Cortical iron mediates age‐related decline in fluid cognition

Abstract Brain iron dyshomeostasis disrupts various critical cellular functions, and age‐related iron accumulation may contribute to deficient neurotransmission and cell death. While recent studies have linked excessive brain iron to cognitive function in the context of neurodegenerative disease, little is known regarding the role of brain iron accumulation in cognitive aging in healthy adults. Further, previous studies have focused primarily on deep gray matter regions, where the level of iron deposition is highest. However, recent evidence suggests that cortical iron may also contribute to cognitive deficit and neurodegenerative disease. Here, we used quantitative susceptibility mapping (QSM) to measure brain iron in 67 healthy participants 18–78 years of age. Speed‐dependent (fluid) cognition was assessed from a battery of 12 psychometric and computer‐based tests. From voxelwise QSM analyses, we found that QSM susceptibility values were negatively associated with fluid cognition in the right inferior temporal gyrus, bilateral putamen, posterior cingulate gyrus, motor, and premotor cortices. Mediation analysis indicated that susceptibility in the right inferior temporal gyrus was a significant mediator of the relation between age and fluid cognition, and similar effects were evident for the left inferior temporal gyrus at a lower statistical threshold. Additionally, age and right inferior temporal gyrus susceptibility interacted to predict fluid cognition, such that brain iron was negatively associated with a cognitive decline for adults over 45 years of age. These findings suggest that iron may have a mediating role in cognitive decline and may be an early biomarker of neurodegenerative disease.

In human development, iron accumulates in basal ganglia and deep gray matter regions through adolescence (Li et al., 2014). Some regions, such as the putamen, continue to accumulate iron throughout later adulthood (Bartzokis et al., 1997;Hallgren & Sourander, 1958;Ramos et al., 2014). While the level of iron is lower in the cortex than in the deep gray matter, age-related iron accumulation is detectable throughout the cerebral cortex (Khattar et al., 2021). The significance of age-related increases in brain iron is not yet entirely understood.
Most studies investigating brain iron in aging and neurodegenerative diseases have analyzed relations within predefined regions of interest (ROIs), particularly those in deep gray matter. This approach is not sensitive to effects beyond these ROIs and involves estimating iron across large cortical surfaces. Thus, whole-brain investigations may reveal novel findings. For example, whole-brain QSM analyses of Alzheimer's disease (Acosta-Cabronero et al., 2013) and Parkinson's disease (Uchida et al., 2019) suggest that iron deposition in distinct cortical regions contributes to neuropathology. Additionally, in healthy aging, Acosta-Cabronero, Betts, Cardenas-Blanco, Yang, and Nestor (2016) reported age-iron relations in basal ganglia and frontal motor regions by utilizing a whole-brain approach. Investigation of iron in cortical as well as deep gray matter regions may consequently help to clarify the neural mechanisms of age-related decline in cognition, which is most pronounced for speed-dependent (fluid) and sensorimotor functioning (Baltes & Lindenberger, 1997;Craik & Salthouse, 2000;Salthouse, 1992Salthouse, , 1996Salthouse & Madden, 2007). Zachariou et al. (2020), for example, found that within a group of older adults, high cortical iron concentration was associated with disrupted functional connectivity of frontoparietal networks and reduced working memory performance. To our knowledge, however, there has been no whole-brain or voxelwise analysis investigating the relation of brain iron to age-related differences in cognitive function.
Here, we used a whole-brain, voxelwise QSM analysis to investigate age-related differences in relation between brain iron and fluid cognition, in healthy adults. We expected to replicate previous studies demonstrating age-related brain iron accumulation within frontal sensorimotor regions (Acosta-Cabronero et al., 2016;Zachariou et al., 2020). Additionally, we hypothesized that cortical brain iron, particularly within visual and sensorimotor regions, mediates agerelated decline in fluid cognition, in view of the strong dependence of age-related cognitive decline on sensory functioning (Baltes & Lindenberger, 1997;Monge & Madden, 2016;Schneider & Pichora-Fuller, 2000). Further, we hypothesized that brain iron is a relatively early biomarker of age-related declines in fluid cognition, such that the brain iron-cognition association is stronger in mid to late life. We first validated our QSM measures with data from histology and then utilized voxelwise QSM to analyze the relation of iron to both age and fluid cognition. We then tested whether brain iron within identified cognition-related clusters has a mediating role in the age-cognition relation. Finally, we investigated differences in the iron-cognition relation across age.

