Quantification of mediation effects of white matter functional characteristics on cognitive decline in aging

Abstract

Cognitive decline with aging involves multifactorial processes, including changes in brain structure and function. This study focuses on the role of white matter functional characteristics, as reflected in blood oxygenation level-dependent signals, in age-related cognitive deterioration. Building on previous research confirming the reproducibility and age-dependence of blood oxygenation level-dependent signals acquired via functional magnetic resonance imaging, we here employ mediation analysis to test if aging affects cognition through white matter blood oxygenation level-dependent signal changes, impacting various cognitive domains and specific white matter regions. We used independent component analysis of resting-state blood oxygenation level-dependent signals to segment white matter into coherent hubs, offering a data-driven view of white matter’s functional architecture. Through correlation analysis, we constructed a graph network and derived metrics to quantitatively assess regional functional properties based on resting-state blood oxygenation level-dependent fluctuations. Our analysis identified significant mediators in the age-cognition relationship, indicating that aging differentially influences cognitive functions by altering the functional characteristics of distinct white matter regions. These findings enhance our understanding of the neurobiological basis of cognitive aging, highlighting the critical role of white matter in maintaining cognitive integrity and proposing new approaches to assess interventions targeting cognitive decline in older populations.

Introduction

Aging is a complex and heterogeneous process that often results in a gradual decrease in physical and cognitive abilities. Such cognitive decline frequently manifests as reduced processing speed, challenges in maintaining attention, and memory loss (Buckner 2004; Harada et al. 2013). Numerous studies have highlighted that alterations in brain structure, including volume reductions, histopathological damage, and neurovascular impairments, are common attributes of aging that influence these cognitive shifts (Raz and Rodrigue 2006; Fjell and Walhovd 2010). Beyond these structural modifications, changes in brain functional activities are considered to be more immediate indicators of cognitive alterations, as they represent the brain’s capacity for information processing and communication, which directly support the execution of cognitive tasks (Park et al. 2004; Rajah and D’Esposito 2005; Spreng et al. 2010). To gain a deeper insight into the processes involved in brain aging, the relationships between age, brain functional metrics, and cognition need to be better understood. One potential approach is to quantify the manner in which aging effects on cognition are modulated by brain functional activities.

Blood oxygenation level-dependent (BOLD) signals measured by functional magnetic resonance imaging (fMRI) provide a basis for mapping neural activities and have enhanced our comprehension of functional changes in the aging brain (Ota and Shah 2022). While extensive studies have been reported of BOLD signals in gray matter (GM), their presence and significance in white matter (WM) have been largely ignored. This has led to an incomplete understanding of if, or how, measures of WM functions are impacted by aging. Nonetheless, based on our findings and other contemporary research, it is evident that although BOLD signals in WM are weaker, they can be reliably detected with appropriate analytical tools (D’Arcy et al. 2006; Fraser et al. 2012; Ding et al. 2013, 2018; Gawryluk et al. 2014; Ji et al. 2017; Peer et al. 2017; Courtemanche et al. 2018; Gore et al. 2019; Li et al. 2019). These signals are modulated by a spectrum of neural activities, such as those induced by differing anesthesia levels, and may be evoked by different stimuli (Wu et al. 2016; Li et al. 2020; Mishra et al. 2020), so similar to cortical BOLD signals they are related to neural activity. In addition, WM BOLD signals in a resting state without any external task or stimuli also manifest signal fluctuations that are notably altered in individuals with neurological or psychiatric conditions (Gao et al. 2020; Huang J et al. 2020a; Lin et al. 2020; Ji et al. 2023). Importantly, our recent study demonstrates that resting-state BOLD signals in WM undergo changes throughout late adulthood (Li et al. 2023). These and other findings suggest the associations between age, WM function, and cognition should be further explored.

