Sub‐Networks of the Brain R₂ Structural Connectome Integrity Matrix (SCIM) and Older Adult Cognitive and Motor Functioning

ABSTRACT

Neuroimaging the structural connectome provides a window into the aging brain; however, few studies address the complexities of mapping late‐life structural connectivity. We defined white matter connections via a structural connectivity‐based atlas and extracted transverse relaxation rates (R₂) forming a structural connectome integrity matrix to neuroimaging data of 1239 participants (age ~ 79 ± 7). Factor analyses revealed four Sub‐Networks (SN) characterized by edge integrity as follows: SN‐1 involved most frontal nodes, all parietal nodes, and key subcortical (basal ganglia) structures; SN‐2 involved most albeit slightly different frontal nodes than SN‐1, nearly all temporal and key subcortical (limbic) nodes; SN‐3 was primarily characterized by edges involving select parietal and temporal and all occipital nodes; SN‐4 was confined to cerebellum, basal ganglia and limbic nodes. A linear mixed‐effects regression model containing weighted composite scores representing each Sub‐Network and adjusting for relevant confounders demonstrated associations of lower R₂ in SN‐1, SN‐2, and SN‐4 with lower baseline global cognition and lower R₂ in SN‐2 and SN‐3 with faster declines in global cognition. Sub‐Networks were also differentially associated with domain‐specific cognitive functions at baseline and over time. Nearly all Sub‐Networks negatively associated with global motor function, dexterity and gait speed at baseline, but only SN‐1 and SN‐4 were associated with change in motor functioning, specifically gait speed, over time. Our approach to the aging structural connectome provides an assessment of its R₂ integrity and underlying sub‐networks as related to critical behaviors associated with normal and pathological aging.

Key Points

• Neuroimaging the structural connectome provides a window into the aging brain; however, few studies address the complexities of mapping late‐life structural connectivity.

• We provided a more holistic approach to measure brain white matter network integrity that considers age‐related alterations in brain tissue in over 1200 older adults.

• Results point to meaningful sub‐network differentiation and unique cognitive and motor associations over time.

In over 1200 older adults, we interrogated structural connectome integrity via transverse relaxation rates of 308 white matter connections and identified four major R₂‐SCIM sub‐networks. These Sub‐Networks differentially contributed to levels of cognitive and motor functioning as well as changes in cognitive more so than motor functioning over time.

1. Introduction

Neuroimaging provides a window into the human brain, shedding light on brain‐behavior relationships previously only understood through animal models and case studies of focal lesions. Historically, most neuroimaging studies involving older adults have not investigated the brain as a connected structure, and too often those that have failed to account for the fact that one age‐related alteration, i.e., a brain change connected with age such as the presence of gray and/or white matter lesions, may complicate the quantification of interest. More work is needed investigating late‐life brain network connectivity using approaches that consider the fact that age‐related alterations in brain tissue may alter that investigation.

For example, the Human Connectome Project Brain Connectivity Toolbox (BCT) has been used to investigate the complex neural circuitry of late‐life brain functional (Pedersen et al. 2021) and structural (Vermunt et al. 2020) network integrity. Few studies, regardless of modality, however, consider complications imposed by known age‐related alterations in brain morphometry. This is even though many connectome studies rely on tractography to assist in quantifying complex neural circuitry; tractography negatively impacted by the presence of gray and/or white matter lesions (Jolly et al. 2021). For example, white matter lesions including infarcts, white matter hyperintensities (WMHs), and hemorrhages obscure fiber tractography and reduce the accuracy of differentiating and reconstructing white matter tracts infiltrated by or neighboring such lesions (Seiler et al. 2018; Rudolph et al. 2024; Fox 2018). Tractography limitations secondary to these alterations also have downstream effects on associations of network integrity with cognitive and motor behavior (Rudolph et al. 2024; Langen et al. 2018). Work is needed that considers how best to assess structural brain connectivity in older adults with age‐related brain alterations.

Several approaches exist to address the challenges involving tractography studies of structural network integrity in older adults. Some studies employ “lesion‐filling” approaches in which lesion voxel intensities are replaced by values of a normal distribution generated from the mean white matter signal intensity (Lawrence et al. 2018); others employ lesion “avoidance” where a WMH lesion mask is applied as regions of avoidance for tractography (Taghvaei et al. 2023). Other studies statistically adjust for lesions (Xu et al. 2018; Tuladhar et al. 2016) or more general gray matter atrophy (Vermunt et al. 2020). While reasonable ways to apply connectome analyses developed on younger adults with little to no brain alterations to older adults with age‐ and/or disease‐related brain abnormalities or focal lesions, they likely underestimate the impact of such alterations on brain connectivity (Rudolph et al. 2024; Taghvaei et al. 2023). Another potential hindrance to widespread use of tractography techniques in older adults is that BCT metrics require sufficient sampling of q‐space, which may entail long scan times or specialized hardware, limiting their utility in studies employing clinical scanners.

