison; the coefficients for the age difference, age difference², and sex difference were originally negative. The adjusted model included the covariates of age difference, age difference.², and sex difference covariates. The error bars represent its 95% confidence intervals. All coefficients are statistically significant (ps < .001)

Supplementary analyses

The above analyses were repeated using FC matrices generated from a more granular cortical parcellation scheme (i.e., Schaefer-200 atlas + 19 subcortical regions) and CT data resampled on a higher resolution template (i.e., fsaverage6). These supplementary analyses (see supplementary materials) revealed mostly similar findings, notwithstanding an age-related convergence CT cluster in the left lingual gyrus within the CAM-CAN dataset.

Discussion

The current study set out to investigate age-related divergence in CT and FC. In both the HCP-Aging and CAM-CAN datasets, the age-related divergence hypothesis was supported—the between-dyad similarity in CT and FC in all fMRI paradigms has consistently decreased with age. Subsequent analyses then ruled out the possibility that this age-related divergence was attributed to the possibility that a 1-year age difference was associated with greater CT/FC changes in the older age groups. Next, to narrow down to the source of this increased variability in the brain, the brain-wide analyses revealed widespread cortical regions and connectomes that exhibited significant effects of age-related divergence. Finally, it was shown that FC patterns were dependent on cortical morphology; CT similarity predicted FC similarity across fMRI paradigms, above and beyond demographical covariates.

Two previous studies had investigated the inter-subject variability in resting-state FC across the lifespan though they were not framed in terms of age-related divergence; both used the CAM-CAN dataset. The first study [38] similarly showed that between-dyad FC similarity decreased with age when the between-dyad comparisons were made within very finely partitioned overlapping age groups as well as within larger non-overlapping age groups. Furthermore, their voxel-wise analyses of FC variability revealed widespread regions of age-related divergence that resemble the cortical regions that demonstrated age-related divergence in CT in the current study. This further reinforces the link between structural similarity and FC similarity. The second study [39] was embedded within a neural fingerprinting context. In this study, between-dyad FC similarity was calculated in the full sample as opposed to within smaller age group partitions. The lifespan trend of these FC similarity correlations appears to fit a U-shaped slope. This slope has a significant negative age coefficient and positive coefficient, albeit magnitudes weaker, for its quadratic term. In other words, there is an overall age-related decline in FC similarity; however, this decline slows down or even reverses to a very minor extent in the older age groups. In relation to these studies, the current results not only corroborated the age-related divergence phenomenon in another dataset (i.e., HCP-Aging) but also showed that this divergence was similarly observed in CT and FC across all included task-based paradigms.

Judging by the rates of divergence across the different tasks, one could deduce that tasks that are more cognitively demanding (e.g., CARIT and MOVIE) will show larger age-related divergence across the lifespan. This can perhaps be explained by the fact that more demanding tasks require greater compensatory activation from auxiliary brain regions in the old [10]. These compensatory mechanisms can be highly variable, even among the healthy population. Aging is associated with a diverse range of preservation and degeneration of brain structures and functions [40, 41]. As the amount of compensation increases accordingly with task demands, so will its associated variability in brain activity, thus resulting in greater age-related divergence in the FC during a cognitively demanding task.

In terms of the specific regions and networks that are susceptible to age-related divergence, the results revealed that both in terms of CT and FC, most of the brain exhibited such an age-related divergence phenomenon. As such, it might be more meaningful to discuss the cortical regions and connectomes that did not adhere to these age-related divergence patterns. In this regard, it was observed in both datasets that the medial occipital lobe in both hemispheres did not show any age-related divergence in CT. In fact, in a supplementary analysis (see supplementary materials) using CT data sampled onto a higher resolution template, a significant age-related convergence effect was observed in the left lingual gyrus within the CAM-CAN dataset. Furthermore, bundles of edges connecting to the visual network were observed to exhibit age-related convergence effects across most task-based and resting state connectomes; in particular, this was most prominently observed during the VISMOTOR task. Taken together, these findings converge to suggest an absence of age-related divergence and instead age-related convergence within visual-related regions and networks. This absence of age-related divergence might be related to the age-related hypoactivity in the visual network. A previous meta-analysis [42] pooled several fMRI task paradigms together and reported significant clusters of hypoactivity mostly in the visual network when healthy older, relative to younger adults, performed the tasks. This consistent hypoactivity would in turn be associated with consistent connectivity patterns and plastic changes, or lack of, in the visual network among the older age groups. This was further supported by a longitudinal study reporting highly homogeneous slopes of decline in cortical thickness, surface area, and volume of the occipital region, relative to the rest of the brain [43]—suggesting little room for neuroplastic mechanisms to alter their downward trajectories.