Our primary hypotheses behind this study are: (i) that age affects human cognition partly through its impact on the functional characteristics of WM and (ii) that age differentially impacts different cognitive domains by modifying the functional characteristics of specific WM regions. To probe these postulates, mediation analysis was employed to elucidate those WM functional properties that propagate the influence of age on cognitive abilities. This method has been extensively applied in the field of neuroscience, yielding significant insights by identifying imaging biomarkers that serve as mediators of the association between 2 distinct variables. For example, Won et al. revealed that basal forebrain connectivity is a mediator in the correlation between cardiorespiratory fitness and cognitive performance among healthy older women (Won et al. 2023). Similarly, Lu et al. found that myelin breakdown mediates age-related slowing of cognitive processing speed in healthy elderly men (Lu et al. 2013). Importantly, the work of Wager et al. has contributed a specialized mediation toolbox designed for brain imaging studies, available at https://github.com/canlab/MediationToolbox, which has been instrumental in identifying neural mediators of cardiovascular reactions to social threats (Wager, van Ast, et al. 2009; Wager, Waugh, et al. 2009). However, the literature contains sparse examples in which brain function is posited as a mediator of age-cognition relationships. Limited studies, such as that by Riva et al. (2022), who identified the connectivity between the right supramarginal gyrus and somatosensory cortices as a partial mediator in the link between age and emotional egocentric bias (Riva et al. 2022), and by Schulz et al., who reported the mediating role of the default network’s connectivity in the relationship between age and executive cognitive function (Schulz et al. 2022), highlight the potential for further exploration. As the majority of functional imaging studies focus on GM only, a notable gap exists in the literature concerning the specific mechanisms by which WM function mediates the impact of age on cognition, underscoring the novelty and necessity of our approach.

We aimed to identify specific WM regions and functional properties within those WM regions that are susceptible to age-related changes, and to quantify how these alterations in turn influence cognitive function. Employing an independent component analysis (ICA) methodology, we segment the WM into spatially independent components (ICs), each made up of voxels that show similar temporal variations (Huang Y et al. 2020b; Li et al. 2023). This stratification permits an in-depth examination of temporal synchronizations among ICs, potentially revealing complex subtleties of neural communication and network structures (Ji et al. 2019). In our analysis, we constructed a graph representing the pairwise connectivities among these ICs, thereby modeling WM as a complex network. This model facilitated the derivation of graph-theoretical metrics such as clustering coefficients, efficiency, and nodal strength, providing a quantitative framework to investigate regional functional properties (Wang 2010). We then applied mediation analysis to systematically search across these regions and their properties, thereby deriving statistically significant mediators that illuminate the pathways through which aging influences cognition. We observed significant influences of WM characteristics on the relationship between age and a set of cognitive scores. We also discovered that aging impacts different cognitive abilities by modifying specific WM regions and their respective functional attributes. These results indicate that different cognitive domains are differentially affected by age-related changes in the functional characteristics of the WM of the brain.

Materials and methods

Data

A cohort of 725 healthy participants, representative of “typical” aging trajectories, was sourced from the Lifespan Human Connectome Project in Aging 2.0 Release (HCP-A) database (Harms et al. 2018; Bookheimer et al. 2019). The term “typical” is in relation to individuals who exhibit typical health for their age in the absence of identified pathological causes of cognitive decline (e.g. stroke, clinical dementia). Of these, 688 participants with complete fMRI datasets and physiological recordings (cardiac and respiratory fluctuations) were included in the analysis, comprising 303 males and 385 females, ranging in age from 36 to 100 yr.

The experimental protocol produced a multi-modal set of images: structural MRI, task-based fMRI, resting-state fMRI, diffusion imaging, and measures of cerebral blood flow via arterial spin labeling (ASL), all conducted across 2 separate imaging sessions. For the purposes of this investigation, we confined our analysis to the structural and resting-state fMRI data acquired in the first session. Detailed methodological descriptions of these imaging protocols can be found in (Harms et al. 2018). Briefly, imaging was performed using Siemens 3T Prisma scanners, each equipped with a 32-channel head coil. The resting-state fMRI acquisition consisted of 2 6-min-41-s runs (left-to-right and right-to-left phase encoding), with the following parameters: a repetition time (TR) of 800 ms, an echo time (TE) of 37 ms, an isotropic voxel size of 2 mm, and a total of 488 volumes per run. During the scans, participants were directed to fixate on a white crosshair against a black backdrop, remain motionless, awake, and blink naturally. Physiological data, including respiratory and cardiac fluctuations, were recorded throughout fMRI scanning. T1-weighted images were obtained with a multi-echo magnetization-prepared rapid gradient echo (MPRAGE) sequence with a TR of 2500 ms, multiple TEs of 1.8/3.6/5.4/7.2 ms, and an isotropic voxel size of 0.8 mm.