We applied an emerging alternative approach (Jolly et al. 2021; Schulz et al. 2021; Mallas et al. 2021) using an atlas‐based definition of the path of white matter connections based on healthy younger adults (Mallas et al. 2021) as previously applied to diffusion data to build a structural connectome integrity matrix (SCIM) for over 1200 individuals. We defined the path of white matter connections or “edges” connecting pairs of gray matter regions or “nodes” using a human brain connectivity atlas developed in‐house (Qi and Arfanakis 2021) and already employed in other studies (Mallas et al. 2021), and characterized the integrity of edges based on R₂ transverse relaxation rates. We chose R₂—as opposed to the previously used diffusion data –because of known links between R₂ and cognition even after accounting for Alzheimer's and cerebrovascular‐related neuropathologies (Dawe et al. 2016, 2018). Furthermore, R₂ is sensitive to various white matter anomalies relevant to the older adult brain including demyelination, axonal loss, edema, and iron deposition (Hikita et al. 2005; House et al. 2008) allowing us to focus on the comprehensive characterization of the integrity of white matter connections using a single modality. Taken together, the unique contributions and comprehensive nature of R₂ (Dawe et al. 2016, 2018; Hikita et al. 2005; House et al. 2008) led us to apply it when building our first SCIM. This is not to suggest the same SCIM approach should not be used for other metrics including, but not limited to, diffusion tensor imaging (DTI) and quantitative susceptibility mapping (QSM). In fact, in this initial study, we investigated the inherent connectivity profile(s) of R₂‐SCIM data in older adults, assessed their robustness to methodological variations, created resulting sub‐network specific composite scores, and examined their associations with levels of and changes in global and domain‐specific cognitive and motor functioning; and this work lays the foundation to extend our approach to other neuroimaging modalities.

2. Materials and Methods

2.1. Study Populations

Individuals contributing to this project were participants in one of five ongoing Rush Alzheimer's Disease Center (RADC) longitudinal, prospective cohort studies; specifically, the Religious Orders Study (ROS, 1994 to present) (Bennett et al. 2018), the Rush Memory and Aging Project (MAP, 1997 to present) (Bennett et al. 2018), the Minority Aging Research Study (MARS, 2004 to present) (Barnes et al. 2012), the African American Clinical Core (AACore, 2008 to present) (Schneider et al. 2009), and the Latino Core (LATC, 2015 to present) (Marquez et al. 2020). The design and operation of these five cohorts are identical in essential details; eligibility required older age, absence of known dementia at study enrollment, and agreement to annual clinical evaluations in the participant's preferred language (i.e., English or Spanish). Recruitment approaches, however, do vary: ROS recruits older Catholic priests, nuns, and lay brothers from religious orders across the United States while all other cohorts recruit older adults from the metropolitan Chicagoland area with or without an ethno‐racial focus. For example, MARS and AACore recruit older African Americans while Latino Core recruits older Latinos. Biennial 3T MRI scans have been conducted since 2012 for ROS, MAP, MARS and AACore, and 2018 for LATC. The Institutional Review Board of Rush University Medical Center approved all studies and participants provided written informed consent for all aspects of the study in accordance with the Declaration of Helsinki.

2.2. Eligible Participants

Of the 5404 participants (1503 ROS; 2356 MAP; 858 MARS; 410 AACore; 277 LATC) who had completed the initial parent study evaluation by the time of these analyses, a total of 1239 participants (67 ROS; 712 MAP; 241 MARS; 128 AACore; 91 LATC) completed at least one in vivo 3T MRI scan with processed R₂‐SCIM data. This group was used to characterize R₂‐SCIM profiles at the first valid 3T MRI (analytic baseline) as described below. For additional analyses involving longitudinal cognitive and motor data, we excluded 192 of the 1239 participants who did not have at least two valid completed cognitive and motor evaluations at or after the first valid 3T MRI. This resulted in a group of 1047 participants for the behavioral analyses.

2.3. MRI Data Acquisition and Processing

Neuroimaging was conducted using one of two 3 Tesla scanners located near participants' neighborhoods. Participant burden was further reduced by providing round‐trip transportation to the MRI facility. Of the 1239 participants with a valid MRI, 484 (39%) were imaged using a 3T Philips Achieva and 755 (61%) using a 3T Siemens Trio scanner. Multi‐echo 2D fast spin‐echo (FSE) MRI data were collected on all participants with similar imaging parameters across the two scanners: five echoes with echo time (TE) = n × 20 milliseconds (ms) where n = 1–5, a repetition time (TR) > 5500 ms, acceleration factor of 2, field of view (FOV) of 230 mm × 184 mm, slice thickness of 3 mm, image matrix 256 × 240, resulting in a voxel size of 1.3 × 0.9 × 3mm³ for a scan time of less than 8 min.

R₂ maps were generated using the multi‐echo 2D FSE data (Arfanakis et al. 2007). In each voxel, we fit the signals generated from the multiple echoes to a mono‐exponential decay, S _i = S ₀ exp(−TE_i·R₂), using a least squares approach, where S_i is the signal measured at TE_i, and S₀ is the fitted signal at TE = 0 ms. Each participant's R₂ maps were transformed to the space of the Illinois Institute of Technology (IIT) Human Brain Atlas (v.5.0) (www.nitrc.org/projects/iit). Specifically, the multi‐echo 2D FSE images collected with TE = 40 ms (and subsequently used to generate the R₂ maps) were registered to the b = 0 s/mm² images of the IIT atlas using ANTs non‐linear registration (Avants et al. 2008). The resulting transformation was then applied to the corresponding R₂ maps with trilinear interpolation.