Finally, it was shown that between-dyad FC similarity was significantly predicted by their CT similarity; this was true even after controlling for the dyad’s age and sex differences. This showed that FC patterns are significantly dependent on cortical structure; a subject will typically perform a task in a manner that is largely consistent with their previous experience which is encoded in their cortical structure [8, 9]. It appears that the strength of this structure-FC relationship varies across different tasks. Furthermore, the fact that this relationship was stronger in the CAM-CAN dataset could also suggest that different MRI acquisition and preprocessing pipelines can inadvertently enhance or attenuate these structure-FC associations. Interestingly, a much smaller spread of between-dyad CT correlations, relative to those of FC, was observed consistently in both datasets. This suggests that the between-dyad FC similarity can vary in a more extreme manner. Speculatively, the larger spread of between-dyad FC similarity could suggest that there are other unstudied between-dyad factors, such as structural connectivity, which could also add to the differences or similarities in FC between dyads. In relation to this, it was previously shown that FC is highly dependent on structural connections [44].

The study presents major implications in various contexts. First, robust evidence derived from two independent samples was presented to debunk the myth that “all old people are the same.” In particular, given the neuroscientific nature of these findings, they can convincingly address stereotypes among the older population relating to certain traits such as personality, which may be perceived to be the consequence of having an aged brain. Next, the reduced between-dyad similarity or increased variability observed among older brains would be of major interest to future researchers for a few reasons. Given that the between-dyad similarity is directly related to neural fingerprinting accuracy, this would mean that neural fingerprinting studies will be able to identify older subjects among their similar-aged peers more accurately or successfully than if such neural fingerprinting was carried out in a younger age group. In other words, the accuracy of neural fingerprinting is dependent on the sample’s age group. Next, the study of brain-behavior associations would certainly benefit from an increased variability across brains; this increased variability among the old would mean that a significant brain-behavior association will be detected more easily. These brain-behavior associations are also likely to be more generalizable given that they are derived from a relatively diverse selection of brains. In another study using the CAM-CAN dataset [45], it was reported that behavioral prediction models trained using multimodal neuroimaging data from an old age group generalized well to age groups beyond its own. Conversely, when a young age group was used to train the behavioral prediction models, the models’ predictions generalized poorly in age groups beyond its own. Taken together, researchers studying neural fingerprinting and brain-behavior associations need to consider the possibility that their results are significantly influenced by the age group of their sample.

The current findings are subject to some limitations. First, the idea of age-related divergence implies that we become more different from our similar-age peers as we age. This should ideally be examined via multiple longitudinal follow-ups of the same birth cohort. Thus, the present cross-sectional approach provides only a rough estimation of this age-related divergence effect and is highly susceptible to cohort-related confounds, such as nutrition and education levels. For instance, it cannot be ruled out that lower nutritional and education levels in the older age-cohorts might have increased their inter-subject brain variability. Second, a large number of participants from both datasets were excluded due to missing or invalid scans in one or more fMRI paradigms. This coupled with the stringent inclusion criteria in both datasets, may inadvertently exert a homogenizing effect on brain variability, especially among the old. Therefore, the estimates of brain variability in both datasets, especially in the older age groups, were likely to be underestimates of the brain variability in the population. Consequently, these underestimates meant that the age-related divergence phenomenon is likely to be stronger in the general population. Third, in order for the age groups to be meaningfully large for between-subject analyses, the age ranges for each of these age groups were defined across different 10-year age bands. These relatively wide age-bands could have inadvertently increased the between-subject variability within the age-groups, on account of the age-related differences in FC and CT. Finally, the different fMRI preprocessing pipeline used in both datasets makes it difficult for one to compare the FC-related results across both datasets and may explain the low consistency in the edge-wise age-related divergence effects across both datasets. Nevertheless, despite the different preprocessing pipelines, age-related divergence effects were similarly observed in the various FC matrices of both datasets.

Supplementary Information

Below is the link to the electronic supplementary material.

Funding

Junhong Yu is supported by the Nanyang Assistant Professorship (Award no. 021080–00001). The HCP-Aging data used in the preparation of this manuscript were obtained from the National Institute of Mental Health (NIMH) Data Archive (NDA). NDA is a collaborative informatics system created by the National Institutes of Health to provide a national resource to support and accelerate research in mental health. Dataset identifier(s): https://doi.org/10.15154/6faj-nf83. This manuscript reflects the views of the authors and may not reflect the opinions or views of the NIH or of the Submitters submitting original data to NDA. The HCP-Aging study was supported by the National Institute On Aging of the National Institutes of Health under Award Number U01AG052564 and by funds provided by the McDonnell Center for Systems Neuroscience at Washington University in St. Louis. The HCP-Aging 2.0 Release data used in this report came from https://doi.org/10.15154/1520707. Data collection and sharing for the CAM-CAN dataset was funded by the UK Biotechnology and Biological Sciences Research Council (grant number BB/H008217/1), together with support from the UK Medical Research Council and the University of Cambridge, UK.

Declarations

Conflict of interest

The author declare no competing interests.