Behavioral tests

The subject tests included a battery of cognitive and physical performance assessments that are particularly pertinent to aging. These evaluations targeted episodic memory, motor speed, sensory perception (including pain tolerance, as well as auditory and visual sharpness), and physical robustness. Within this spectrum of assessments, our study concentrated on 5 cognitive and, for comparison, 1 sensory measure from the NIH toolbox: the Picture Sequence Memory test (PSM) for episodic memory, the Dimensional Change Card Sort (DCCS) test for cognitive flexibility, the Flanker Task for attentional control and inhibitory capability, the Pattern Completion Processing Speed (PCPS) test for processing speed, the List Sorting Working Memory (LSWM) test for working memory capacity, and the visual acuity test for sensory assessment (Hodes et al. 2013). The HCP database also provides a Cognition Composite Score derived by averaging the normalized scores of each of the 5 cognitive measurements introduced above.

Preprocessing

As this study focuses on WM signals, we did not download data that were proposed through ICA-FIX, which has regressed out WM signals (Salimi-Khorshidi et al. 2014). Instead, we used “uncleaned” images that underwent only minimal preprocessing pipelines (MPP) (Glasser et al. 2013). Briefly, T1-weighted images were nonlinearly aligned to the MNI space using FNIRT (Jenkinson et al. 2012) and subsequently processed through the Freesurfer pipeline. This yielded volume and surface parcellations in conjunction with a T1w-to-MNI mapping field (Dale et al. 1999). For fMRI, the processing included head motion removal, distortion correction from susceptibility effects using the FSL software (Jenkinson et al. 2012) based on the 2 runs of data with reversed-phase encoding directions, and linear registration to T1 and then non-linear registration to the MNI space. We carried out additional processing, including regression of confounding variables, including 12 parameters related to head movement (3 translations, 3 rotations, and their respective derivatives), as well as respiratory and cardiac variances modeled by the RETROICOR algorithm (Glover et al. 2000). This was succeeded by correction for linear trends and temporal filtration using a band-pass filter with a range of 0.01–0.1 Hz. A group WM mask was generated by averaging the Freesurfer-derived WM parcellations (transformed to MNI space) from all participants and applying a threshold of 0.95, ensuring computations were confined to only WM regions. After that, the fMRI data within the WM mask were spatially smoothed with a 4-mm full width at half maximum (FWHM) Gaussian kernel.

Group ICA

Spatio-temporal data were decomposed into spatial components by ICA. These ICs serve as a basis set that represents the original data with lower dimensions, following an undefined linear mixing procedure. In our study, we used Group ICA of the FMRI Toolbox (GIFT) as described by (Calhoun et al. 2001). While several toolbox parameters were retained at default values, we changed the mask option to confine the ICA calculation within the WM mask we created and adjusted the number of ICs and principal components (PCs). Component numbers were estimated from the fMRI data using the MDL criteria (Calhoun et al. 2001) and determined to be 72 based on our data. Initiating the Group ICA process, the temporal dimension for each participant was reduced from 488 to 108 (1.5× the anticipated IC numbers) via spatial principal component analysis (PCA). These PCs were then temporally concatenated across all subjects, resulting in a combined time course of 688 × 108 (= #subjects × #PCs) matrix for every voxel. These group data were subjected to a second round of PCA, reducing the dimension from 108 to 72. This procedure generated PCs that accounted for the maximal group-level variances. Subsequently, 72 ICs were derived from these PCs using the Infomax algorithm (Bell and Sejnowski 1995). The spatial maps for each IC, at the group scale, were reconstructed, converted to z-scores, and subjected to a threshold of z > 2. It is important to mention that the z-score, in this context, is primarily descriptive and does not assert any statistical significance (Mckeown et al. 1998). Finally, these ICs were superimposed as masks on individual fMRI datasets to extract the average time series of interest. Based on these, functional networks were established by assessing pair-wise correlations between the time series of the 72 ICs, as elaborated in subsequent sections.

Network measurements

Functional connectivity (FC) matrices were formulated by calculating Pearson’s correlation coefficients between the time courses of paired ICs for each participant. We derived 2 primary types of network metrics: the within-IC FC of each IC and graph-theoretical metrics that represent the global and/or local functional properties of the network. The within-IC FC is obtained from the mean z-score derived from the group ICA, indicating the synchrony of BOLD fluctuations within each IC. For network architecture, we employed the Brain Connectivity Toolbox (Rubinov and Sporns 2010) to produce 4 important graph-theoretical measures:

1. Global characteristic path length: This represents the average minimum pathway length throughout the network, indicating the efficiency of the entire brain. A reduced path length signifies expedited information transfer and optimized network functionality. A single measurement of the global characteristic path length is associated with the neural network of an individual.

