Aging has long been suggested to be accompanied by cognitive decline, even in the absence of dementia or other neurological insults. This decline is more pronounced in cognitive abilities such as processing speed, working memory, and encoding new information into episodic memory, whereas other capacities, particularly semantic knowledge, and emotional processing, seem to remain relatively stable along the adult lifespan (Park et al., 2002; Buckner, 2004; Hedden and Gabrieli, 2004; Park and Payer, 2006; Salthouse, 2009). The precise determination of age-related cognitive changes remains a challenge due to high variability across cognitive skills and research designs. Largely, a monotonous, linear to a first approximation, decline has been reported, starting in the third decade (Buckner, 2004; Park and Reuter-Lorenz, 2009), with some cognitive abilities exhibiting an accelerated deterioration in the seventh decade (Hedden and Gabrieli, 2004).

Age-associated cognitive impairment has frequently been linked to structural modifications in both gray matter (GM) and white matter (WM) (Buckner, 2004; Madden et al., 2009; Bergfield et al., 2010; Giorgio et al., 2010; Thambisetty et al., 2010), though temporal characteristics of these changes and the exact relations with cognitive performance have not been defined (Ferreira et al., 2014). GM volume loss has typically been reported as linear, beginning early in life (Good et al., 2001; Ge et al., 2002; Giorgio et al., 2010; Chen et al., 2013) and only few studies indicated non-linear trends (Terribilli et al., 2011; Lebel et al., 2012; Ziegler et al., 2012). Conversely, WM volume dynamics has frequently been described as non-linear with an inverted-U trajectory over the entire lifespan peaking in the fourth to fifth decade (Giorgio et al., 2010; Chen et al., 2013). Further, fractional anisotropy (FA), a diffusion tensor imaging (DTI) index thought to reflect WM integrity, was shown to deteriorate earlier than WM volume (Salat et al., 2005; Giorgio et al., 2010), quickly after peaking between 20 and 42 years of age (Lebel et al., 2012). Interestingly, a few DTI studies reported a considerable FA decrement in the transition between early and middle adulthood, whereas no significant difference was observed between middle-aged and older adults (Salat et al., 2005; Giorgio et al., 2010).

In addition to structural modifications, alterations in neurotransmitter function have been described in healthy aging, particularly reductions in the binding potential of monoaminergic transmitters (Hedden and Gabrieli, 2004; Park and Reuter-Lorenz, 2009; Bäckman et al., 2010). Age-related cognitive deficits have been most associated with dopaminergic dysfunction (Volkow et al., 1998; Bäckman et al., 2000, 2006; Erixon-Lindroth et al., 2005), though the exact trajectory of dopaminergic decline is still disputed with both linear and non-linear reductions documented (Bäckman et al., 2006). Of note are a few reports highlighting a dramatic decline by midlife in the availability of dopamine receptors (Antonini et al., 1993), dopamine transporters (Mozley et al., 1996), and serotonin receptors (Sheline et al., 2002); yet, these are awaiting confirmation by larger and wider age-range studies.

Importantly, several other neurobiological measures potentially related to cognitive aging are greatly modified by middle-age. For instance, deficits in adult dentate and subventricular zone neurogenesis, considered to play a role in aging, were reported as already established in middle-aged animals (Kuhn et al., 1996; Hamilton et al., 2013). The levels of brain-derived neurotrophic factor (BDNF), phosphorylated cAMP response element binding protein (p-CREB) and neuropeptide Y (NPY), recognized as positive regulators of dentate neurogenesis and memory function, significantly declined as early as middle-age in the rat hippocampus (Hattiangady et al., 2005). Moreover, age-related synaptic plasticity changes, such as long-term potentiation deficits, were evident in middle-aged animals (Rex et al., 2005; Lynch et al., 2006). In humans, signal pathway activation (e.g., ERK-, JNK-, mTOR-, MAPK-, mitochondrial apoptosis, and caspase cascade) in fibroblasts derived from young, middle-aged and old healthy adults as well as Progeria patients showed more pronounced alterations in the transition between the young and middle-aged groups than among other groups (Aliper et al., 2015). Further, a set of genes related particularly to synaptic plasticity, vesicular transport and mitochondrial function showed reduced expression in the frontal cortex after age 40 (Lu et al., 2004). Collectively, this evidence from animal models and humans implies that some age-related decline may present in middle-age.

