Decreased Functional Connectivity by Aging Is Associated with Cognitive Decline

Aging is related to cognitive decline, and brain functions such as those involved in attention, memory, motor control, and emotional control are compromised in elderly people. Over the past two decades, many imaging techniques have been applied to investigate the neural correlation of the aging effects on cognitive function. Structural neuroimaging studies including voxel-based morphometry and diffusion tensor imaging indicate that the gray matter of the human brain selectively shrinks with age and that this shrinkage is associated with the selective disruption of anatomical connections (Raz & Rodrigue, 2006). Specifically, association cortices are more profoundly affected by aging than primary sensory–motor cortices. Not only regional changes but also alterations in the connections between brain regions may underlie age-related cognitive decline (O'Sullivan et al., 2001).

Recently, resting state fMRI (rs-fMRI) has generated new research findings. The rs-fMRI studies have suggested that brain activities are organized by multiple subsystems, resembling specific neuroanatomical systems (Damoiseaux et al., 2006). These subsystems are called resting state networks (RSN), and the activities of brain regions involved in a given RSN show high temporal correlations. Some of the rs-fMRI studies have used a model-free approach involving independent component analysis (ICA) to describe the RSN. This ICA statistical technique allows extraction of brain functional networks with distinct spatio-temporal patterns by identifying spatially independent and temporally synchronous brain regions. Thus, many RSNs such as those involving default mode processing, sensation, motor behaviors, and attention have been identified so far. Although the functional connectivity on the basis of the RSNs would enhance the understanding of aging, there are relatively few studies investigating aging effects on the RSNs.

The default mode network (DMN) is one of the salient RSNs; it is known to be active at rest (Raichle, 2006). The main regions in the DMN are the medial pFC, posterior cingulate cortex, and inferior parietal lobe. Damoiseaux and his colleagues have reported decreased activation of the DMN in older subjects compared with younger ones and its significant relationship with performance in a trail making test (Damoiseaux et al., 2008). Koch and his colleagues also reported similar aging effects on DMN coactivation (Koch et al., 2010). Esposito et al. have shown that the overall connectivity of the DMN was negatively correlated with age, and the negative correlation existed for all DMN regions in variable degrees (Esposito et al., 2008). This finding was confirmed by a large-scale study using the 1000 Functional Connectomes Project data set (Biswal et al., 2010). A different approach using a method assessing functional connectivity density suggested that long-range connections in DMN may be more vulnerable to aging than short-range connections (Tomasi & Volkow, 2011).

According to a study using graph theory (Achard & Bullmore, 2007), the global efficiency of functional connectivity was reduced in older people, and the detrimental effects of age on efficiency were specific to the frontal and temporal cortical and subcortical regions. In addition, the number of intermodular connections with frontal modular regions was reduced in older people compared with young people, whereas the number of connection nodes in posterior and central modules was increased (Meunier, Achard, Morcom, & Bullmore, 2009). The previously mentioned study using the functional connectivity density reported age-related changes in the dorsal attention network (based on the lateral prefrontal and parietal cortex), as well as in the somatosensory and subcortical networks, in addition to those in the DMN (Tomasi & Volkow, 2011). These studies suggest that not only the DMN but also the other RSNs are affected by aging, and these disrupted RSNs could be the neural basis of cognitive decline in elderly people.

Previous ICA studies investigating aging effects on the RSNs have directly compared young with elderly subjects (Koch et al., 2010; Esposito et al., 2008). The sample size in each of these studies was too small for parametric analysis. In the current study, we explored the effect of age on the functional connectivity of RSNs and its relationship with cognition in a large number of healthy subjects. We hypothesized that the connectivity of some RSNs, including that of the DMN, would decrease with age, and that the change is correlated with cognitive decline. We also performed seed-based analyses to corroborate the evidence from ICA analyses.