2.4. Creation of the Structural Connectome Integrity Matrix (SCIM) and Its Application to R₂

The path of white matter connections, or “edges”, connecting pairs of gray matter regions or “nodes” was defined for each participant using the structural connectome of the IIT atlas (Qi and Arfanakis 2021). The tractogram of the IIT Human Brain Atlas v.5.0 was generated using probabilistic tractography on the IIT HARDI template and used to define a symmetric 88 × 88 connectivity matrix, each element of which includes the total number of streamlines connecting two gray matter nodes. Track density images for all connections are also available. We thresholded the IIT connectivity matrix by 5% of the maximum number of streamlines in the matrix and identified the 308 strongest unique white matter connections. Also consistent with the literature (Qi and Arfanakis 2021; Rubinov and Sporns 2010), we then thresholded the track density images of each of the 308 connections by 5% of the corresponding maximum track density value to identify the most likely path of each connection and converted the thresholded track density images to masks.

As previously stated, this atlas‐based approach was used in light of the inherent challenges of person‐specific tractography in the presence of white matter lesions such as those related to aging (Seiler et al. 2018; Rudolph et al. 2024; Fox 2018; Langen et al. 2018). Thus, we used the masks of the 308 connections as described above to extract the median R₂ value for each connection of each participant in the IIT atlas space, choosing median values to mitigate the influence of any potential spikes in R₂ values. Finally, an R₂ structural connectome integrity matrix (R₂‐SCIM) was generated for each participant that included the median R₂ values for the 308 white matter connections. For quality assurance (outlined in the Statistical Analysis section below), we also calculated participant‐specific global R₂‐SCIM scores by averaging the 308 values of each participant's R₂‐SCIM.

2.5. Cognitive and Motor Assessment

Details of the annual cognitive evaluation have been described elsewhere (Bennett et al. 2018; Barnes et al. 2012; Marquez et al. 2020). Briefly, in all cohorts, each annual evaluation included the administration of a comprehensive neuropsychological protocol. Eighteen test measures were used to create a global cognitive function composite score with these same test measures further divided into five cognitive domains. Specifically, episodic memory (Wechsler Memory Scale—Revised Logical Memory Story A: Immediate and Delayed Recall; CERAD Word List Memory: total recall Trials 1 through 3, Delayed Recall, and Recognition; East Boston Story: Immediate and Delayed Recall), semantic memory (Category Fluency: sum of animals and fruits/vegetables; 15‐item CERAD‐version of the Boston Naming Test; 10‐item National Adult Reading Test), perceptual orientation (15‐item Judgment of Line Orientation; 16‐item Raven's Standard Progressive Matrices), perceptual speed (Number Comparison; Symbol Digit Modality Test; Stroop: Word Reading and Color Naming), and working memory (total correct for: Digit Span Forward; Digit Span Backward; Digit Ordering). The global composite cognitive function score was calculated by converting individual raw scores to z‐scores using the baseline mean and standard deviation from the combined parent studies; at each assessment, the z‐scores were averaged, with higher scores indicating better function. The composite scores for the five cognitive domains were calculated similarly to the global composite using the specific items outlined above. No normative data were used in the calculation of these composite scores which have been validated across RADC cohorts and racial groups (Wilson et al. 2002; Barnes et al. 2016).

All participants were also evaluated on motor and gait performance as part of their annual evaluations (Buchman et al. 2019). Hand strength was evaluated via grip and pinch, measured bilaterally and computed separately, using the Jamar hydraulic hand dynamometer (Lafayette Instruments, Lafeyette, IN, USA). Upper extremity dexterity was based on the average of four trials (2 right, 2 left) of both the Purdue Pegboard and an index finger tapping task registered via an electronic device (Western Psychological Services, Los Angeles, CA, USA). Gait was evaluated by the time (in seconds) and the number of steps taken to execute an eight‐foot walk and 360° turn, respectively. Balance was also measured through leg and toe stand tasks. All 10 test items were scaled by the sex‐specific mean at initial (i.e., parent study) visit and then averaged to obtain a global motor functioning score and three motor domains (hand strength, dexterity, and gait) as outlined above and previously validated (Buchman et al. 2011).

2.6. Experimental Design and Statistical Analysis

Statistical summaries of participant characteristics were calculated including means and standard deviations as well as frequency distributions, as appropriate. All 308 unique connections and the global R₂‐SCIM score were reviewed with statistical (means, SDs, ranges, quantiles, skewness and kurtosis coefficients; Table S1) and graphical summaries (scatterplot matrices, QQ plots, and boxplots; Figures S1–S3) to ensure no technical malfunctions resulting in spurious data were missed, and to establish that the data were appropriate for the planned analysis. As all 308 connections were available for each participant, no imputation was needed. Distributions of data were compared by scanner to check for scanner differences (Figure S4). All summary computations and statistical models were programmed in SAS (version 9.4, SAS system for Linux) and supplemented with R (R Foundation for Statistical Computing). All data used in this manuscript may be requested at www.radc.rush.edu.