The timing of age-related structural modifications was reported to vary across brain regions with earlier deterioration in high-order association areas relative to primary sensory and motor cortices (Raz, 2000). Association regions are known to develop relatively late, both ontogenetically and phylogenetically, compared to primary sensory and motor areas (Tau and Peterson, 2010; Fjell et al., 2014); therefore age-related brain changes were suggested to follow the “last-in-first-out” rule (Raz, 2001). Today, brain aging is thought to involve not only regional dysfunction but also alterations in the communication between remote brain areas (Andrews-Hanna et al., 2007; Antonenko and Floel, 2014). In addition to WM structural connections evaluated by means of, for example, DTI, functional connections between distinct GM regions are also explored. Analysis of coherent spontaneous fluctuations in the fMRI blood oxygenation level-dependent (BOLD) signal, termed intrinsic functional connectivity MRI (fcMRI; for reviews, see Fox and Raichle, 2007; Buckner et al., 2013) now allows studying the brain’s connectional architecture across the lifespan. Heretofore, several fcMRI studies have documented age-related reduced connectivity within large-scale intrinsic connectivity networks (ICNs) (Andrews-Hanna et al., 2007; Damoiseaux et al., 2008; Koch et al., 2010; Tomasi and Volkow, 2012; Ferreira and Busatto, 2013), and altered connectivity between ICNs has also been proposed (Onoda et al., 2012; Antonenko and Floel, 2014). However, this literature refers to only a few ICNs, lifespan trajectories of intra-network connectivity decline and differences between networks are not well characterized and results regarding inter-network connectivity changes are variable, with both increases (Meunier et al., 2009; Betzel et al., 2014; Chan et al., 2014; Geerligs et al., 2015; Grady et al., 2016; Spreng et al., 2016) and decreases (Meunier et al., 2009; Allen et al., 2011; Onoda et al., 2012) reported.

The present study aims to fill this gap and extend our view of intra- and inter-network functional reorganization during aging. The authors believe that a comprehensive outlook of the lifetime dynamics of brain functional connectomics is crucial for deciphering the mechanisms underlying brain aging, determining intervention targets and their optimal timing. Herein, an extensive fcMRI cohort of 887 healthy individuals between the ages of 21 and 85 was used to characterize functional connectivity changes across the adult lifespan in seven robust well-established ICNs: the default mode (DMN), salience (SN), dorsal attention (DAN), fronto-parietal control (FPCN), auditory (AN), visual (VN), and motor (MN) networks (Damoiseaux et al., 2006; Van den Heuvel and Hulshoff Pol, 2010; Van Dijk et al., 2010; Yeo et al., 2011; Brier et al., 2012). Consistent with previous lifespan studies (Giorgio et al., 2010; Van Helmond et al., 2010; Park et al., 2013; Ferreira et al., 2015; Kennedy et al., 2015), the cohort was divided into young, middle-aged, and old groups. Consistent with earlier imaging literature (Li et al., 2011, 2013; Montembeault et al., 2012; Belcher et al., 2013; Geerligs et al., 2015) the seven networks were divided into high-order cognitive networks (DMN, SN, DAN, and FPCN), known to involve associative brain areas and play a role in high-order cognition (Van den Heuvel et al., 2009; Menon, 2011; Braga and Leech, 2015) and primary sensory and motor networks (AN, VN, and MN). Differential response of these network groups to aging was assumed based on recent fcMRI reports (Chan et al., 2014; Geerligs et al., 2015) and the “last-in-first-out” theory (Raz, 2001). Taken together, we expected functional connectivity decline by middle-age within high-order cognitive networks, as well as early between-network connectivity alterations, where high-order cognitive networks are involved.

