Decreased segregation of brain systems across the healthy adult lifespan

Significance

The brain is a large-scale network, not unlike many social or technological networks. Just like social networks, brain networks contain subnetworks or systems of highly related or interacting nodes (in the case of brains, nodes may represent neurons or brain areas). Using functional MRI to measure functional correlations between brain areas during periods of rest, we describe differences in brain network organization in a large group of individuals sampled across the healthy adult lifespan (20–89 y). We characterize a measure of system segregation, reflecting the degree to which the systems share connections among one another. Increasing age is accompanied by decreasing segregation of brain systems. Importantly, system segregation is predictive of measures of long-term memory function, independent of age.

Abstract

Healthy aging has been associated with decreased specialization in brain function. This characterization has focused largely on describing age-accompanied differences in specialization at the level of neurons and brain areas. We expand this work to describe systems-level differences in specialization in a healthy adult lifespan sample (n = 210; 20–89 y). A graph-theoretic framework is used to guide analysis of functional MRI resting-state data and describe systems-level differences in connectivity of individual brain networks. Young adults’ brain systems exhibit a balance of within- and between-system correlations that is characteristic of segregated and specialized organization. Increasing age is accompanied by decreasing segregation of brain systems. Compared with systems involved in the processing of sensory input and motor output, systems mediating “associative” operations exhibit a distinct pattern of reductions in segregation across the adult lifespan. Of particular importance, the magnitude of association system segregation is predictive of long-term memory function, independent of an individual’s age.

Healthy adult aging is characterized by a progressive degradation of brain structure and function associated with gradual changes in cognition (see reviews in refs. 1, 2). Among the age-accompanied functional changes, one prominent observation is a reduction in the specificity with which distinct neural structures mediate particular processing roles [i.e., a reduction in functional specialization, or “dedifferentiation” (3)]. A reduction in functional specificity has been observed across multiple spatial scales of brain organization, ranging from the firing patterns of single neurons (e.g., refs. 4, 5) to the evoked activity of individual brain areas (6–10). (For additional discussion see ref. 11.)

Despite the compelling evidence for age-accompanied reductions in functional specialization across numerous brain areas, the relationship between neural specialization and cognition generally is weak. This likely is related to the fact that broad cognitive domains such as “long-term memory” and “executive control” are mediated by distributed and interacting brain systems, each consisting of multiple interacting brain areas. Thus, relating functional specialization in a single brain area to general measures of cognition likely will be unsuccessful. Such an argument is consistent with the view that severe impairment in cognitive function due to injury or insult typically is a consequence of damage to multiple brain locations (e.g., refs. 12, 13). Based on these considerations, it seems plausible that the cognitive decline evident even in healthy older adults may be related to decreased functional integrity at a systems level of organization. The present report approaches healthy aging from this systems-level perspective in an effort to relate systems-related functional specialization to age-accompanied differences in cognition.

Before proceeding, it is important to clarify the meaning of system. The term “system” often is used in relation to brain organization when referring to any group of areas that subserve common processing roles. For example, the visual system comprises brain areas primarily defined by their role in processing visual stimuli (e.g., ref. 14), and the frontal–parietal task control system consists of brain areas involved mainly in adaptive task control (15). Identifying distinct brain systems and defining their functional roles by examining how their constituent areas are modulated by experimental testing or are impaired by brain damage is not an easy endeavor; systems of brain areas typically mediate processing roles that span multiple stimulus and task demands. This reality makes assessing changes in the functional specialization of systems across cohorts of individuals extremely challenging.

An alternative formal and complementary approach to defining a brain system involves understanding how brain areas relate to one another via their patterns of shared functional or anatomical relationships in the context of a large-scale network (16, 17). Like many other complex networks, brain networks may be analyzed as a graph of connected or interacting elements. When a brain network graph represents the interaction of areas, one prominent feature is the presence of subsets of areas that are highly interactive with one another and less interactive with other subsets of areas. This pattern of organization reflects the presence of distinct “modules” or “communities” (e.g., ref. 18). Importantly, numerous connectivity-defined human brain modules have been shown to overlap closely with functional systems as defined by other methods of assessing information processing [e.g., task-evoked activity, lesion-mapping (19, 20)]. The close correspondence between differing methods of system identification provides a basis for using connectivity to understand the organization of brain systems and how they may differ with age.

Modular brain networks are characterized by a fine balance of dense within-system relationships among brain areas (nodes) that have highly related processing roles, as well as sparser (but not necessarily absent) relationships between areas in systems with divergent processing roles. This pattern of system segregation facilitates communication among brain areas that may be distributed anatomically but nevertheless are in the service of related sets of processing operations, and simultaneously reinforces the functional specialization of systems that perform different sets of processing operations (21). Importantly, segregated systems can communicate with one another via the presence of the sparser connections between them. As such, any deviation in the patterns of within- and between-system connectivity may be considered evidence for a change in the system’s specialization. Furthermore, if aging is associated with changes in functional specialization at the level of brain systems, this may be revealed by examining the differences in patterns of within- and between-system areal connectivity across age.