To identify and characterize R₂‐SCIM sub‐networks underlying the 308 unique connections at analytic baseline given the large number of connections, we applied principal component analysis (PCA) with orthogonal rotation as a data reduction technique. We defined the final number of components by assessing how each additional component contributed to the total variance explained as well as by examining the contributions of individual connections for each component using standard cut‐off ≥ 0.6 to identify high, absolute values reflecting high contributions to the component. For each defined component, PCA‐derived weighted composite scores were calculated consisting of the sum of the 308 R₂‐SCIM values weighted by the component‐specific (i.e., positive or negative) coefficient for each participant (higher composite score = higher R₂ integrity weighted toward the highest contributing connections of a given component). To assess whether the selected final sub‐networks were robust to variations, we replicate the PCA first, for each scanner separately at analytic baseline (Figure S5) and second, considering participants' 2nd and 3rd valid 3T MRI scans (separately) as available (Figure S6). To examine the associations between the PCA‐derived R₂‐SCIM weighted composite scores with decline in cognitive and motor function over time (global scores and domain scores as distinct outcomes collected from the first valid MRI to the last follow‐up visit), we applied separate linear mixed‐effects models with random terms for person‐specific level and slope. The models included a linear function of time (in years since first MRI), the continuous weighted component scores as main predictors, a term to adjust for scanner, and terms to adjust both level and slope for sex, education, and age at first MRI. For cognition only, a time varying term for mode of cognitive testing (typically in person, however, sometimes via telephone, especially during COVID) was included. Significance was set a p ≤ 0.05.

3. Results

3.1. Participant Characteristics

At first valid MRI scan, the 1239 participants contributing to the PCA were, on average, 79 years of age with around 16 years of education. This group was predominantly female (79%) and approximately 40% self‐identified as non‐Latino Black or Latino. Characteristics of the 1047 participants within longitudinal cognitive and motor data were similar and may be found in Table 1. The global R₂‐SCIM at baseline was 0.013 ms⁻¹ on average (SD = 0.0005 ms⁻¹; range = [0.010, 0.014 ms⁻¹]; kurtosis = 3.21; skewness = −0.04). Since the distribution of scores was similar across scanners (Figure S4), we combined data for all analyses.

TABLE 1.

Characteristics of participants at baseline contributing to analyses involving longitudinal cognitive and motor data (n = 1047).

Characteristics
Demographics
Age, mean (SD), years	78.6 (7.5)
Women, no. (%)	828 (79)
Education, mean (SD), years	15.7 (3.6)
Ethno‐racial group, no. (%)
Non‐Latino White	641 (61)
Black/African American	307 (29)
Latino	91 (9)
Other	8 (1)
Cognitive scores, mean (SD)
Global cognition	0.19 (0.57)
Episodic memory	0.25 (0.73)
Semantic memory	0.22 (0.76)
Perceptual orientation	0.21 (0.75)
Perceptual speed	0.13 (0.76)
Working memory	0.12 (0.78)
Motor function scores, mean (SD)
Global	0.97 (0.20)
Hand strength	0.99 (0.28)
Dexterity	1.02 (0.17)
Gait speed	0.90 (0.20)
Scanner, no. (%)
3T Philips	389 (37)
3T Siemens	658 (63)

Note: Baseline was defined as first eligible MRI.

3.2. Principal Component Analysis (PCA) of R₂ ‐SCIM

The first principal component explained 66% of the total variance in R₂‐SCIM. When considering additional components beyond four, the variance explained did not substantially increase; furthermore, loading values for an additional fifth component did not exceed 0.6. Lastly, the four R₂‐SCIM Sub‐Networks reflected interpretable albeit distinct contributions from the highest loading connections, all of which had positive loading values ≥ 0.6 (Figure 1). Thus, we calculated PCA‐derived weighted composite scores focusing on the first four components accounting for 78% of the data variability.

FIGURE 1.

Connectograms representing the loading values of the R₂‐SCIM unique connections contributing with greatest weight (all loading values depicted were positive and ≥ 0.60) to each of the four selected principal components (n = 1239). The number of connections displayed by panel are n = 136 (panel A, Subnetwork 1), n = 87 (panel B, Sub‐Network 2), n = 81 (panel C, Sub‐Network 3), and n = 27 (panel D, Sub‐Network 4). Each of the four sub‐networks is characterized by specific R₂‐SCIM values; only three connections with high contributions (i.e., loadings ≥ 0.6) were found within two different sub‐networks.

Focusing on connections with high contributions (i.e., loading values ≥ 0.6), each of the four components (Figure 1) was characterized by a relatively symmetrical and unique R₂‐SCIM sub‐network. Sub‐Network 1 (SN‐1) was characterized by connections between mostly frontal nodes, all parietal nodes, and key subcortical structures especially within the basal ganglia. These connections most likely represent U‐fibers within frontal and parietal lobes, long association connections between frontal and parietal nodes, and projections connecting individual nodes from each of these lobes to subcortical structures (Figure 1A). Sub‐Network 2 (SN‐2) was characterized by connections between most albeit slightly different frontal nodes than SN‐1, nearly all temporal nodes, and key subcortical structures including those of the limbic system. These connections appeared to represent U‐fibers within frontal and temporal lobes, connections between frontal nodes and the cingulate, as well as projections from frontal and temporal nodes to subcortical structures (Figure 1B). Sub‐Network 3 (SN‐3) was primarily characterized by connections between select parietal and temporal nodes and all occipital nodes including U‐fibers within each lobe as well as association connections between these lobes (Figure 1C). Finally, Sub‐Network 4 (SN‐4) was confined to connections involving the cerebellum, basal ganglia and limbic structures only (Figure 1D). Sub‐networks remained largely similar when fitting the PCA for each scanner separately at analytic baseline (Figure S5) and when considering participants' 2nd and 3rd valid 3T MRI scans (separately) as available (Figure S6).