The present findings indicate that functional connectivity (FC) decline in high-order cognitive networks is already evident by middle-age, as expressed by fcMRI measures. The time course of FC alterations within a motor network is characterized by an early increment followed by a decrement. Finally, most inter-network connectivity changes occur in the transition from young to middle adulthood, highlighting that several potentially meaningful FC changes onset early in the adult lifespan.

Growing evidence now suggests that age-related cognitive decline is associated with FC alterations within large-scale brain networks (Andrews-Hanna et al., 2007; Damoiseaux et al., 2008; Ferreira and Busatto, 2013). The DMN has been the focus of much of this literature, most studies reported FC decline with age and associated FC decline with reduced cognitive performance (e.g., Andrews-Hanna et al., 2007; Damoiseaux et al., 2008; Ferreira and Busatto, 2013; Vidal-Piñeiro et al., 2014). More recently, age-related connectivity decreases have also been noted in other ICNs (Andrews-Hanna et al., 2007; Onoda et al., 2012; Tomasi and Volkow, 2012; Ferreira and Busatto, 2013; Geerligs et al., 2015). Importantly, most of the above mentioned studies directly compared young and old adults, while only a few studied young and middle-aged individuals and actually reported diminished FC in middle-adulthood (Bluhm et al., 2008; Allen et al., 2011; Evers et al., 2012). Here, FC decline by middle-age is observed in an extensive cohort of individuals within the DMN, SN, DAN, and PFCN. This early FC decline accords with previously suggested trajectories of cognitive and structural deterioration with aging (Buckner, 2004; Park and Reuter-Lorenz, 2009; Giorgio et al., 2010; Chen et al., 2013). Furthermore, several recent studies explicitly underscored a detectable cognitive decline by middle-age in both humans (Singh-Manoux et al., 2012; Ferreira et al., 2015) and animals (Moore et al., 2006; Stouffer and Yoder, 2011; Shoji et al., 2016), supporting the view that connectivity decreases in specific functional brain systems may be associated with particular behavioral changes (Andrews-Hanna et al., 2007; Onoda et al., 2012).

Even more remarkable is the fact that within high-order cognitive networks young to middle-age FC decline is more pronounced than the changes observed between middle-aged and older adults. While this finding conforms with earlier structural connectivity findings (Salat et al., 2005; Giorgio et al., 2010) as well as with age-related alterations in genetic, biochemical, and neurophysiological variables (Antonini et al., 1993; Kuhn et al., 1996; Mozley et al., 1996; Sheline et al., 2002; Lu et al., 2004; Hattiangady et al., 2005; Rex et al., 2005; Lynch et al., 2006; Hamilton et al., 2013; Aliper et al., 2015), it appears inconsistent with the progressive course of cognitive decline. Although the source for this lifespan FC dynamics cannot be determined by the current study, compensation and dedifferentiation theories might be proposed to account for the apparent late FC plateau. Park and Reuter-Lorenz (2009) suggested in their “scaffolding theory of aging and cognition” (STAC) that although the brain is subjected to multiple neural challenges throughout the lifespan, adaptive compensatory processes help it maintain homeostatic cognitive function up to a certain point. The absence of further FC decline and even connectivity increment in several node pairs in older age might be related to such compensatory mechanisms, as prudent observation of a few behavioral studies seems to reveal a hitherto unreported cognitive plateau around middle-age (Villardita et al., 1985; Park et al., 1996; Buckner, 2004; Park and Payer, 2006). Dedifferentiation, the loss of functional specialization in brain activity, is an alternative theory commonly proposed to account for cognitive aging (Park and Reuter-Lorenz, 2009). One model which was proposed to represent dedifferentiation is the hemispheric asymmetry reduction in older adults (HAROLD) (Cabeza et al., 2002; Dolcos et al., 2002; Li et al., 2009). In line with this model and in accordance with our findings, Zuo et al. (2010) showed that FC between geometrically corresponding interhemispheric regions follows a quadratic U-shaped trajectory across the lifespan with a turning point in middle-age in heteromodal regions. Future insights into the physiological origins of BOLD-based FC and their age-related changes (e.g., dopaminergic transmission; Ferreira and Busatto, 2013) as well as future studies concurrently evaluating FC and cognitive performance may inform on the underlying mechanisms of the lifetime FC variations observed here.