We use functional connectivity, as measured by blood oxygen-level–dependent (BOLD) functional MRI (fMRI) during rest [resting-state functional correlations (RSFCs), see ref. 22], to assess age-related differences in the organization of brain systems. Changes in RSFC patterns between sets of areas have been observed following extensive directed training (23–25), and differences in RSFC patterns also have been reported in cross-sectional comparisons spanning from childhood to older age (e.g., refs. 26–29). The extant data suggest that RSFCs are malleable and reflect sensitivity to a history of coactivation: changes in the processing roles of areas may be characterized by changes in their RSFCs with other areas (for discussion, see ref. 17). This feature makes RSFCs particularly useful in assessing differences in the organization and specialization of brain systems.

In the present study, the age-accompanied differences in the functional specialization of brain systems are revealed by examining patterns of within- and between-system areal RSFCs in a large healthy adult lifespan sample (n = 210; age range, 20–89 y). The inclusion of subjects distributed across each decade of adulthood not only allows us to assess how older and younger adults differ in their organization of brain systems, but also provides insight as to whether there is a critical stage of the adult lifespan when differences in system organization may appear. Previous reports attempted to address related questions by examining end points of the adult aging spectrum, focusing on the organization within specific systems (e.g., refs. 26, 28, 30), or using area nodes that are not representative of functional areas [e.g., structural parcels (31–34)]. The latter feature likely contributes to the inconsistent findings observed in the relationship between summary network measures and age groups (e.g., refs. 31, 35 vs. refs. 30, 36). In addition to examining age-related differences in system organization developed from a biologically plausible cortical parcellation of the human brain network, we also relate systems-level differences in organization to differences in general measures of cognitive ability. To foreshadow the results that follow, we report that aging is associated with differences in patterns of connectivity within and between brain systems, that these differences are not uniform across all systems, and that the segregation of brain systems has a direct relationship to measures of cognitive ability independent of age.

Results

Defining Biologically Plausible Brain Network Nodes Using RSFC-Based Area Parcellation.

Understanding the organization of brain systems and how areas within and between brain systems interact with one another mandates the identification of a set of nodes representing areas in the overall brain network. The identified nodes must be biologically constrained and represent meaningful areal distinctions (for discussion, see ref. 17). Network nodes were constructed by processing further (Experimental Procedures, Area Node Definition) a recently published RSFC-based area parcellation map. This parcellation map identified cortical locations where patterns of RSFCs exhibited abrupt transitions in a large group of subjects [i.e., putative area borders; Fig. 1A (37)]. Fixed-radius disks (3-mm radius) were built around putative area centers, defined in relation to the transitions, along the cortical surface (Fig. 1B).

Fig. 1.

Network nodes are defined using putative area centers from RSFC-boundary mapping and labeled by RSFC systems. (A) RSFC-boundary mapping (37) parcellation map depicts the probability of RSFC pattern transitions across the cortical surface. (B) The local minima of the parcellation map were identified, and 3-mm–radius disks were created around these positions to represent the locations of putative area centers along the cortical surface. Disks served as network nodes (n = 441). (C) RSFC systems defined by Power et al. (19). (D) Surface disks labeled by RSFC system membership.

A total of 441 nonoverlapping disks were created across the cortical surface of the two hemispheres; each of these disks served as a node in the construction of a graph representing an individual's brain network. Each node was labeled according to a published functional system map, defined by consensus of system assignments using community detection of RSFCs across multiple thresholds (Fig. 1C) (19). The final system labels of each node are depicted in Fig. 1D, and SI Appendix, Table S6 lists the node count for each functional system. Using independent datasets to define nodes and assign system labels to these nodes allowed interrogation of connectivity within and between systems in an unbiased fashion. Brain graphs were constructed for each participant as a 441 × 441-node graph, labeled by RSFC-defined functional system. Edge weights were calculated as the Fisher z-transformed correlation (Pearson’s r) between each pair of nodes, and negatively weighted edges were removed from each correlation matrix to eliminate potential misinterpretation of negative edge weights (see Experimental Procedures for further details).

Healthy Adult Aging Is Accompanied by Decreasing Connectivity Within Functional Brain Systems, and Increasing Connectivity Between Functional Brain Systems.

To explore whether the relationships between areas of distinct functional systems may differ across healthy adult aging, area-to-area relationships (graph edges) first were characterized according to whether they connected areas within a functional system or between distinct functional systems. Fig. 2A depicts the mean within-system and between-system correlations across age (n = 210; age range, 20–89 y). There is clear evidence that the mean within-system correlation is greater than the mean between-system correlation, independent of age. Importantly, we also observed two different patterns with respect to the types of relationships as a function of age: within-system correlations decreased with increasing age, and between-system correlations increased with increasing age. These observations were statistically supported in a general linear model demonstrating that although the mean correlation of all nodes in the network was predicted by age [F(1,208) = 4.16, P = 0.043], there was also a main effect of type of edge [within- vs. between-system; F(1,208) = 4114.99, P < 0.001] and an interaction between age and edge type [F(1,208) = 45.80, P < 0.001]. This interaction was a result of significant age-related decreases in within-system correlations (r = −0.28, P < 0.001) and significant age-related increases in between-system correlations (r = 0.32, P < 0.001), which are significantly different from one another: z = −7.77, P < 0.001.