3.3. R₂ ‐SCIM Sub‐Networks and Cognition

For cognition, the mean follow‐up time from first valid MRI to last visit was 6 ± 3 years, with a mean of 7 ± 3 cognitive assessments over time. The adjusted linear mixed‐effects model examining the associations between the continuous PCA‐derived R₂‐SCIM weighted composite scores at baseline and level of and change in global cognition resulted in associations between the mostly frontal SN‐1, fronto‐temporal SN‐2, and predominantly sub‐cortical SN‐4 with initial level of global cognition (Table 2). There was no association with the parieto‐temporo‐occipital SN‐3. Only SN‐2 and SN‐3 were associated with change in global cognition over time (Figure 2). Thus, lower levels of R₂‐SCIM involving SN‐1, SN‐2, and SN‐4 were associated with lower baseline global cognition while lower levels of R₂‐SCIM involving SN‐2 and SN‐3 were associated with faster rates of global cognitive decline (see Table 2 for adjusted mean differences and p‐values). Observed individual trajectories of global cognition for a subset of randomly selected individuals with low versus high PCA‐derived R₂‐SCIM weighted scores were compared to provide a visualization only, as depicted in Figure S7.

TABLE 2.

Multivariable‐adjusted mean differences in initial level and change of cognitive function, according to continuous PCA‐derived R₂‐SCIM Sub‐Network scores (reversed and standardized for interpretation purposes) at baseline (n = 1047).

	Initial level	p	Annual rate of decline	p
	Mean difference (95% CI)		Mean difference (95% CI)
Global cognition
Sub‐network 1	−0.041 (−0.072; −0.011)	0.01	0.001 (−0.005; 0.008)	0.7
Sub‐network 2	−0.072 (−0.109; −0.035)	0.0002	−0.009 (−0.017; −0.002)	0.02
Sub‐network 3	−0.014 (−0.044; 0.016)	0.4	−0.011 (−0.017; −0.004)	0.001
Sub‐network 4	−0.033 (−0.066; −0.001)	0.045	−0.004 (−0.011; 0.003)	0.2
Episodic memory
Sub‐network 1	−0.038 (−0.077; 0.002)	0.1	0.001 (−0.007; 0.010)	0.8
Sub‐network 2	−0.053 (−0.101; −0.004)	0.03	−0.007 (−0.017; 0.003)	0.2
Sub‐network 3	−0.030 (−0.069; 0.009)	0.1	−0.008 (−0.016; 0.001)	0.1
Sub‐network 4	−0.028 (−0.071; 0.014)	0.2	−0.002 (−0.011; 0.006)	0.6
Semantic memory
Sub‐network 1	−0.034 (−0.074; 0.007)	0.1	−0.004 (−0.012; 0.005)	0.4
Sub‐network 2	−0.103 (−0.153; −0.054)	< 0.0001	−0.007 (−0.017; 0.002)	0.1
Sub‐network 3	0.008 (−0.032; 0.048)	0.7	−0.013 (−0.021; −0.004)	0.003
Sub‐network 4	−0.008 (−0.052; 0.036)	0.7	−0.007 (−0.016; 0.002)	0.1
Perceptual orientation
Sub‐network 1	−0.036 (−0.075; 0.002)	0.1	0.0002 (−0.007; 0.007)	0.9
Sub‐network 2	−0.063 (−0.110; −0.016)	0.01	−0.005 (−0.013; 0.003)	0.3
Sub‐network 3	−0.021 (−0.059; 0.017)	0.3	−0.001 (−0.008; 0.006)	0.7
Sub‐network 4	−0.039 (−0.081; 0.002)	0.1	−0.005 (−0.012; 0.002)	0.2
Perceptual speed
Sub‐network 1	−0.058 (−0.098; −0.017)	0.01	−0.005 (−0.011; 0.002)	0.2
Sub‐network 2	−0.116 (−0.166; −0.067)	< 0.0001	−0.003 (−0.011; 0.005)	0.5
Sub‐network 3	−0.035 (−0.075; 0.006)	0.1	−0.004 (−0.010; 0.003)	0.3
Sub‐network 4	−0.072 (−0.116; −0.028)	0.001	−0.007 (−0.014; −0.00002)	0.049
Working memory
Sub‐network 1	−0.022 (−0.063; 0.020)	0.3	0.002 (−0.004; 0.009)	0.5
Sub‐network 2	−0.055 (−0.105; −0.005)	0.03	−0.011 (−0.018; −0.003)	0.004
Sub‐network 3	0.001 (−0.034; 0.037)	0.9	−0.007 (−0.013; −0.001)	0.04
Sub‐network 4	−0.002 (−0.045; 0.040)	0.9	−0.005 (−0.012; 0.001)	0.1

Note: All analyses adjusted for sex, scanner, mode of cognitive assessment, education, and age at first valid 3T MRI. Sub‐Networks reflect continuous PCA‐derived weighted composite scores of the sum of the 308 R₂‐SCIM values weighted by the component‐specific coefficient (higher score = higher R₂ integrity weighted toward the highest contributing connection of a given component). Bolded values signify p < 0.05.

FIGURE 2.