| Participants
Ninety-eight community-dwelling individuals between 18 and 78 years of age gave written informed consent and were enrolled in the study protocol approved by the Duke University Institutional Review Board. All participants were compensated for their time and completed an initial screening (Session 1) that included three psychometric tests: the vocabulary and logical memory subtests of the Wechsler Adult Intelligence Scale-Revised (WAIS-R; Wechsler, 1981), the Mini-Mental State Exam (MMSE; Folstein, Folstein, & Mchugh, 1975), and the Beck Depression Inventory (BDI; Beck, 1978). Participants were excluded if the scaled score on the WAIS vocabulary subtest was less than 10 (50th percentile), the MMSE score was less than 27, or the BDI score was greater than 16.
See Table 1 for participant demographics and inclusion measures.
During Session 1, participants also performed 12 tests of fluid cognition, detailed below. Imaging was conducted approximately 1 month later in Session 2. Eleven participants were excluded based on Session 1 screening criteria. Six individuals declined to participate in the subsequent MRI session, and five additional participants failed MRI safety screening. Finally, nine participants were excluded for poor image quality (due to either motion or missing slices due to technical errors).

| Cognitive assessments
Cognitive testing assessed three domains of fluid cognition: perceptual-motor speed, executive function, and episodic memory.
Twelve neurocognitive tests were administered, comprising four tests for each domain. These were a combination of computer-based tests of reaction time (RT) and standardized psychometric tests. Computeradministered tests were developed in-house and described in more detail in Madden et al. (2017a). Each domain included one test from the cognition section of the NIH Toolbox (Gershon et al., 2013 Reitan, 1986), and the computed score on the NIH Toolbox Dimensional Change Card Sort Test. We assessed episodic memory from two computer-administered tests (a 6-item shape change detection task; Saults & Cowan, 2007; and 20-min delayed recall of 16 words), one psychometric test (WAIS logical memory delayed; Wechsler, 1997), and the computed score for the NIH Toolbox Picture Sequence Memory Test.
We used a factor analytic approach (Dagley et al., 2017;Hedden et al., 2012Hedden et al., , 2014Madden et al., 2017b;Salthouse et al., 2015) to define a latent construct for perceptual speed, executive function, and memory in terms of the first factor for the four tests in each of these domains. We extracted the first unrotated factor from a factor analysis of the four tests in each domain. We conducted a principal axis factor analysis on the relevant indicator variables (for all participants) and used the factor score (for each participant) from the first unrotated factor as the summary measure. We used a principal factor analysis rather than a principal components analysis because our interest was in the shared variance among the indicator variables rather than the mathematically independent components (Salthouse et al., 2015). The factor scores served as summary measures for data reduction. An overall measure of fluid cognition was derived from the first factor derived from all 12 tests and the factor scores were covaried for sex and WAIS vocabulary.

| Image acquisition
Anatomical imaging data were collected on a 3T GE Signa Ultra High Performance (UHP) Signa MR360 whole-body 60 cm bore MRI scanner (GE Healthcare, Waukesha, WI) equipped with a peak strength of 113 mT/m gradients and a 260 T/m/s slew rate. The scanner possessed a 48-channel head coil that was used for radiofrequency reception. Participants wore earplugs to reduce scanner noise, and foam pads were used to minimize head motion. Three-plane (straight axial/coronal/sagittal) localizer fast spin-echo images were acquired at the start of the scan to define the data collection volume. Global field homogeneity was ensured by the use of a semi-automated high-order shimming program. One run of T1-weighted anatomical images, one resting-state run of T2*-weighted (functional) BOLD (blood-oxygenlevel-dependent) contrast imaging, one run of Susceptibility Weighted Angiography (SWAN) sequence, four runs of event-related T2*weighted imaging, one run of DWI, and one run of T2-weighted FLAIR imaging were recorded. The DWI, task, and resting-state imaging data are not reported here. Anatomical T1-weighted images included 292 straight axial slices that were attained using a 3D fast inverse-recovery-prepared spoiled gradient recalled (

| Susceptibility estimation
Susceptibility map reconstruction was performed with STI Suite v 3.0 (https://people.eecs.berkeley.edu/$chunlei.liu/software.html) in MATLAB (version 2017a) and included phase unwrapping, background field removal (necessary because of the air-tissue interface), and susceptibility reconstruction. The STI Suite utilized the 3D Axial SWAN scans, and a binary brain mask was extracted from the average magnitude image using functional magnetic resonance imaging of the brain (FMRIB) software brain extraction tool (Jenkinson, Beckmann, Behrens, Woolrich, & Smith, 2012). Using the unwrapped phase image and the binary brain mask, the improved sparse linear equation and least-squares (iLSQR) (Li, Wu, & Liu, 2011) algorithm was used to reconstruct the susceptibility map.