2. Local clustering coefficients: This metric assesses the fraction of triangular connections surrounding an IC, equivalent to how frequently a given IC’s neighbors are interconnected. It provides insights into the propensity of ICs to form cohesive clusters, which is instrumental in discerning a network’s small-world nature. A local clustering coefficient was measured for each IC, so there are 72 measurements for each individual.

3. Local efficiency: This quantifies the capacity for efficient information exchange among an IC’s neighbors in its absence. In essence, it measures how seamlessly a network can adapt and maintain connectivity when an IC is excluded. A local efficiency was measured from each IC so that there are 72 measurements for each individual.

4. Local strength: This denotes the sum of the weights of connections associated with an IC, illuminating its influence or centrality within the overall network. A local strength was measured from each IC so that there are 72 measurements for each individual.

It should be emphasized that these metrics are derived from the weighted FC matrix, and only weights >0.2 were retained for computation to increase the signal-to-noise ratio.

Mediation analysis

Mediation analysis serves to elucidate pathways that link variables, which may have different units or features. Whereas a direct association between 2 variables (x and y) can be depicted as [x -> y], mediation introduces an intermediate variable, m, forming a sequence [x -> m -> y]. In this framework, x is the initial variable and y is the outcome or dependent variable, the primary endpoint. m, the mediator, is important in transmitting the effect of x on y. It is essentially a bridge in the causal pathway from x to y. In this study, we hypothesize that the impact of age (x) on cognitive ability (y) is mediated by certain functional properties in specific WM regions. To quantitatively assess mediation, 2 linear regression equations are employed. Their slopes, namely path coefficients, represent the strength of associations. As shown in Fig. 1, path coefficient a represents the effect of x on m. The path coefficient b represents the effect of m on y. The path coefficient c’ reflects the direct effect of x on y when we control for the mediator m. The total relationship c represents the effect of x on y when we do not control for the mediator m. The essence of mediation analysis lies in discerning if c-c’ equals zero. This compares the effect of x on y across 2 scenarios: one where the mediator can vary (or can be ignored), and another where it is statistically accounted for (or held constant). Through algebraic reconfiguration, it is evident that c-c’ = a*b. Consequently, the mediation effect is given by the product of the coefficients a and b. Bootstrapped distributions yield confidence intervals and P-values for the mediation effect, aiding in determining its significance. In our study, bootstrapping (1,000 samples) was performed for every candidate mediator, producing a collection of P-values that were then corrected for multiple comparisons using the false discovery rate (FDR) approach. Note that in mediation analysis, we used the young index (calculated by subtracting the age of each individual from the maximum age found in the group) as a substitute for age. Our objective was to consistently arrange all variables in an order that reflects a progression toward healthier or more optimal states, for clarity in interpretation. For instance, behavioral performance metrics range from low to high, and network metrics follow a similar low-to-high progression.

Fig. 1.

Experiment hypothesis. This mediation analysis investigates whether the impact of age on cognitive abilities is channeled through its effects on the WM function, as indicated by functional network metrics derived from BOLD measurements in fMRI.

Results

Relationship between age and 6 different clinical scores

Figure 2 illustrates the relationship between age and 6 different clinical scores, including 5 cognitive scores (Flanker, DCCS, PCPS, LSWM, and PSM) and 1 sensory score (visual acuity test). Each scatter plot consistently indicates a negative correlation, signifying a decline in scores as age increases, despite varying in their specific cognitive or sensory domains. The PCPS score presents the strongest negative relationship with age, closely followed by the visual acuity score.

Fig. 2.

Relationship between age (in mo) and clinical scores, including 5 cognitive scores—Flanker, DCCS, PCPS, LSWM, and PSM—and 1 sensory score from the visual acuity test. The correlation coefficient (r) and the P-value are displayed above each scatter plot.

ICs derived from ICA, and correlations within and between them

The spatial maps of the 72 ICs are depicted in Fig. 3a, which shows the spatially distinct clusters in the brain, often manifesting with bilateral symmetry. The ICs are located throughout the brain, and their anatomical descriptions can be found in the Supplementary Materials. The within-IC FC shown in Fig. 3b represents the average correlation between the mean signal and the signals of all voxels within a particular IC. The inter-IC correlation matrix presented in Fig. 3c captures the relationships between the different ICs. Together, these data provide the basis for our subsequent mediation analysis.

Fig. 3.