Mean estimated trajectories of change in global cognition by levels of PCA‐derived R₂‐SCIM Sub‐Networks at baseline (n = 1047). Estimates are derived from a single linear mixed model assuming a linear function of time (years since first valid MRI) and controlling for the four continuous PCA‐derived weighted scores, scanner, sex, age, and education, on both the intercept and the slope, as well as mode of cognitive assessment; within‐participant correlation was captured by correlated random intercept and slope. Mean estimated trajectories (solid lines) with 95% confidence intervals (indicated with shading) are represented for a typical profile of covariates (a woman from the Rush Memory and Aging Project, with 16 years of education, and aged 79 years at first 3T Siemens MRI); the choice of profiles has no influence on the differences in trajectories estimated by the model. We chose two representative levels of weighted composite scores: Low = 10th percentile versus high = 90th percentile. Note: Examples of observed individual trajectories of global cognition may be seen in Figure S7.

We further investigated the associations between all four R₂‐SCIM Sub‐Networks and the five cognitive domains in separate linear mixed‐effects models adjusted for relevant confounders. Only SN‐2 was associated with the initial level of cognition for all domains; SN‐1 and SN‐4 were associated with the initial level of perceptual speed only. SN‐4 was also associated with changes in perceptual speed over time. Additional associations were noted for SN‐2 with change in working memory, and for SN‐3 with change in both semantic and working memory. All associations were positive, suggesting that lower levels of R₂‐SCIM in relevant composite scores were associated with lower baseline levels and faster rates of decline in associated cognitive domains (see Table 2 for adjusted mean differences and p‐values).

3.4. R₂ ‐SCIM Sub‐Networks and Motor Function

For motor function, the mean follow‐up time from first valid MRI to last visit was 5 ± 3 years, with a mean of 5 ± 2 repeated measures over time. The adjusted linear mixed‐effects model examining the associations between the PCA‐derived R₂‐SCIM Sub‐Networks and global motor function resulted in associations of the mostly frontal SN‐1, fronto‐temporal SN‐2, and predominantly sub‐cortical SN‐4 with baseline performance only. There was no association with the parieto‐temporo‐occipital SN‐3. There were no significant associations between any sub‐network and change in global motor function over time (see Table 3 for adjusted mean differences and p‐values).

TABLE 3.

Multivariable‐adjusted mean differences in initial level and change of motor function, according to continuous PCA‐derived R₂‐SCIM Sub‐Network scores (reversed and standardized for interpretation purposes) at baseline (n = 1047).

	Initial level	p	Annual rate of decline	p
	Mean difference (95% CI)		Mean difference (95% CI)
Global motor function
Sub‐network 1	−0.019 (−0.028; −0.010)	< 0.0001	−0.001 (−0.003; 0.001)	0.2
Sub‐network 2	−0.040 (−0.051; −0.029)	< 0.0001	0.001 (−0.001; 0.002)	0.4
Sub‐network 3	−0.006 (−0.015; 0.002)	0.2	0.0005 (−0.001; 0.002)	0.6
Sub‐network 4	−0.025 (−0.034; −0.015)	< 0.0001	−0.001 (−0.002; 0.001)	0.4
Hand strength
Sub‐network 1	−0.010 (−0.025; 0.006)	0.2	−0.002 (−0.004; 0.001)	0.2
Sub‐network 2	−0.033 (−0.051; −0.015)	0.0004	−0.002 (−0.005; 0.002)	0.3
Sub‐network 3	−0.005 (−0.021; 0.010)	0.5	0.0004 (−0.002; 0.003)	0.8
Sub‐network 4	−0.008 (−0.025; 0.008)	0.3	−0.002 (−0.005; 0.001)	0.2
Dexterity
Sub‐network 1	−0.013 (−0.022; −0.005)	0.003	0.001 (−0.001; 0.002)	0.5
Sub‐network 2	−0.034 (−0.044; −0.024)	< 0.0001	0.001 (−0.001; 0.002)	0.5
Sub‐network 3	−0.007 (−0.016; 0.001)	0.1	−0.001 (−0.002; 0.001)	0.4
Sub‐network 4	−0.025 (−0.034; −0.016)	< 0.0001	0.001 (−0.001; 0.003)	0.2
Gait speed
Sub‐network 1	−0.018 (−0.028; −0.008)	0.0004	−0.002 (−0.004; −0.000004)	0.049
Sub‐network 2	−0.028 (−0.040; −0.017)	< 0.0001	−0.001 (−0.003; 0.001)	0.3
Sub‐network 3	−0.005 (−0.014; 0.005)	0.4	0.0002 (−0.002; 0.002)	0.8
Sub‐network 4	−0.025 (−0.035; −0.014)	< 0.0001	−0.002 (−0.004; −0.003)	0.02

Note: All analyses adjusted for sex, scanner, education, and age at first eligible 3T MRI. Sub‐Networks reflect continuous PCA‐derived weighted composite scores of the sum of the 308 R₂‐SCIM values weighted by the component‐specific coefficient (higher score = higher R₂ integrity weighted toward the highest contributing connection of a given component). Bolded values signify p < 0.05.

The three R₂‐SCIM Sub‐Networks (i.e., SN‐1, SN‐2, and SN‐4) that were positively associated with baseline global motor function were also all positively associated with baseline dexterity and gait speed (Table 3). Only SN‐2 was also associated with hand strength at baseline. SN‐1 and SN‐4 were associated with change in gait speed; specifically, lower levels of R₂‐SCIM involving these sub‐networks were associated with faster rates of decline in gait speed. No other associations demonstrated significance.