This study included 73 normal individuals (43 men, 30 women). The mean age of the participants was 60.2 ± 12.8 (SD) years, and the range was 36–86 years old. All participants had no history of neurological or psychiatric disorders, including dementia. All participants underwent a neuropsychological assessment and participated in MRI examination. We confirmed that all subjects showed no abnormalities in MRI, including apparent brain atrophy, silent brain infarction and pathological subcortical white matter lesions. All participants gave informed consent; the Shimane University Medical Ethics Committee approved the study.

All participants were assessed using neuropsychological test batteries including the Mini Mental State Examination (MMSE; Folstein, Folstein, & McHugh, 1975), the Frontal Assessment Battery (FAB; Dubois, Slachevsky, Litvan, & Pillon, 2000), the Kohs' Block Design Test (KBDT), the Verbal Fluency Test (vegetable for the semantic category and “shi” for the phonemic letter), the Self-rating Depression Scale (Zung, 1965), and the Apathy Scale (Okada, Kobayashi, Aoki, Suyama, & Yamagata, 1998; Starkstein et al., 1992). We performed correlation analyses between these neuropsychological indices and age, and the statistical threshold was corrected using the Bonferroni method (0.05/7 = 0.007).

Imaging data were acquired using a Siemens AG 1.5 T scanner. Twenty axial slices parallel to the plane connecting the anterior and posterior commissures were measured using a T2*-weighted gradient-echo spiral pulse sequence (repetition time = 2000 msec, echo time = 46 msec, flip angle = 90°, scan order = interleave, matrix size = 64 × 64, field of view = 256 × 256 mm², isotropic spatial resolution = 4 mm, slices = 20, slice thickness = 5 mm, gap = 1 mm). All participants underwent a 5-min rs-fMRI scan after only being instructed to remain awake with their eyes closed. After the functional scans, T1-weighted images of the entire brain were measured (192 sagittal slices, repetition time = 2170 msec, echo time = 3.93 msec, inversion time = 1100 msec, flip angle = 15°, matrix size = 256 × 256, field of view = 256 × 256 mm², isotropic spatial resolution = 1 mm).

A problem with using imaging techniques to determine functional connectivity is brain atrophy, which occurs in some elderly subjects. As the atrophy can reduce blood flow (Schlageter et al., 1987), we needed to correct for the effects of atrophy on functional connectivity. To obtain gray matter probability maps for atrophy correction in functional imaging, T1-weighted magnetic resonance images underwent an diffeomorphic anatomical registration through exponentiated lie algebra (DARTEL; Ashburner, 2007) using Statistical Parametric Mapping version 8 (SPM8, www.fil.ion.ucl.ac.uk/spm/). DARTEL is an efficient diffeomorphic framework for registering images; the optimization was performed using a Levenberg–Marquardt strategy. First, the gray was segmented from the white matter in the images for each subject. The gray matter images were warped to the Montreal Neurological Institute template space. Second, the deformations that best aligned the images were estimated by iteratively registering the images with their average. Third, using the deformations from the previous step, spatially normalized and smoothed Jacobian scaled gray matter images were generated. The smoothed images represent the regional volume of gray matter tissue (range = 0–1). The images were resampled at a voxel size of 3 × 3 × 3 mm³.

SPM8 was used for preprocessing. The first 10 functional images of each subject were discarded for magnetic field stabilization. The remaining 140 functional images were realigned to remove any artifacts from head movement and to correct for differences in image acquisition time between slices. Next, the functional images were normalized to the standard space defined by a template T1-weighted image and resliced with a voxel size of 3 × 3 × 3 mm³ to agree with the gray matter probability maps. Then spatial smoothing was applied with the FWHM equal to 8 mm, and temporal smoothing was performed by using the band-pass filter: 0.01–0.08 Hz.