Visualization of ICs, and the functional connectivity within and between them. a) Spatial maps of the 72 ICs. b) An exemplification of within-IC functional connectivity, detailing the average correlation between the mean signal and signals of all voxels within a particular IC. c) The inter-IC correlation matrix showcases synchronization between different ICs, further producing network metrics important for mediation analysis.

Correlation among age (young index), within-IC FC, and behavioral scores

Figure 4 shows the significant correlations (P < 0.05, FDR corrected) between age (represented by the young index), within-IC FC, and a set of 6 behavioral scores. We observed a consistent positive correlation between the young index and within-IC FC among the 72 ICs, reinforcing the prominent influence of age on functional homogeneities in the ICs. The Flanker and visual panels stand out, with the presence of right-aligned blue bars indicating positive correlations between within-IC FC and respective behavioral scores. For a functional measurement to be considered a potential mediator between age and behavior, 2 criteria must be met. First, there should be a significant correlation between that measurement and age, and second, it must also exhibit a significant correlation with behavior. In other words, only those ICs that concurrently display both blue and red bars can be deemed as candidate mediators. In the results, the absence of blue bars in the DCCS, PCPS, LSWM, and PSM panels suggests a lack of significant correlation between the within-IC FC and the respective behavioral scores for these categories, precluding the application of mediation analysis for these specific behavioral scores.

Fig. 4.

The correlation coefficients between age, within-IC FC, and diverse behavioral scores across 6 distinct panels. Only significant correlations (P < 0.05, FDR correction) are shown. Each panel is uniquely aligned with a behavioral score: Flanker, DCCS, PCPS, LSWM, PSM, and Visual. Within each panel, the x-axis indicates the correlation coefficients, the y-axis indicates the IC index, and the bars depict the relationship of the 72 ICs with age and the specified behavioral score highlighted above. It’s important to note the utilization of the young index to represent age, ensuring clarity in interpretation. The left-aligned red bars elucidate the correlation of age with the within-IC FC for the 72 ICs. In contrast, the right-aligned blue bars illustrate the correlation between a designated behavioral score and the within-IC FC for these ICs. Instances where red bars appear on the right and blue bars on the left signal negative correlations.

It is also important to note that the distribution of red bars should remain largely consistent across all 6 panels, as the age and the within-IC FC for the 72 ICs, as well as their correlations, should be invariant across all 6 panels. However, subtle variations may be observed due to differences in the sample groups for each panel. This is because some individuals might lack specific behavioral test data, and the study samples regarding each panel are not completely the same, resulting in minor discrepancies in the distributions.

Mediation effects of within-IC FC

Figure 5 presents the mediation effects of within-IC FC on the association between age and specific behavioral scores. In panel (a), the mediating influence of within-IC FC between age (designated as the young index) and the Flanker behavioral score is depicted. Notably, significant mediation is observed in ICs 8, 9, 10, 11, 14, 22, 23, 26, 27, 32, 34, 42, 43, 61, 63, 67, and 70. These regions were observed bilaterally across the occipital, temporal, and parietal lobes, mainly along the superior longitudinal fasciculus, posterior thalamic radiation, superior and posterior corona radiata, sagittal stratum, the splenium of the corpus callosum, and the retrolenticular part of the internal capsule. The ICs that show the strongest mediation are ICs 22, 23, and 42, which are all located along the bilateral sagittal stratum (inferior longitudinal fasciculus and inferior fronto-occipital fasciculus) and the left posterior thalamic radiation proximal to the fusiform gyrus, superior temporal gyrus, and inferior temporal gyrus. Panel (b) shows the mediation effect of within-IC FC between age (young index) and the visual acuity score. Correspondingly, the radar chart emphasizes the significant mediation brought about by ICs 42 and 67, which are constrained to the right hemisphere in the temporal and occipital areas, respectively, along the right posterior thalamic radiation, optic radiation, sagittal stratum, and the superior longitudinal fasciculus.

Fig. 5.

Mediation effects of within-IC FC on the relationship between age and behavioral scores. Panel a) depicts the mediating influence of within-IC FC between age (young index) and the Flanker behavioral score. The 3D visualization at the top highlights the distribution of ICs with significant mediation effects. The radar chart on the right provides a quantitative visualization of the mediation effects across the ICs, with longer radiations symbolizing stronger mediation effects. Panel b) shows the mediation of within-IC FC between age (young index) and the visual behavioral score. As before, the distributions of significant ICs are displayed on the left, with the radar chart detailing the mediation magnitudes.