Fig. 2.

Increasing adult age is associated with decreasing segregation of brain systems. (A) Mean within-system RSFC decreases with age, and mean between-system RSFC increases with age. (B) Mean node-to-node correlation matrix (10 systems) of each age cohort. Nodes are grouped according to system labeling (Fig. 1D); color bars along axes represent system labels (see legend to the right). Within-system RSFC (on matrix-diagonal) exhibits decreasing strength across cohorts, whereas between-system RSFC (off matrix-diagonal) exhibits increasing strength across cohorts. The latter pattern is apparent particularly for a subset of between-system relationships. For example, RSFCs of the frontal–parietal control or the ventral attention system (highlighted with a white box) with other brain systems (e.g., default system; white arrows) are increasingly greater (yellow/orange colors) from younger to older adult cohorts. (C) Mean system segregation decreases with age, reflecting proportionally greater between-system correlations relative to within-system correlations. (D) Mean network participation coefficient exhibits an age-associated increase, supporting observations related to system segregation. For each scatterplot, a line reflecting the linear regression between age and the dependent variable is depicted.

The results depicted in Fig. 2A reflect node-to-node relationships collapsed across all functional systems identified by Power et al. (19). We focused on 10 consistently identified systems for subsequent analyses (system legend in Fig. 1C; for additional analyses, see SI Appendix, Supplemental Results and Table S5) and evaluated their age-related changes in within- and between-system correlations (SI Appendix, Fig. S3A and Table S7). Although varying in the degree of age-related change, almost all the systems demonstrated comparable patterns of decreasing correlations within systems and increasing correlations between systems, suggesting a general nature to this prominent pattern.

The patterns described above may be appreciated further by examining the 441 node cross-correlation matrices. To do so, participants first were divided into four age-based cohorts. Cohorts were constructed with the aim of maintaining a relatively consistent age range across subjects while ensuring a roughly equivalent number of participants in each cohort. The four adult lifespan categorical divisions were “younger” (YA; 20–34 y; n = 61), “middle-early” (ME; 35–49 y; n = 46), “middle-late” (ML; 50–64 y; n = 43), and “older” adults (OA; 65–89 y; n = 60). For each cohort, a mean matrix was constructed in which the nodes were ordered by their predefined community labels (from Fig. 1D). The mean matrices that depict the 10 systems (Fig. 2B) reinforce the two patterns we noted: (i) correlation values along the diagonal of the matrix reflect within-system relationships, and these decrease with increasing age, and (ii) correlation values off-diagonal reflect between-system relationships, and these increase with increasing age. It is important to note that these patterns are not homogenous across all systems or areas, suggesting that age-associated differences in connectivity may be specific to certain types of relationships. For example, the connectivity between the frontal–parietal control system and other systems increases strongly with age (r = 0.36, P < 0.001), whereas the auditory system does not exhibit as strong an age-related increase in between-system connectivity (r = 0.10, P = 0.156; see SI Appendix, Table S1 for details on the 10 systems). We revisit these important observations in a subsequent section.

Age-Accompanied Differences in Brain Connectivity Reflect a Decreased Segregation of Functional Brain Systems.

The differences in the patterns of within- and between-system RSFCs suggest that aging is accompanied by decreased independence of brain systems (by way of exhibiting weaker connectivity among areas within systems and greater connectivity between areas of distinct systems). As a way of summarizing and quantifying the pattern of differences in the within-system correlations in relation to the between-system correlations, we created a measure of system segregation. This measure was calculated as the difference between the mean magnitudes of between-system correlations from the within-system correlations as a proportion of mean within-system correlation (see Experimental Procedures). Accordingly, values greater than 0 reflect relatively lower between-system correlations in relation to within-system correlations (i.e., stronger segregation of systems), and values less than 0 reflect higher between-system correlations relative to within-system correlations (i.e., diminished segregation of systems). This formulation of system segregation results in a theoretical maximal segregation value of 1, which would reflect a system with greater than 0 mean within-system connectivity but absent connectivity with all other systems.

Mean system segregation is plotted as a function of participant age in Fig. 2C. Across age, segregation values are greater than 0, demonstrating that mean within-system correlations are stronger than mean between-system correlations, regardless of age. Despite this commonality, older age is associated with decreasing system segregation (r = −0.53, P < 0.001). Moreover, this pattern was observed in 8 of the 10 systems of interest when interrogated individually (SI Appendix, Fig. S3B and Table S7).