| Validation of susceptibility data
To validate the voxelwise susceptibility analyses, we compared the susceptibility data to iron concentrations, in deep gray matter ROIs, derived from postmortem histology (Hallgren & Sourander, 1958 The comparison data were iron concentrations derived from postmortem histology (Hallgren & Sourander, 1958 were also included in our analysis. Next, we examined the degree to which the age trend in our data set corresponded to that reported by Hallgren and Sourander (1958).
These authors observed that the adult age-related increase in deep gray matter iron was pronounced for the putamen. We derived an estimated iron value in the putamen for each participant, based on their age and the Hallgren and Sourander regression equation, which related adult age to putamen iron.
In Equation (1), y is the nonhaemin iron in the putamen (mg/100 g fresh weight), and x is the individual age in years. We then correlated the predicted values for our sample, based on the Hallgren and The resulting coregistration file was used to warp the magnitude image to the 3D T1-weighted structural image using FLIRT in FSL.
The 3D T1-weighted structural image was then registered to the MNI template using FSL FLIRT and warped using FNIRT. The resulting warp transformation parameters were then used to normalize the segmented gray matter images, magnitude images, and reconstructed susceptibility maps for each into MNI space. Visual inspection of the reconstructed susceptibility map revealed oscillatory susceptibility estimation around the edges of the brain. Because this would cause unreliable whole-brain voxel-wise analysis, we eroded the binary brain masks (previously generated by the average magnitude images) by 3 mm. We then masked the reconstructed susceptibility images by the binary, eroded brain masks of each participant to ensure reliable analyses. Susceptibility map and gray matter images were smoothed with an 8-mm full width at half maximum Gaussian kernel. These previously validated parameters were validated in voxelwise structural analyses, including voxelwise QSM (Uchida et al., 2019). The gray matter images were then thresholded to exclude nongray matter voxels with an 80% confidence level, binarized, and then used to mask the susceptibility images. An overlap image was created, representing the number of participants containing gray matter in each voxel. This overlap image was thresholded and used to mask the analytical area such that only voxels that included at least 20% of the participants (n = 13) were analyzed. To replicate previous findings indicating increased brain-iron accumulation in sensory-motor regions (Acosta-Cabronero et al., 2016), a multiple regression was performed to identify voxels associated with age. Next, a multiple regression was performed to identify voxels associated with general fluid cognition. The resulting beta images were analyzed using the SPM toolbox. Voxels significant at the uncorrected threshold of p <.0001 were cluster corrected at 500 cubic voxels. Voxels that survived family-wises error rate (FWER) correction at p <.05 were also reported ( Figure 1).

| Mediation and moderation analyses
To determine whether susceptibility mediates the relation between age and general fluid cognition, we performed a mediation analysis  cluster's susceptibility mediated the age and fluid cognition relation.
This mediation is a path model in which the relation of age to each of the clusters is a separate a path, the relation of each of the clusters to fluid cognition (controlling for age) is a separate b path, the relation of age to fluid cognition (the total effect of age) is the c path, and the direct effect of age (controlling for all of the cluster mediators) is the c 0 path. We modeled the mediators as operating in parallel; thus each mediator was covaried for the others. A significant a Â b path interaction for a mediator variable would imply that the predictor (age) effect on the outcome variable (fluid cognition) is indirect, operating through that variable rather than direct. The a Â b path interaction effects were tested with bootstrap confidence intervals. To confirm that results were reliable and did not vary due to cluster exclusion for multicollinearity, we also performed separate mediation analyses with each cluster tested as an independent mediator (see Table 4).
Additionally, the moderation effects of age and susceptibility in each cluster were calculated using the PROCESS macro. Finally, for clusters with significant interactions (p <.01), we performed a Johnson-Neyman test to identify the particular point, on the age continuum, at which susceptibility predicts fluid cognition (Johnson & Fay, 1950).