Correlation among age (young index), network metrics, and behavior scores

Figure 6 illustrates the significant correlations (P < 0.05, FDR correction) between age, symbolized by the young index, network metrics, and an assortment of behavioral scores. Specific panels, like Flanker, DCCS, and visual, are particularly noteworthy due to the presence of right-aligned blue bars. Conversely, certain panels regarding PCPS, LSWM, and PSM are devoid of blue bars, highlighting the lack of significant mediators in terms of network metrics for specific ICs.

Fig. 6.

A detailed representation of the significant correlations between age (substituted by the young index), network metrics, and 6 distinct behavioral scores across 18 panels. Each column corresponds to a specific behavioral score, while each row signifies a network metric. Within each panel, the x-axis indicates the correlation coefficients and the y-axis indicates the IC index. The relationship between each network metric of 72 ICs, age, and each behavioral score is depicted in every panel. Red bars on the left denote the correlation of age with the metrics for the 72 ICs, whereas blue bars on the right show the correlation with specific behavioral scores. Negative correlations are indicated when red and blue bars swap positions.

Mediation effects of network metrics

Figure 7 shows the mediation effects of various network metrics on the association between age and behavioral scores. In panel (a), the mediating influences of 3 network metrics, clustering coefficient, efficiency, and strength, on the relationship between age (represented as the young index) and the Flanker behavioral score are illustrated. Given the numerous significant ICs, the effects can be described as comprehensive. ICs 22, 23, and 27 exhibit the overall strongest mediation effects, located along the bilateral sagittal stratum that includes the inferior longitudinal fasciculus and right inferior fronto-occipital fasciculus in the posterior temporal and occipital lobes, as well as the superior longitudinal fasciculus proximal to the inferior precentral gyrus and posterior insula.

Fig. 7.

Mediation effects of network metrics on the relationship between age and behavioral scores. Panel a) depicts the mediating influence of the 3 network metrics between age (young index) and the Flanker behavioral score. 3D visualization highlights ICs with significant mediation effects. The radar chart on the right provides a detailed view of the mediation effects across the ICs, with longer radiations symbolizing stronger mediation effects. In panel b), the focus shifts to the mediation of the 3 network metrics between age (young index) and the visual behavioral score. As before, significant ICs are displayed on the left, with the radar chart detailing the mediation magnitudes. c) Illustrate the mediation of the network strength between age (young index) and the visual behavioral score.

Panel (b) focuses on the mediation effects of the same 3 network metrics on the relationship between age (young index) and the DCCS behavioral score. The significant ICs for the clustering coefficient include IC 23, 38, 50, 61, and 67. For efficiency, the significant ICs are 22, 38, 64, and 67, which are regionally similar to the significant ICs for the clustering coefficient. Combined, these regions include the bilateral forceps, majorly connected by the splenium of the corpus callosum, the inferior longitudinal fasciculus, and inferior fronto-occipital fasciculus, extending from the occipital to parietal and temporal lobes. The significant regions extend more anteriorly along the temporal lobe in the right hemisphere for clustering coefficient and the left hemisphere for efficiency. There is also an isolated cluster in the superior longitudinal fasciculus proximal to the middle cingulate gyrus and posterior middle frontal gyrus in the right hemisphere only. Moreover, IC 58 stands out for the strength metric, located in the posterior corona radiata, located proximal to the inferior precuneus and posterior cingulate gyrus. In general, IC 22 and 23, located along the bilateral sagittal stratum that includes the inferior longitudinal fasciculus and right inferior fronto-occipital fasciculus in the posterior temporal and occipital lobes, exhibit the strongest mediation effects. Finally, panel (c) emphasizes the mediation of the network strength between age (young index) and the visual behavioral score. The significant ICs in this mediation process are IC 23, 34, 42, 56, and 67. This covers the superior and inferior longitudinal fasciculus, as well as the inferior fronto-occipital fasciculus and forceps, major extending from the occipital lobe to the middle of the temporal lobe and to the angular gyrus. It also includes an isolated region in the right anterior thalamic radiation and forceps minor proximal to the anterior portions of the superior and middle frontal gyri. IC 56 and 67, located in the right frontal and occipital lobes, exhibit the overall strongest mediation effects.

Mediation effects of global network metrics

In our investigation of the potential mediating role of global network metrics in the relationships between age and cognitive composite scores, we focused on the characteristic path length as the representative metric. As shown in Fig. 8, although a pronounced correlation was observed between age and the cognitive composite score (r = −0.34, P = 7.75e-17), the mediation effect via the characteristic path length was not substantiated. Specifically, the connection between this global network metric and the cognitive composite score was not found to be significant (P = 0.39).