Our measure of segregation is intimately related to the graph-theoretic concept of participation coefficient. A node’s participation coefficient is a measure of the extent to which a given node connects to nodes in systems (communities) other than its own. Higher values indicate that the node is connected to many nodes in other systems, whereas lower values indicate that the node’s interactions are limited largely to its own system. Based on our findings of decreased segregation with age, we predicted that participation coefficients would increase with increasing age. The mean participation coefficient across all nodes increases with increasing age (r = 0.46, P < 0.001; Fig. 2D; also see SI Appendix, Fig. S2 and Supplemental Results for further analyses on participation coefficients and discussion regarding the distinctions between multiple measures).

To summarize, evaluating multiple measures that quantify intersystem relationships reveals that older adults exhibit proportionally greater connectivity between nodes in different functional systems. Of particular importance, the relationship between age and system segregation was evaluated across multiple choices related to matrix thresholding (i.e., inclusion of negative edges, thresholded vs. unthresholded graphs; SI Appendix, Supplemental Results and Table S2), area node sets [surface based vs. volume based (including subcortical nodes); SI Appendix, Supplemental Results and Table S1], and system labeling (defined by previously published system maps, cohort defined and individually defined; SI Appendix, Supplemental Results, Fig. S1, and Table S1). The negative relationship between age and system segregation was found to be robust in all instances.

Association and Sensory-Motor Systems Exhibit Distinct Patterns of Age-Related Differences in Segregation.

Thus far, we have described system-related changes largely by collapsing measures across all systems. However, there also is evidence that different functional systems exhibit differences in their specific patterns of age-related changes in segregation. For example, it appears that areas in the frontal–parietal control system exhibit greater connectivity to areas in several other systems, including the ventral attention and cingulo-opercular control systems, in OA compared with YA (Fig. 2B). One broad yet useful distinction that characterizes functional systems and their constituent areas is whether they are involved primarily in sensory-motor processing or in more associative processes (e.g., refs. 38, 39). Sensory-motor systems are engaged in neural coding and transformation of incoming sensory and outgoing motor information (e.g., the visual system, the motor system), whereas “association” systems typically direct and integrate information in a wide range of tasks and across multiple modalities (e.g., the frontal–parietal control system and the “bottom-up” ventral visual-spatial attention system). Consistent with their purported roles in coordinating diverse sets of operations with other systems, areas in association systems exhibit widespread and diverse anatomical projections with distributed brain systems (e.g., refs. 38, 40) and have been demonstrated to exhibit greater RSFC with areas in other systems compared with areas in sensory-motor systems (19).

We examined the sensory-motor/association system distinction to determine whether there were age-accompanied differences in segregation as a function of system type. Systems first were classified according to whether they primarily fit a sensory-motor or an association role. Sensory-motor systems included the auditory, hand somatomotor, mouth somatomotor, and visual systems. Association systems included the cingulo-opercular control, dorsal attention, frontal–parietal control, salience, ventral attention, and default systems.

Overall, the results show that aging is associated with decreasing segregation of both association and sensory-motor systems; however, their patterns of age-related changes in segregation differ (Fig. 3 A and B). When we applied linear and nonlinear fits (first-, second-, and third-degree polynomials) to sensory-motor system segregation, we found that the age function for these systems was fit significantly only by a linear model [t(208) = −8.28, P < 0.001, adjusted R² = 0.24], whereas the second- and third-degree polynomials were not statistically reliable [t(208) = 1.13, P = 0.261 and t(208) = 0.01, P = 0.099, respectively]. In contrast, association system segregation was fit significantly by both a linear (t = −8.55, P < 0.001) and a quadratic model [t(208) = −2.80, P < 0.001], but not by a third-degree polynomial [t(208) = −0.09, P = 0.933]. The linear and quadratic models of association system segregation were significantly different from one another [F(1,207) = 7.84, P = 0.006], with the quadratic model having a higher adjusted R² than the linear model (0.28 vs. 0.26, respectively). The significant quadratic function is illustrated by nonparametric local smoothing (local spline approximation; Fig. 3 A and B), which revealed a steeper decline in segregation of association systems at later ages (after an approximate age of 50) compared with earlier ages (see SI Appendix for additional analyses). The different patterns of age-related changes in system segregation as a function of system type can be appreciated by viewing spring-embedded layouts of each cohort’s areal network graph (Fig. 3C). Association systems exhibit less within-system connectivity and greater between-system connectivity, particularly in the oldest age cohort.

Fig. 3.

Sensory-motor and association systems exhibit distinct patterns of age-associated differences in segregation. Locally weighted scatterplot smoothing (LOESS) graphs depict (A) the linear association between decreasing sensory-motor system segregation and increasing age and (B) the quadratic association between decreasing association system segregation and increasing age. Decreases in association system segregation exhibit an inflection point reflecting accelerated reductions starting at an approximate age of 50 y (red dotted line). (C) Spring-embedded layouts of the 10 systems (4% edge density) of the four cohorts’ mean correlation matrices (Fig. 2B). Sensory-motor systems exhibit progressive age-accompanied reductions in both within-system correlations and segregation with other systems (e.g., the visual system, highlighted by the arrows). Association systems exhibit prominent and sudden decreases in segregation with other systems starting in middle-late adulthood [e.g., the frontal–parietal control (in yellow) and cingulo-opercular control systems (in purple) exhibit less within-system connectivity and greater between-system connectivity in middle late and older adult cohorts, highlighted by the circle].