After the preprocessing, we performed spatial ICA using Group ICA of the fMRI Toolbox (GIFT, icatb.sourceforge.net/) to isolated independent component maps (Calhoun, Adali, Pearlson, & Pekar, 2001). ICA is a data-driven, multivariate signal processing approach. In the applications of ICA to fMRI data, the four-dimensional signals are usually modeled as linear mixtures of unknown and spatially independent components (van de Ven, Formisano, Prvulovic, Roeder, & Linden, 2004). This method consisted of three major steps: data reduction, estimation of independent components, and back-reconstruction. First, single-subject data were combined together, and the aggregated data were reduced by principle component analysis to decrease the computational load. Second, ICA for the group data was performed using the Infomax algorithm. The number of independent components to be extracted from the data was set at 20, which is consistent with the minimum description length estimate. Third, the individual spatial maps and time courses of the independent components were reconstructed by using the aggregated components and the results from the data reduction. In the ICA analysis, the signal observed at a given voxel is assumed to be the sum of the contributions of all the independent components (McKeown et al., 1998). The special distributions of voxel values in the ICs are statistically independent from each other to a very high degree, and the degree of contribution reflects the functional connectivity of the IC network. The GIFT can provide contributions from all the voxels to each IC as whole-brain images of z scores (Beckmann, DeLuca, Devlin, & Smith, 2005). The maps were averaged to produce each component map used in the random-effects analyses. Here, the z scores reflect the degree to which a given voxel's time series correlates with the time series corresponding to the specific independent component. This process was performed to visualize which brain regions were statistically significant for each component; the threshold of the one-sample t test was p < .05 with family-wise error (FWE) rate correction for searching the whole brain. Each of the spatial components was manually inspected for the presence of obvious artifacts that get effectively separated out by ICA (e.g., motion, ventricles). We retained eight RSNs.

The ICA algorithm assumes that the time courses of cortical areas within one component are synchronous. Although the components are spatially independent, significant temporal correlations can exist between them (Sakoglu et al., 2010; Jafri, Pearlson, Stevens, & Calhoun, 2008). The intercomponent correlation reflects functional connectivity between the RSNs. The temporal correlations between all pair-wise combinations [number: 8 (8 − 1)/2 = 28] were calculated for each participant. We converted the correlation coefficient into a z score by Fisher's transfer and regarded the resultant z score as our measure of internetwork functional connectivity. Using a one-sample t test, we tested whether the z score of each pair differed from baseline (z = 0). The statistical threshold was corrected by the Bonferroni method (0.05/28 = 0.0018).

First, we performed whole-brain correlation analyses between each functional connectivity and age. Similar correlation analyses between z scores for connectivity and explanatory factors have been frequently used to analyze task performance (Garcia-Garcia et al., 2012; Jin, Pelak, & Cordes, 2012) and clinical symptoms (Assaf et al., 2010; Rocca et al., 2010; Greicius et al., 2007). The tests were masked by the map of each RSN (FWE corrected p < .05). To determine whether the aging effects resulted from underlying gray matter atrophy, we conducted reanalyses after adding the voxel-wise gray matter probability map as a covariate using the biological parametric mapping (BPM; Casanova et al., 2007). The BPM allows solving a general linear model by incorporating information obtained from other modalities (i.e., structural data). Here we incorporated a gray matter probability map for the voxel-wide SPM analysis. Therefore, we could investigate the aging effects after excluding the effects of brain atrophy. The statistical threshold was set at uncorrected p < .001 at the voxel level and corrected p < .05 at the cluster level. Then, to investigate the aging effects on the internetwork functional connectivity, we performed correlation analyses between age and connectivity. These analyses were performed on the significant 18 functional connectivities among the RSNs; therefore, the statistical threshold was set at 0.05/18 = 0.0028.

Whole-brain correlation analyses between the functional connectivity of each component and the neuropsychological tests were also performed. In these analyses, the voxel-wise gray matter probability map was regressed out as a covariate using the BPM. The statistical threshold was set at uncorrected p < .001 at the voxel level and p < .05 at the cluster level. Furthermore, we investigated the correlation between the internetwork functional connectivity and neuropsychological test scores.