4. Discussion

In this study of over 1000 older adults, we interrogated structural connectome integrity via transverse relaxation rates of 308 white matter connections, identified major R₂‐SCIM sub‐networks and their robustness to methodological variations, and tested associations of these sub‐networks with cognitive and motor functioning. Analyses characterized four robust R₂‐SCIM sub‐networks that differentially contributed to levels of and changes in cognitive and motor functioning. Taken together, it appears that R₂‐SCIM results in meaningful sub‐network differentiation that is robust to methodological variation and presents with unique cognitive and motor associations in older adults.

Results of this study contribute to the literature in several ways. First, we extended the atlas‐based approach, previously used in traumatic brain injury (Jolly et al. 2021; Mallas et al. 2021) and stroke (Schulz et al. 2021) to relatively more healthy older adults who may have subtle brain alterations that nonetheless complicate quantification for connectome studies relying on person‐specific tractography. Second, by using this atlas‐based definition of the path of white matter connections previously employed with diffusion‐based data, we extracted connection‐specific R₂ values reflecting demyelination, axonal loss, edema, and iron deposition (Hikita et al. 2005; House et al. 2008) that have robust associations with cognition even after accounting for neuropathology (Dawe et al. 2016, 2018). Not only does this extend our understanding of the structural connectome beyond the typically utilized diffusion‐weighted MRI data, it lays the foundation for future studies. Specifically, our atlas‐based definition of white matter connections may be used to extract connection‐specific characteristics from other MRI and even non‐MRI modalities (e.g., T₁ relaxation times, magnetic susceptibility, positron emission tomography, etc.). Such applications, particularly when combined within the same study, will provide a more holistic view of the integrity of the structural connectome and related sub‐networks of the aging brain.

Transverse relaxation rates of white matter connections were the basis of the four R₂‐SCIM Sub‐Networks outlined in this study suggesting similar underlying tissue abnormalities contributing to the highest connections per Sub‐Network. For example, SN‐1 predominantly involved frontal lobe connections. Frontal white matter is known to be most vulnerable to normal aging (Bouhrara et al. 2021), often showing changes in myelin water content earlier than other regions of brain (Bouhrara et al. 2021; Khodanovich et al. 2023). Thus, a commonality of the highest contributing connections of SN‐1 may be that they are myelin rich and late‐myelinating, making them more susceptible to earlier damage than fibers in other brain regions (Gao et al. 2011). While frontal connections are also represented in SN‐2, what distinguishes SN‐2 from SN‐1 is the preponderance of connections in the temporal instead of the parietal lobe, and inclusion of connections from frontal and temporal nodes to limbic structures. SN‐2 may represent R₂ abnormalities that extend beyond vulnerability of myelin rich and late‐myelinating fibers and involve alterations (e.g., axonal loss) associated with neurodegeneration of limbic and temporal lobe structures. SN‐3 appears to reflect common R₂ alterations of shorter U‐ and association fibers given its highest contributing connections are those among select temporal, parietal, and occipital regions. Lastly, the brain alterations driving SN‐4 may be the amalgamation of cerebrovascular anomalies and higher metal deposition in the putamen, caudate, and globus pallidus (Bartzokis et al. 1997; Bartzokis and Tishler 2000) as well as the associated damage (e.g., demyelination, etc.) to their structural connections that may ensue (Pu et al. 2020). Future work should investigate R₂‐SCIM sub‐networks in younger adults to compare their resulting sub‐network profiles to those represented in the current study.

The unique neuroanatomical sub‐divisions reflected by the four R₂‐SCIM Sub‐Networks found in this study were reinforced by their robustness to methodological variation as well as their differential associations with cognitive and motor functioning. For instance, regardless of whether we analyzed the 308 nodal pairs at first valid 3T MRI (i.e., analytic baseline) stratified by scanner or at participants' 2nd and 3rd valid 3T MRI scans (separately) as available, the PCA‐derived profiles remained relatively stable. Furthermore, SN‐1, mainly characterized by connections involving the frontal lobes, was associated with global cognitive and motor performance at baseline, in particular, levels of perceptual speed, dexterity, and gait; SN‐1 was also associated with change in gait speed over time. Relatively similar associations were seen between the predominantly sub‐cortical SN‐4 and cognitive and motor functioning with the addition that SN‐4 was also associated with changes in perceptual speed over time. These associations confirm long‐standing reports (Gunning‐Dixon et al. 2009) that alterations in or disconnections of frontal white matter and, separately, subcortical structures are detrimental to speeded cognitive (Madden et al. 2009) and motor (Kannan et al. 2022) functions for older adults. They further suggest that R₂ alterations of the predominantly sub‐cortical connection contributions of SN‐4 may be more sensitive to changes in perceptual speed over time in our analytic group than the decomposition of the more frontally‐mediated connection contributions of SN‐1. Sensitivity to changes in cognition (i.e., global cognition, semantic and working memory) were most evident for SN‐3, a finding consistent with our previous R₂ studies reporting associations of parietal, temporal, and occipital regions with these same cognitive measures (Dawe et al. 2021; Lamar et al. 2025).

In contrast to all other Sub‐Networks, SN‐2 was associated with level of global and all domain‐specific cognitive and motor scores as well as change in global cognition, especially working memory. Such ubiquitous associations with cognition, as well as baseline motor functioning, provide evidence for the importance of frontal and temporal intra‐lobar connections as well as frontal and temporal connections with the limbic system to age‐related behavior. In fact, previous studies have found that alterations in these regions may signal a level of dysconnectivity that is associated with elevated risk of incident mild cognitive impairment and dementia (Lindemer et al. 2017; Liang et al. 2023). We are actively investigating these outcomes and dementia‐related neuropathologies as related to all R₂‐SCIM Sub‐Networks.