Segregation of Association Systems Predicts Memory Function Across the Adult Lifespan.

We hypothesized that decreasing system specialization may relate to relevant behavioral variables that reflect cognitive ability, and that measurements of between- and within-system RSFC may be leveraged to reveal this relationship. To test this prediction directly, RSFC system segregation was compared with offline measurements of cognitive function collected a week before MRI scanning. Cognitive function was quantified with a series of standardized behavioral tests that spanned multiple cognitive domains (see SI Appendix, Supplemental Experimental Procedures for exact cognitive measures). Given the breadth of the cognitive measurements, participant scores first were submitted to a factor analysis to identify behavioral factors that may represent distinct general functions. An examination of the factor-loading structure revealed that the three factors that emerged were related to measurements of episodic memory, fluid processing, and verbal ability (SI Appendix, Table S3). A significant age-related decrease was observed in the episodic memory (r = −0.20, P = 0.003) and fluid processing components (r = −0.75, P < 0.001), whereas the verbal ability component increased with age (r = 0.32, P < 0.001). This general pattern is consistent with previous characterizations of age-accompanied differences in cognition (41, 42).

We related the three cognitive factor scores to the measurements of system segregation. Critically, because age correlated with all measurements of interest, regression models were computed by first removing the effect of age on the measurements of cognition and segregation (i.e., partial correlations). After we controlled for age, segregation of association systems was significantly related to episodic memory (r = 0.18, P = 0.007): individuals with greater association system segregation exhibited higher episodic memory scores (Fig. 4A). Of particular interest, the high degree of segregation among different association systems was significantly related to memory function, whereas the relationship between memory and segregation of association from sensory-motor systems only approached significance (i.e., association-to-association segregation and memory: r = 0.22, P = 0.001; association-to-sensory-motor segregation and memory: r = 0.13, P = 0.063; Fig. 4 B and C). The significant relationships between memory and association-to-association system segregation remained after Bonferroni correction for the 18 performed comparisons. After controlling for age, no relationship was found between sensory-motor system segregation and memory (r = −0.09, P = 0.203; see SI Appendix, Fig. S5, Table S8, and Supplemental Results for additional analyses).

Fig. 4.

Greater association system segregation is associated with superior long-term episodic memory, independent of age. (A) Episodic memory scores are predicted by participants’ association system segregation. Data points are color coded by participants’ age cohort to demonstrate that the relationship between memory and association system segregation is independent of age. (B) Relationship between episodic memory and segregation of association systems from other association systems and (C) segregation of association systems from sensory-motor systems. For each scatterplot, a line reflecting the linear regression between episodic memory scores and system segregation is depicted.

Discussion

The present findings indicate that increasing age is associated with decreasing segregation of functional brain systems, as defined by patterns of RSFC between brain areas. Compared with sensory-motor systems, association systems exhibit greater decreases in segregation from ∼50 y of age onward. Of particular importance, even after controlling for age, individuals with less segregated association systems exhibited the poorest memory ability. We consider each of these observations and their implications for understanding the organization and function of brain networks across the adult lifespan.

RSFC-Defined System Segregation as a Metric of Functional Specialization.

There is accumulating support that healthy aging is accompanied by changes in the information processing of brain areas as evidenced by task-evoked activity. In some instances, these changes reflect quantitative young/old differences in comparable sets of areas (e.g., refs. 43–45), whereas in other instances, there is evidence for qualitative differences in the areas recruited across cohorts (e.g., refs. 6, 46). It is uncertain whether these changes reflect age-associated strategic differences in task engagement, or they are obligatory differences in response to the ongoing cascade of white-matter and gray-matter brain changes that occur in even very healthy older adults (47, 48). Further, it currently also is uncertain whether any of the qualitative changes are constrained by system distinctions. Despite these uncertainties regarding the cause and nature of changes in task-related activation, given the strong link between patterns of RSFC and task-related activity, we hypothesize that RSFC differences reflect a statistical marker of the activity changes that accompany aging. What is particularly intriguing is that RSFC differences were present both within and between systems.

We have suggested that the functional specialization of a system may be characterized by the balance of connections between areas within the system and limited interaction with areas in other systems. In younger adults, this modular architecture facilitates the common processing roles of areas within similar systems while allowing communication between systems with distinct processing roles (21). Consistent with this, there is evidence for a relationship between measures of modularity and changes in task performance and learning in healthy young adults (49, 50). Conversely, comparisons of younger and older adults have revealed differences in brainwide modularity in a direction consistent with the present observations (30, 32, 33, 36). Importantly, we believe our measure of segregation is more sensitive to age-related differences in network organization than modularity (SI Appendix, Fig. S6) and allows clearer insight regarding the underlying changes contributing to the measure. We hypothesize that age-associated decreases in system segregation may indicate decreased functional specificity of system-based processing roles. Brain areas in distinct systems exhibit greater interaction with continual aging, as reflected in patterns of RSFC. Although this may be an adaptive response to ongoing anatomical and biochemical alterations (1, 2), it does not appear that the increasing interaction between areas in distinct systems confers a benefit to the individual; rather, the increased “blurring” across systems and decreased communication within systems may reflect a fundamental age-related mechanism that negatively affects cognitive function. We return to this point in the following section.