In addition to temporal aspects of connectivity decline, spatial features should be considered. Our study supports the view that different brain systems are differentially affected by aging. Here, the MN showed a significantly different FC time course compared with high-order cognitive networks (early FC increment followed by a late decrement). Age-related changes in FC within the MN have been studied only rarely and results are conflicting with both decreases (Wu et al., 2007) and increases (Tomasi and Volkow, 2012; Solesio-Jofre et al., 2014) reported. However, consistent with our finding, a previous analysis of brain structural covariance networks (Li et al., 2013) uncovered a difference between motor and high-order cognitive networks. Li et al. (2013) report that non-motor networks demonstrate a distributed topology in the young group, shrinkage into a more localized topology in the middle-aged group, and maintained localized topology in the older group, while the MN shows increased spatial distribution in middle-age and decreased distribution in older participants. Moreover, a recent study by Song et al. (2014) reported an equivalent opposing effect of aging on FC within the DMN and a sensorimotor network.

The observed difference between the motor and high-order cognitive networks is in line with the “last in, first out” hypothesis. Moreover, differential vulnerability to aging is also noted within the same network. Interestingly, the vmPFC, LTC and frontal pole, here showing maximal connectivity reduction among DMN regions, were previously reported as being regions of advanced age-related atrophy and greater expansion during evolution (Hill et al., 2010; Bludau et al., 2014; Fjell et al., 2014). As for the MN, the reduced connectivity in an anterior cerebellar region, probably associated with motor function (Bernard et al., 2013), is consistent with previous reports of reduced cortico-cerebellar connectivity with aging (Bernard et al., 2013; Bernard and Seidler, 2014). The concurrent increased connectivity in cortical MN regions is consistent with the hypothesis that shorter connections are enhanced in older individuals, while longer connections are diminished (Rowe et al., 2006). Other early changes observed in the MN are probably due to increased anticorrelations of subcortical regions with the LMC.

Age-related alterations in interrelations between networks represent another central finding of the current study. Most prominent alterations were increased correlations (e.g., DAN-MN) and decreased anticorrelations (e.g., DMN-SN) between networks in the transition from young to middle adulthood. Our observations strongly support accumulating evidence for enhanced communication between functional networks with aging (Chan et al., 2014; Geerligs et al., 2015; Turner and Spreng, 2015; Gallen et al., 2016; Spreng et al., 2016; Tsvetanov et al., 2016). Increased correlations and decreased anticorrelations between networks both reflect reduced selectivity and specificity in the brain’s intrinsic functional architecture. These changes have been associated with compromised cognitive performance (Kelly et al., 2008; Hampson et al., 2010; Antonenko and Floel, 2014; Chan et al., 2014; Tsvetanov et al., 2016) and hence comply with the dedifferentiation theory of aging; yet, a role in compensation has also been suggested (Gallen et al., 2016; Grady et al., 2016). It should be mentioned that GSR used as a preprocessing step in our study has been reported to artifactually enhance anticorrelations between networks (Fox et al., 2009; Murphy et al., 2009). However, a repeated analysis without GSR confirmed age-related reduced anticorrelation between the DMN and SN (Supplementary Figure S3). This finding is consistent with recent aging studies omitting GSR (Keller et al., 2015; Spreng et al., 2016) and supports previous literature suggesting neural rather than artifactual origin of anticorrelations (Chang and Glover, 2009; Fox et al., 2009; Carbonell et al., 2011; Chai et al., 2012; Keller et al., 2013; Power et al., 2014). Additionally, in line with a few previous reports (Allen et al., 2011; Onoda et al., 2012), our analysis also reveals diminished between-network connectivity with aging for some network pairs. Undoubtedly, the complexity of age-related changes in integration of information between networks warrants further investigation. The importance of the present findings is in pointing out decreased segregation of brain systems by middle-age.