Fig. 8.

An exploration of the mediating role of global network metrics, specifically the characteristic path length, in the connection between age and cognitive composite scores. The mediation through the characteristic path length is not evident, as evidenced by the nonsignificant correlation between the global network metric and the cognitive composite score, in spite of the figure shown below suggesting a significant correlation with the cognitive composite score.

Discussion

We aimed to explore the effects of aging on brain functions and human cognition, hypothesizing that age-related changes in WM function are crucial mediators of cognitive performance. Our findings substantiate this hypothesis, demonstrating that as age increases, cognitive performance declines, particularly in memory, executive function, processing speed, and visual ability. This cognitive decline is closely associated with decreased functional properties within and across different WM regions, suggesting that age-related changes in WM function are indeed a significant factor affecting cognition. Moreover, our results support the second hypothesis that aging has a differential impact on various cognitive domains; we found that specific WM regions exhibit unique functional alterations corresponding to their cognitive domain, thereby confirming that the age-related cognitive decline is not uniform but is modulated by distinct functional characteristics of particular WM areas.

The within-IC FC signifies the degree of synchronous neural activity within specific brain regions. Reduced within-IC FC has been linked to declines in cognitive performance and is characteristic of the aging brain (Andrews-Hanna et al. 2007). The Flanker task, a measure of attention, executive function, and response inhibition, activates an extensive brain network responsible for attentional control and higher executive processes. Two important regions are known to be associated with this task, including the dorsolateral prefrontal cortex (DLPFC), which facilitates executive functions, such as working memory and strategic planning, and the anterior cingulate cortex (ACC), which plays a role in error monitoring, attentional control, and the modulation of cognitive and emotional processing (Casey et al. 2000). Other regions, such as the parietal, temporal, premotor, and visual cortex, are also considered associated with the Flanker task. In our investigation, we identified the ICs in the frontal lobe as significant mediators between age and Flanker performance, which is consistent with previous findings. However, we observed that the ICs in the temporal lobe exhibit the strongest mediation effects, specifically the posterior thalamic radiation and bilateral sagittal stratum, which includes the inferior longitudinal fasciculus and the inferior fronto-occipital fasciculus, running along the length of the temporal lobe and extending through the temporal stem. The temporal lobe modulates attention, especially in a flexible task-dependent manner (Rusnáková et al. 2011). The WM pathways of the temporal lobe play an important role in linking the temporal and occipital lobes to the frontal lobe within the temporal cortex attention network, and their altered FC may reflect broader neural network changes that detrimentally impact cognitive functioning (Ramezanpour and Fallah 2022). Literature, such as (Sarabin et al. 2023), illuminates the substantial relationship between cortical thickness in temporal areas and Flanker task performance, particularly evident in comparative studies of ADHD and healthy control groups. This body of work suggests that ADHD may induce neural loss within the temporal lobes, thus diminishing Flanker’s task performance. Analogously, we suggest that aging may similarly impair temporal lobe function, potentially related to neural attrition, which in turn could affect the performance of cognitive tasks dependent on these regions. For the visual score, ICs in occipital and temporal areas, which are respectively responsible for basic and high-level visual processing, show a mediation effect, possibly indicating that age-related changes in visual network connectivity influence visual ability. The implication here is that visual processing, a critical aspect of cognitive performance, is not immune to the effects of aging, with alterations in within-IC FC signaling potential disruptions in how visual information is processed and interpreted.