Distinctions Between Association and Sensory-Motor System Segregation May Reflect Important Differences in Age-Accompanied Changes in Information Processing.

Systems involved in “associative” operations exhibited greater age-accompanied decreases in segregation compared with systems involved in processing sensory input and motor output. This distinction was informed further by the observation of different patterns of age-related segregation: decreasing association system segregation was fit better by a quadratic than a linear function, distinguished by a sharper rate of decline from age 50 onward (Fig. 3). This distinction between association and sensory-motor systems may reflect different trajectories in the patterns of decreasing functional specialization. Operations mediated by association systems, including but not limited to the maintenance and execution of task set, allocation of attention, controlled mnemonic retrieval, and executive control, likely are performed by multiple brain systems and require substantial interaction between areas in these systems (39, 51, 52). The decreasing executive ability that characterizes older age (42, 53, 54) may be a consequence of reduced functional specialization of systems that mediate these abilities, as indexed by decreasing association system segregation.

Although executive abilities exhibit a particular age-associated decline, there is substantial evidence that aging is associated with an acceleration of changes that affects both sensory and more associative or “cognitive” functions jointly (53). Lindenberger and Baltes (55) suggested that the interrelationships they observed among different cognitive and sensory behavioral systems in very old adults were mediated by some basic brain mechanism or “common cause” that was deteriorating with age. In fact, they speculated that the declining specificity of behavioral systems that occurs with age was the result of “dedifferentiation” of the brain. Based on our present findings, we speculate that the underlying substrate of the observations leading to the common-cause hypothesis may be the degree of between-system interactions occurring in the adult brain. Specifically, we suggest that the tighter link between sensory and cognitive function in older age is intimately related to the decreasing RSFC-defined system segregation observed here.

System segregation was predictive of a summary measure of memory. The age-invariant relationship between association system segregation and long-term memory scores suggests that our measurements of network properties exhibit a much broader relationship to behavior than that which simply characterizes differences present across adult aging. The direction of the relationship suggests that decreased system segregation has negative consequences for behavior (see SI Appendix, Supplemental Results and Discussion for additional analyses related to relationships between segregation and cognitive measures) and that measurements of segregation may be used as an important neural measure across a range of cross-sectional and longitudinal comparisons related to health and pathology (also see SI Appendix, Supplemental Discussion), but also experimental and interventional manipulations (e.g., refs. 50, 56).

The Use of Graph Theory to Study Brain Networks.

Two authors of the present report (G.S.W. and S.E.P.) previously stressed the importance of examining brain network nodes that represent biologically meaningful entities, and our current efforts have attempted to satisfy this requirement by building brain networks from an area parcellation method that appears neurobiologically plausible [brain areas (37); also see SI Appendix, Supplemental Discussion]. Although relatively accurate node representation is an important constraint in generating valid brain networks, it also is important to understand the nature and interpretation of the derived network measures (e.g., refs. 37, 57). We briefly describe an illustrative and cautionary example.

We have described the use of a novel measure, termed system segregation, which characterizes the amount of within- and between-system connectivity in brain systems. This measure is strongly associated with many traditional graph measures (e.g., participation coefficient, modularity) but can capture the trend of aging more strongly than the other measures (SI Appendix, Supplemental Results and Fig. S6B). In addition, we also quantified a prominently examined measure, global efficiency (GE), in each of our participants (see SI Appendix, Supplemental Experimental Procedures for details on calculation). GE is a measurement related to the average shortness of paths between nodes of a network, where networks with shorter average paths have higher GE. Several previous studies examining differences in GE as a function of age cohort found greater GE in older vs. younger adults (31, 35). In the present study, examining the relationship between GE and age also reveals that increasing age is accompanied by increasing GE. Critically, this observation is trivial and logical when one recognizes that the introduction/strengthening of connections (edges) between systems (clusters) quickly decreases average path length (58). Given the negative relationship between age and system segregation, it is not surprising that GE increases with age. Importantly, accounting for system segregation eliminates the significant age–GE relationship (SI Appendix, Fig. S7). Observations related to differences in other summary measures, such as “small-worldness” (e.g., ref. 33), may be similarly sensitive to these forms of basic underlying properties.

Concluding Comments.

Using a network-based approach revealed that healthy aging is accompanied by decreased segregation of brain systems defined by their patterns of resting-state correlations. We hypothesize that system segregation is an important measure and guide toward understanding functional specialization of areas within distinct brain systems, and it will be important to examine how changes in segregation of specific systems affect the functional roles of their areas. The age vs. system segregation relationship was most prominent for association systems following 50 y of age. After controlling for age, we also found that the degree of association system segregation was predictive of offline measures of memory ability, suggesting that system segregation may be an age-invariant marker of individual differences in cognition.