| ROI analysis and validation
Pearson correlations revealed that average susceptibility in the red nucleus and putamen correlated positively with age after Bonferroni correction for multiple comparisons (p <.0056). Age correlations for all other ROIs were not statistically significant (Table 2). For a subset of these ROIs corresponding to those reported by Hallgren and Sourander (1958), a correlation indicated a significant relationship between our susceptibility values and the histologically defined iron ( Figure 3a). Finally, our observed putamen susceptibility aligned very closely with the estimated values based on the Hallgren and Sourander age regression equation (Figure 3b).
F I G U R E 2 Fluid cognition (y-axis) negatively correlates with age (x-axis), p <.0001 T A B L E 2 Susceptibility and age correlation for each of the subcortical brain regions

| Mediation analysis
The mediation analysis revealed that the effects of age on all of the mediators (a paths) were significant (see scatter plots in Figure 5), as well as the total effect of age (c path) on overall fluid cognition. Susceptibility in the right inferior temporal gyrus was significantly related to fluid cognition controlling for age (b path). Additionally, a significant a Â b interaction indicated that susceptibility in the right inferior temporal gyrus was a significant mediator of the relation between age and fluid cognition. The mediation effect was significant but partial, in that the direct effect of age remained significant with the mediators taken into account (Tables 3 and 4). understand the nature of the interaction, we performed a Johnson-Neyman test to identify the age at which the inferior temporal gyrus susceptibility negatively predicted fluid cognition. The results revealed that inferior temporal gyrus susceptibility negatively predicted fluid cognition at 46 years of age and above (p <.05) ( Figure 6). There were no significant moderation effects for any other cluster (all p-values >.05). Although global gray matter volumes did not significantly correlate with global QSM values (r = À.173, p = .162), we performed post hoc analyses to determine if our effects were attributable to regional gray matter volume. We added gray matter volumes of the significant right inferior temporal cluster were added to the mediation and moderation models as covariates. In these models the effects for both mediation (see Table S1) and moderation, F (1, 62) = 7.40, p = .009, R 2 change = .027, remained significant. Note: Interaction terms tested using confidence intervals (CIs). Zachariou et al., 2020). This study also presents the novel finding that iron in several other cortical regions, including the postcentral gyri, posterior cingulate, and middle and inferior temporal gyri, are associated with age and fluid cognition. In particular, iron within the right inferior temporal cortex was significant mediator of age-related decline in fluid cognition, such that increasing iron in this region was associated with decreasing cognition, particularly for individuals beyond 45 years of age ( Figure 6).

| DISCUSSION
Results from our ROI analysis show that QSM, a relatively recent MRI technique, is a valid measure for estimating brain iron. We were able to validate the striatal distribution across regions ( Figure 3a) and the histology-predicted degree of age-related increase of iron in the putamen (Figure 3b). While in-vivo estimates of brain iron using the ROI approach have elucidated its relation to cognition (Daugherty et al., 2015;Penke et al., 2012;Sullivan et al., 2009;Van Bergen et al., 2018;Zachariou et al., 2020), our results suggest that the voxelwise approach, which is not constrained to predefined areas or averaging over large cortical structures, reveals novel relations.

| Age-related brain iron accumulation
Age-related iron accumulation in the putamen and sensorimotor regions is well established in previous findings that utilize various iron-measurement techniques, including histology (Hallgren & Sourander, 1958;Ramos et al., 2014), in vivo ROI (Daugherty et al., 2015), and whole-brain analyses (Acosta-Cabronero et al., 2016). In line with these findings, our age-related, voxelwise QSM analysis found that iron was positively associated with age in several sensorimotor regions (putamen, sensorimotor cortices, and the cingulate gyrus). We also found age-related iron deposition in nonmotor areas (visual cortex and middle temporal gyrus) (Figure 4a).
The detection of cortical regions outside sensory areas suggests an advantage of the current whole-brain approach, revealing novel relations beyond predefined ROIs. However, it should be noted that the current approach has some limitations. In particular, the variability of QSM susceptibility in some structures, such as the hippocampus (Daugherty et al., 2015;Rodrigue, Daugherty, Haacke, & Raz, 2013), may mask age-related effects, especially for whole-brain, non-ROI based approaches (Acosta-Cabronero et al., 2016).