This study points toward critical and relatively robust R₂‐SCIM sub‐network integrity related to normal and, perhaps, pathological aging; however, limitations should be acknowledged. First, we studied only the 308 strongest connections of the IIT atlas. While this selection may be seen as limiting, by targeting only major connections present in all participants we reduced the contribution of smaller or unreliable connections on the definition of sub‐networks, which allowed us to perform more meaningful comparisons. Second, although we believe that conclusions of our work would not have changed, we did not address potential stimulated‐echo contamination. Third, our participants are highly educated and predominantly female; these characteristics may limit generalizability. Fourth, because participation in any neuroimaging study requires a certain degree of mobility and health, our participants do not include the complete range of cognitive performance. Finally, an external cohort validation of our PCA‐profiles was not possible given that, to our knowledge, quantitative R₂ data are not available in another cohort of older adults; however, we tested the robustness of our PCA‐derived sub‐networks via split samples based on scanner and within the same participants at later scanning time points. Additional work is needed, when data are available, to test resulting sub‐networks in other older adult cohort studies.

Despite these weaknesses, strengths of this study include the large analytic group that included non‐Latino White, non‐Latino Black, and Latino older adults. Further strengths include our comprehensive examination of both white matter integrity via R₂ transverse relaxation and behavior. Specifically, we leveraged, on average, 7 years of annual cognitive testing and 5 years of annual motor testing. Furthermore, this study lays the foundation to extend our SCIM portfolio to other MRI modalities (e.g., DTI or QSM) for a multi‐modal interrogation of late‐life brain network integrity. Taken together, our work suggests that there is variation in the R₂ integrity of the structural connectome and that this variation is important for the cognitive and motor functioning of older adults.

Funding

This work was supported by the National Institute on Aging (R01 AG076143, R01 AG022018, R01 AG056405, P30 AG010161, P30AG072975, R01 AG017917, R01 AG062711) and National Institute of Neurological Disorders and Stroke (UH3 NS100599, UF1NS100599).

Disclosure

The Institutional Review Board of Rush University Medical Center approved these cohort studies as well as neuroimaging and participants gave written informed consent for all study procedures in accordance with the Declaration of Helsinki.

Supporting information

Table S1: Descriptives of five key summary statistics (mean, standard deviation [SD], tertiles) of the distributions of the 308 unique R₂‐SCIM at first eligible MRI (n = 1239).

Figure S1: Scatterplot matrix of key summary statistics of the distributions of the 308 unique R₂‐SCIM at first eligible MRI (n = 1239).

Figure S2: Normal quantile‐quantile (Q‐Q) plot of key summary statistics of the distributions of the 308 unique R₂‐SCIM at first eligible MRI (n = 1239).

Figure S3: Boxplots of the 308 unique R₂‐SCIM at first eligible MRI (left panel; descending order of the median values from the entire group) and distribution of the global R₂‐SCIM values (right panel) in the entire group (n = 1239). The red dashed lines indicate the mean value (=0.013) of global R₂‐SCIM.

Figure S4: Distributions of the mean of 308 R₂‐SCIM connections per scanner at first eligible MRI.

Figure S5: Connectograms representing the loading values of the R₂‐SCIM unique connections contributing most (loading values ≥ 0.60) to each of the four selected principal components obtained when fitting the principal component analysis among 484 participants with data from the 3T Philips scanner (left columns) and separately among 755 participants with data from the 3T Siemens scanner (right columns) at first valid 3T MRI. The number of connections displayed by panel are listed directly below each Sub‐Network profile (represented by the four rows) as is the degree of consistency to the “main” (i.e., original) corresponding Sub‐Network profile that considered all scanners at first 1st valid 3T MRI collectively.

Figure S6: Connectograms representing the loading values of the R₂‐SCIM unique connections contributing most (loading values ≥ 0.60) to each of the four selected principal components obtained when fitting the principal component analysis among participants with a second (left column: n = 706; including 243 Philips and 463 Siemens scanners) and a third (right column: n = 409; including 127 Philips and 282 Siemens scanners) 3T MRI assessed over time. The number of connections displayed by panel are listed directly below each Sub‐Network profile (represented by the four rows) as is the degree of consistently to the “main” (i.e., original) corresponding Sub‐Network profile that considered only the 1st valid 3T MRI. All participants were scanned at baseline and each follow‐up using the same scanner (Philips/Siemens); mean follow‐up duration between the first valid 3T MRI and the second and third 3T MRI was 3 (SD = 1) years and 5 (SD = 1) years, respectively.

Figure S7: Individual observed trajectories of global cognition during follow‐up by levels of continuous PCA‐derived R₂‐SCIM Sub‐Networks at first valid 3T MRI. For each panel, and to facilitate visualization only (i.e., not as the result of any statistical comparison between these groups), we chose to contrast cognitive trajectories of 25 randomly selected participants with low (< 10th percentile) and 25 randomly selected participants with high (> 90th percentile) levels of PCA‐derived weighted composite score.

Data Availability Statement

The data supporting the findings of this study, as well as all data for RADC cohort studies, are available at www.radc.rush.edu.