Of additional interest, there is evidence for developmental differences in patterns of within- and between-system RSFC, wherein young adults appear more segregated than children and adolescents [(59), and this is also the case following more stringent movement correction (60)]. It will be important to examine system segregation in the context of early brain development more closely to understand how the mechanisms giving rise to changes in segregation differ across the entire lifespan. By measuring functional relationships in the absence of overt cognitive tasks, the present approach seems particularly well suited toward characterizing and understanding the complex organization of brain networks across various cohorts and species and how this organization develops, differs, and evolves in relation to behavior.

Experimental Procedures

Participant Demographics.

Healthy adults from the Dallas Lifespan Brain Study (DLBS) who completed a resting-state fMRI scan were included in the present study (n = 268). Participants were recruited from the Dallas–Fort Worth community and provided written consent before participating. All study procedures were reviewed and approved by the Institutional Review Boards at The University of Texas at Dallas and The University of Texas Southwestern Medical Center. A final sample of 210 participants met the minimum requirements of RSFC data quality (See SI Appendix, Supplemental Experimental Procedures for details and additional exclusion criteria). Table 1 summarizes these participants’ demographics, broken down by age cohorts.

Table 1.

Demographic information

Measure	Younger	Middle-early	Middle-late	Older	P
N	61	46	43	60	NA
Age range, y	20–34	35–49	50–64	65–89	NA
Female, %	77	65	72	55	ns
Education, y (SD)	14.90 (3.82)	14.67 (4.19)	15.49 (3.94)	13.25 (3.86)	0.026*
MMSE score	28.67 (1.19)	28.74 (1.12)	28.51 (1.16)	27.68 (1.17)	<0.001

MMSE, mini-mental state examination; NA, not available; ns, not significant.

^*Mean differences were tested with χ² test for sex distribution and ANOVA for years of education and MMSE.

Experimental Design and Data Acquisition.

The DLBS consists of multiple data acquisition sessions that include cognitive and neuropsychological testing and MRI scanning. The MRI scanning session consists of a series of anatomical MRI and fMRI scans acquired using a Philips Achieva 3T scanner. See SI Appendix, Supplemental Experimental Procedures for image acquisition details of the anatomical T1 scan and resting-state functional scan.

Data Preprocessing.

Adult-lifespan atlas construction.

Typical registration targets for MRI (e.g., MNI152) are based on a young adult sample, which systematically introduces greater registration error among older adults (61). To avoid this issue, following the procedure outlined in Buckner et al. (61), an adult-lifespan sample-representative template was created (see SI Appendix, Supplemental Experimental Procedures for details) to incorporate typical structural variation present across our healthy adult lifespan sample (n = 268; age range, 20–89 y).

Standard fMRI preprocessing.

Functional images first were processed to reduce artifacts (62), register each individual’s magnetization-prepared rapid gradient echo (MP-RAGE) image to the adult-lifespan sample atlas, and resampled to 3-mm isotropic atlas space. See SI Appendix, Supplemental Experimental Procedures for details.

RSFC preprocessing.

Several additional preprocessing steps were used to reduce spurious variance unlikely to reflect neuronal activity in RSFC data, which included in the following order: (i) multiple regression of the BOLD data to remove variance related to the whole brain signal (cf. ref. 63), ventricular signal, white matter signal, six detrended head realignment parameters obtained by rigid-body head motion correction, and the first-order derivative terms for all aforementioned nuisance variables; (ii) band-pass filtering (0.009–0.08 Hz); and (iii) volumetric spatial smoothing (6-mm full width at half maximum in each direction). To reduce the effect of motion artifact on RSFCs, data were processed following the recently described “scrubbing” procedure (64). See SI Appendix, Supplemental Experimental Procedures for details and Supplemental Results for additional analyses related to movement.

Surface preprocessing.

There is growing evidence that landmark-based registration of a participant’s cortical surface exhibits more accurate alignment of cortical anatomy than either linear or nonlinear volume-based registration techniques (65). As this may confer particular benefits toward minimizing variation in anatomical registration in the present adult-lifespan sample, FreeSurfer 5.0 was used to process volumetric images into fsaverage surface images. The left and right hemisphere fsaverage surfaces were brought into register with each other by mapping onto a hybrid left–right fsaverage atlas (“fs_LR”) (66). See SI Appendix, Supplemental Experimental Procedures for details.

Area node definition.