The mediating influence of network metrics on the association between age and cognitive performance reflects a nuanced interplay between neurobiological changes and behavioral outcomes. Performance on the Flanker task appears to be influenced by age-related changes in the network metrics of clustering coefficient, efficiency, and strength, particularly within the temporal and frontal lobes, suggesting that these regions become less optimally configured with age. Such reconfiguration could implicate a decline in the local connectivity and overall network integration necessary for the rapid response inhibition and attentional control required by the Flanker task. The frontal lobe is well-established as a mediator in tasks requiring executive control, such as the Flanker task (Miller and Cohen 2001). Age-related declines in frontal lobe function, as evidenced by alterations in network metrics like clustering coefficient and efficiency, could directly affect the cognitive control required to perform the Flanker task effectively. In addition, the role of the temporal lobe in cognition is not just about traditional sensory processing but also executive functions and the ability to quickly access memory stores that can influence performance on Flanker tasks (Rusnáková et al. 2011; Grady 2012). In the context of the DCCS task, which measures cognitive flexibility, the study highlights the 3 network metrics in different brain regions, particularly temporal areas, that show age-related mediation effects and point to a distributed pattern of mediation. These results may reflect the necessity for both localized and network-wide efficient processing to perform the DCCS task successfully, supporting theories that emphasize the importance of network integration in executive functions. Several frontal and parietal ICs were identified as significant mediators, which are known to be associated with executive functions (Miyake et al. 2000; Leber et al. 2008). The involvement of the temporal ICs, which are usually recognized for their roles in language and auditory processing as well as cognitive flexibility (Ramos et al. 2013), underlines their significance in the impact of aging on the performance of DCCS. For the visual behavioral score, the mediation analysis reveals a prominent role of network strength, particularly in the frontal and occipital lobes. The strength of connections in these regions is essential for visual processing, and the findings suggest that the robustness of these connections is compromised with age, which may contribute to a decline in visual abilities. These results are in line with the understanding that the occipital lobe as well as frontal areas are important for visual processing and that age-related changes in their connectivity can influence visual perception and related cognitive functions (Barceló et al. 2000; Li et al. 2001; Owsley 2011).

We have pinpointed several regions that are recurrently recognized as significant mediators between aging and various cognitive functions, particularly within the temporal and occipital lobes. The temporal lobe, extensively documented as susceptible to the aging process, may offer an explanation for these findings. Previous research has consistently shown that the temporal lobe undergoes significant age-related alterations, notably characterized by a reduction in cortical volume (Bartzokis et al. 2001; Raz et al. 2005). Further studies using diffusion MRI techniques have reported a decline in WM integrity within these regions (Gunning-Dixon et al. 2009). This suggests that neuronal loss, along with a reduction in myelination, could correlate with a decreased need for neural communication. Given the networked nature of the brain, impairment in any single node can potentially reduce the functional capacity of the entire network, adversely influencing cognitive performance across various domains. In the case of the occipital regions, the rationale may lie in their integral role in visual processing. Most cognitive tasks engage visual functions to some degree, and as the function of the visual region declines with age, this could correspondingly impact performance across multiple cognitive domains.

In our findings, a greater number of network metrics have been identified as significant mediators compared to within-IC FC. This pattern may indicate that the mediation processes predominantly operate at the network level, potentially relating to the architecture of the network rather than the functional homogeneity within each functional region. Notably, despite the emphasis on network-level mediation, the global network metric, namely, characteristic path length, did not exhibit a significant mediating effect. This suggests that the mediating influence is specific to localized networks. Thus, the configuration of particular regions is more crucial in mediating the effects of age on specific behavioral outcomes than a broad global metric influencing overall cognitive decline.

Conclusion

In conclusion, our study offers valuable insights into the aging process and its effects on human cognition. Through mediation analysis, we have substantiated the hypothesis that cognitive decline with age is mediated by alterations in WM functional properties. Our investigation further confirms the hypothesis that the impact of aging on cognition is not homogeneous. Specific WM regions are associated with unique functional changes that correlate with the cognitive domains they influence, illustrating the differential effects of aging across various cognitive functions. By illuminating the complex interplay between age, WM function, and cognition, our work advances our knowledge of the aging brain’s functional interactions and of the critical role played by the neurovascular response of WM.

Author contributions

Muwei Li (Conceptualization, Formal analysis, Investigation, Methodology, Software, Validation, Visualization, Writing—original draft), Kurt Schilling (Formal analysis, Investigation), Fei Gao (Conceptualization, Methodology), Lyuan Xu (Formal analysis, Investigation), Soyoung Choi (Formal analysis, Investigation), Yurui Gao (Formal analysis, Investigation), Zhongliang Zu (Formal analysis, Investigation), Adam Anderson (Supervision), Zhaohua Ding (Supervision, Writing—review & editing), Bennett Landman (Funding acquisition, Supervision, Writing—review & editing), and John C. Gore (Conceptualization, Funding acquisition, Supervision, Writing—review & editing).

Funding

This work was supported by the National Institutes of Health (NIH) grants RF1 MH123201 (JCG and BAL), R01 NS113832 (JCG), R01 NS129855 (ZD), T32 EB001628 (SC), and K01 EB032898 (KGS).

 

Conflict of interest statement: None declared.