The use of a dataset such as the ‘1000 Functional Connectomes’ Project inherently imposes a source of variability among participants in terms of demographic parameters, cognitive intactness, arousal level, eye opening condition, head movement and technical acquisition parameters, features that have been reported to affect resting-state correlations (Power et al., 2012; Van Dijk et al., 2010, 2012). The statistical analysis was designed to account for these potential biases, as far as the data allowed. Inclusion of participants of different age groups further augments variability in terms of resting metabolic rate (Peng et al., 2014), resting cerebral blood flow (Lu et al., 2011), vascular CO₂ reactivity (Lu et al., 2011; Murphy et al., 2013); blood pressure (AlGhatrif et al., 2017); and hemoglobin concentration (Patel, 2008) all known to affect BOLD fMRI signal and/or the coupling of neural activity to the BOLD signal (D’Esposito et al., 1999; Levin et al., 2001; Patel et al., 2012; Murphy et al., 2013; Mark et al., 2015). Nevertheless, the uneven effect of aging on large-scale brain networks (e.g., DMN vs. MN), emphasized here and elsewhere (Song et al., 2014; Marstaller et al., 2015), probably indicates that age-related FC alterations cannot be merely attributed to changes in neurovascular coupling. Moreover, the GSR step applied in data preprocessing is expected to lessen the confounding effects of the above mentioned physiological factors as well as of cardiac and respiratory cycles (Van Dijk et al., 2010), which have not been directly measured. Though a matter of long-standing debate (Fox et al., 2009; Murphy et al., 2009; Murphy and Fox, 2016), GSR seemed warranted under the current study due to several important reasons: it was shown to enhance the detection of network-specific seed-based correlations (Fox et al., 2009), which are at the core interest of this work; reduce motion and hardware artifacts (Power et al., 2016), an important advantage when evaluating a large multi-center dataset; and enhance the neuronal-hemodynamic correspondence (Keller et al., 2013). Taking into consideration the interpretive complexity imposed by GSR (Murphy and Fox, 2016), we repeated the analyses without GSR, confirming that the study’s main outcomes are present. Another limitation of this study is the absence of data regarding WM hyperintensity burden which has been shown to increase with age (Nyquist et al., 2015) and affect FC (Liang et al., 2016). As both FC decrements and increments have been reported in the presence of WM lesions (Liang et al., 2016), it is difficult to appreciate this parameter’s influence on our results. Finally, the absence of cognitive data for this cohort limits the ability to contemplate regarding the mechanisms underlying current observations. Future single-site longitudinal studies combining detailed analysis of behavioral performance, elaborated demographic information, laboratory data, structural and functional imaging data, as well as measures to control for non-neuronal physiological parameters, might better address the above mentioned confounds and expand the scope of our findings.

In summary, our results confirm the hypothesis that normal brain aging involves reorganization of large-scale functional brain systems. Furthermore, it underscores the early occurrence of FC alterations where high-order cognitive networks are involved. In an era in which we seek ways of augmenting cognition throughout healthy aging, deferring the onset of connectivity decline, slowing down its pace and prolonging the proposed subsequent plateau should be considered potential targets for intervention.

The imaging data of the current study were obtained from 17 research sites through the online dataset of the International Neuroimaging Data-sharing Initiative (INDI). Each center’s ethics committee approved submission of de-identified data as detailed in the NITRC: ‘1000 Functional Connectomes’ Project site (http://fcon_1000.projects.nitrc.org/). The institutional review board of Rambam healthcare campus approved the receipt and analysis of these data.

TS-T, NB, AE, JA-P, and IK contributed to the conception and design of this work. TS-T, NB, and IK collected, analyzed, and interpreted the data. RP contributed to data analysis. ES designed and performed the statistical analyses. TS-T, NB, ES, JA-P, and IK contributed to in-depth discussions about methods and results. TS-T and IK wrote the paper. All authors revised and approved the manuscript.

The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest. The reviewer NE and handling Editor declared their shared affiliation, and the handling Editor states that the process nevertheless met the standards of a fair and objective review.