| Brain iron accumulation relates to fluid cognition
The fluid cognition-related voxelwise QSM analysis found that iron was negatively associated with fluid cognition in many regions related to age (putamen, sensorimotor cortices, and cingulate gyrus) and one cluster that we did not detect in the age-related results (right inferior temporal gyrus). Since fluid cognition highly negatively correlates with age (Figure 4b), the overlapping results are not surprising. The mediation analysis revealed that brain iron in the right inferior temporal gyrus cluster was associated positively with age p = .0001 (Table 3).
Therefore, we suspect that we did not find this result in the agerelated voxelwise analysis due to cluster thresholding. Indeed, relaxing the cluster-size threshold from 500 mm 3 to 450 mm 3 reveals an agerelated cluster in the right inferior temporal gyrus (cluster size = 476, peak t = 4.69). Further relaxing cluster thresholds also yielded significant clusters in the superior temporal lobe, parahippocampus, cerebellum, and left inferior temporal gyrus. We recognize that voxelwise approaches to structural data have the risk of false negatives (Whitwell, 2009), and that cluster thresholding, while necessary to avoid spurious results, can be arbitrary and mask effects in small clusters.

| Significance of the inferior temporal gyri
While we expected a negative relation between fluid cognition and brain iron, we did not predict that fluid cognition would specifically relate to brain iron in the inferior temporal gyrus. Previous ROI studies have indicated iron in striatal regions, including the putamen, is predictive of dementia ratings (Sullivan et al., 2009), verbal working memory (Daugherty et al., 2015), and motor function (Daugherty et al., 2015).  Fuller, 2000). Thus, we expected that brain iron in primary visual and sensorimotor regions would mediate the age-cognition relation. While we did observe that visual and sensorimotor areas accumulate iron with age, iron in these regions does not mediate the relationship between age and fluid cognition. One possibility for this discrepancy is that brain iron mediation effects in these regions may be more subtle or only present for older adults. Future studies with a larger sample of the older adult population should attempt to replicate these findings.
Previous studies have reported increased inferior temporal iron in patients with Alzheimer's disease (Tao et al., 2014). The increase of iron in Alzheimer's disease patients has been found to interact with increased neuropathology (e.g., neurofibrillary tangles) and predict cognitive decline rate and magnitude (Ayton et al., 2019).
Since the current study did not measure these neuropathologies, we cannot rule out that they could mediate our findings. However, all the present participants exhibited MMSE scores of 27 and above (Table 1), so the presence of undetected Alzheimer's disease is unlikely.
The mechanism specific to the inferior temporal cortex that causes brain iron to impact cognition in aging or neuropathology remains unknown. As discussed previously, iron accumulation may contribute to cell death (Block, Zecca, & Hong, 2007)  Importantly, it is unlikely that true effects are lateralized, and that the right inferior temporal cortex cluster is the only region in which brain iron acts as a mediator. As previously noted, voxelwise approaches to structural data risk false negatives, and cluster thresholding can be arbitrary and mask effects in small clusters. In fact, a sub-threshold cluster (200 mm 3 ) in the left inferior temporal cortex also exhibited a negative relation to fluid cognition (see Figure S1), and with confidence intervals relaxed to 92.5, mediation of the age-fluid cognition relation was also significant for this left inferior temporal region (Table 5). Future studies should also probe how brain iron in the inferior temporal cortex and other regions interacts with other determinants of age-related declines in fluid cognition, such as white matter integrity and functional connectivity.

| Brain iron as a biomarker
Previous studies have suggested that brain iron accumulation may be a biomarker of mild cognitive impairment (Sun et al., 2017).
Although previous work targeted brain iron in subcortical regions, the present findings demonstrated that iron in supratentorial cortical regions, in this case inferior temporal cortex, may have a causal role in cognitive decline in the context of healthy aging. Additionally, we found that increasing iron in the right inferior temporal cortex predicts decreasing fluid cognition beyond 45 years of age.
Data from fMRI also suggest that middle adulthood is a period during which a decline in functional connectivity in visual sensory regions contributes to age-related differences in task-related brain activation (Madden et al., 2017b;Monge et al., 2017). Thus, like declines in functional connectivity, brain iron accumulation is detectable relatively early in healthy aging and neurodegenerative disease and has predictive potential. However, the cross-sectional nature of the current study limits our conclusions regarding agespecificity. Longitudinal studies tracking brain iron in conjunction with age-related cognitive decline and MCI onset should lend further insights.

| CONCLUSIONS
This study demonstrates that QSM is a valid measure of brain iron that is useful for understanding the interaction of adult age, iron, and cognition. Here, we found that increasing brain iron in the right inferior temporal cortex contributed to increased decline in fluid cogni- Note: Interaction terms tested using confidence intervals (CIs) of 0.925.