Network nodes were constructed by processing further a recently published RSFC-based area parcellation map. This parcellation map identified cortical locations where patterns of RSFC exhibited abrupt transitions in a large group of subjects (i.e., putative area borders; Fig. 1A) (37). There is evidence that RSFC transitions are least pronounced at locations corresponding to an area’s putative center (37); accordingly, to avoid locations with greater ambiguity as to the area’s identity, locations where RSFC maps exhibited high local stability, local minima of RSFC transition, were identified in the RSFC parcellation map (in fs_LR space) (37). Minima were required to be at least 8 mm apart along the cortical surface, and minima identified within the FreeSurfer medial wall were excluded. Fixed-radius disks (3-mm geodesic radius) were built around each of the minima locations (Fig. 1B). By representing nodes as fixed-radius disks around the probabilistically defined center of an area, this method of node definition avoids locations of greater uncertainty. In addition, the small fixed-radius disks have the benefit of ensuring equivalent node size across the entire node set and minimizes the likelihood of creating nodes that capture information from multiple adjacent areas (for discussion, see refs. 17, 67, 68). Small fixed-size disks also are beneficial if the nodes are intended to sample information across individuals who may have variation in their area parcellation (69–71). This technique, by definition, limits use of information from the full extent of a given area, but the many advantages noted above outweighed this concern. A total of 441 nonoverlapping disks were created across the two cortical hemispheres (L = 221, R = 220). Visual inspection confirmed that the disks did not cross any strong RSFC boundaries on the original parcellation map.

Disks were labeled according to a published functional system map defined by voxelwise community detection of RSFC, in which systems were assigned based on consensus across multiple thresholds (19) (Fig. 1C). As the RSFC parcellation and system maps both are represented on the fs_LR cortical surface, we could combine information from both approaches and interrogate DLBS data in the same cortical space. All fs_LR vertices first were labeled according to RSFC system membership. The system labels of all vertices within a disk then were identified, and each disk was labeled with a system by a winner-take-all approach. The final system labels of each disk are depicted in Fig. 1D, and SI Appendix, Table S6 lists the disk count for each functional system. Single-system analyses in this study focused on 10 consistently identified systems (see legend in Fig. 1C) and include at least four nodes per system.

Basic RSFC and Graph-Theoretic Analysis.

Preparing RSFC data for connectivity analysis.

For each participant, the resting-state fMRI time series of vertices within each of the 441 nodes was extracted, and the vertex-mean time course was computed for each node. The cross-correlation of each node’s time course with every other node’s time course was calculated, forming a node-to-node correlation matrix. Correlation coefficients were converted into z-values using Fisher’s equation (72). The resulting Fisher z-transformed r-matrix (z-matrix) is a fully connected, weighted relatedness graph. Although it is possible to include negative ties in a network analysis (73), because of the present ambiguity regarding the meaning of negative correlations (63, 74, 75), negative z-values were excluded from the data matrix. The final data matrix for each participant was a 441 × 441 z-matrix with the diagonal and negative values set to zero.

Within-system and between-system connectivity.

Nodes were labeled by functional systems (Fig. 1D). For a given system, within-system connectivity was calculated as the mean node-to-node z-value of all nodes of that system to each other (e.g., the mean of the z-values between all default system nodes to all other default system nodes). Conversely, between-system connectivity was calculated as the mean node-to-node z-value between each node of a system and all nodes of all other systems (e.g., mean of z-values between all default system nodes and all other nodes in the brain).

System segregation.

A measure of system segregation was computed to summarize values of within-system correlations in relation to between-system correlations. Specifically, this measure takes the differences in mean within-system and mean between-system correlation as a proportion of mean within-system correlation, as noted in the following formula:

$$\text{System segregation}=\frac{{\overline{\text{Z}}}_{\text{w}}-{\overline{\text{Z}}}_{\text{b}}}{{\overline{\text{Z}}}_{\text{w}}},$$

where ${\overline{\text{Z}}}_{\text{w}}$ is the mean Fisher z-transformed r between nodes within the same system and ${\overline{\text{Z}}}_{\text{b}}$ is the mean Fisher z-transformed r between nodes of one system to all nodes in other systems. Importantly, our measure of system segregation retains the weight of all positive edges in a graph, allowing weak connections to contribute to the characterization of system interactions.

Participation coefficient.

A node’s participation coefficient measures to what extent a node interacts with nodes in other systems, in relation to the total number of connections it possesses (total degree). Participation coefficient results presented in the main manuscripts were calculated from each subject’s z-matrix, where negatives and the diagonal are coded as 0. Participation coefficients for each node were calculated according to the following equation (76, 77):

$${\text{P}}_{\text{i}}=1-{\displaystyle \sum }_{m\in M}{\left(\frac{k_i^w\left(m\right)}{k_i^w}\right)}^2,$$

where $k_i^w\left(m\right)$ is the weighted connections of node $i$ with nodes in system $m$ (a system to which node $i$ does not belong ) and $k_i^w$ is the total weighted connections node $i$ exhibits. Higher participation coefficient values indicate proportionally greater communication with nodes in other systems.

Behavioral Battery and Analysis.

Exploratory factor analysis (EFA) was performed on 18 cognitive tasks (22 dependent variables; see task list in SI Appendix, Table S3) to find the common variance between multiple variables, thereby reducing the number of outcome variables and tests to be performed. The EFA (varimax rotation, three factors, as determined by parallel analysis; see SI Appendix, Supplemental Experimental Procedures for details) resulted in groupings of variables we identified as representing measures related to fluid processing, episodic memory, and verbal ability. The tasks that contributed to each factor are detailed in SI Appendix, Table S3. Factor scores were produced by using the regression method of Thomson (78).