ICA is a useful method for detecting networks because, when using ICA, it is not necessary to develop a prior hypothesis (van de Ven et al., 2004). However, there are indications that ICA cannot discriminate between inter- and intraregion functional connectivity. To separate the alteration of inter- and intraregion functional connectivity, we performed additional seed-based analyses, in which seeds that were the main regions showed significant contributions for eight independent component networks. The criterion for selecting ROIs was set at t > 20 (Table 2), and 24 ROIs were selected. The range of ROI was defined as a sphere with a 6-mm radius. We used CONN toolbox (web.mit.edu/swg/software.htm) for the seed-based analyses. The correlation coefficients between each ROI time series were transferred by Fisher's z transformation. We adopted t tests for comparing with the baseline (z = 0) and corrected the statistical criteria by Bonferroni method (p < .00018). For significant connections, we performed partial correlation analyses with age and cognitive indices. In the analyses, average gray matter probabilities of the ROIs were considered to be nuisance covariates, and the effects of atrophy were removed. The statistical threshold of the partial correlation analyses was first set at p < .0007 (corrected by Bonferroni method). However, connections that exceeded this strict criterion were not observed. Therefore, we reported connections that exceeded the uncorrected criteria (p < .05).

Figure 1 shows the distribution of age. The neuropsychological data were summarized in Table 1. The performance was declined with aging for most cognitive indices. The scores of MMSE, FAB, KBDT, and phonemic Verbal Fluency Test were negatively correlated with age (−.36 ≦ rs ≦ −.40, ps < .002). The mood indices including Self-rating Depression Scale and Apathy Scale were not affected by age.

Results of the ICA are depicted in Figure 2, where each map shows the result of one-sample t test for the back-reconstructed individual IC patterns (FWE corrected p < .05). We obtained eight IC patterns, which represent functionally relevant RSNs. The IC pattern of Figure 2A includes the medial pFC (BA 9/10/32), left dorsolateral pFC (BA 9/10), peripheral ACC (BA 24/32), posterior cingulate cortex (BA 23/31), precuneus (BA 7), bilateral inferior parietal lobe (IPL; BA 39), and middle temporal gyrus (MTG; BA 21). This pattern corresponds to the DMN. Figure 2B shows an IC pattern including mainly the left lateral pFC (BA 8/9/10/44/45/46/47) and superior parietal lobe (SPL; BA 7/40), which called the left frontoparietal network. In contrast, the IC pattern shown in Figure 2C includes mainly the right lateral pFC (BA 8/9/10/44/45/46) and SPL (BA 7/40), which called the right frontoparietal network. In Figure 2D, the IC pattern covers the peripheral and dorsal ACC (BA 24/32) and bilateral insula (BA 13). This pattern corresponds to the salience network, which was first proposed by Seeley et al. (2007). The IC pattern of Figure 2E coincides with the visual cortex (BA 17/18/19). The bilateral temporal cortices (BA 21/22) are the auditory network shown in the IC pattern of Figure 2F. The IC pattern of Figure 2G includes the bilateral primary motor cortex (BA 4), premotor cortex (BA 6), supplementary motor cortex (BA 6), and primary/secondary somatosensory cortex (BA 1/2/3/5/7). This pattern constitutes the sensory–motor network. The detail information of each IC was summarized in Table 2.

Figure 3 shows relationships among the RSNs; each network had some significant connection with the other networks. The bilateral frontoparietal networks strongly connected with each other and were negatively correlated to the posterior networks, which include the sensory–motor, auditory, visual, and cerebellum network. In contrast to the frontoparietal networks, the salience network that consisted of the insula and dorsal ACC positively connected with the posterior networks. The DMN had positive connections with both the frontoparietal and posterior networks. The sensory–motor, auditory, and visual networks positively connected with each other.

The most notable finding is that the connectivity of the bilateral insula and ACC (both being parts of the salience network) was negatively correlated with age. Figure 4A shows the negative correlation map and the scatter plot for the functional connectivity within the salience network versus age. In the visual network, the functional connectivity of the bilateral parahippocampal gyri was also modulated by age. As with the salience network, the functional connectivity decreased with age (Figure 4B). We did not observe a significant correlation between the DMN connectivity and age. However, when a less stringent criterion (voxel level p < .001) was applied, a significant correlation was obtained, that is, the posterior cingulate cortex (PCC) in the DMN showed negative correlation with age (Figure 4C). In cases without the correction for the gray matter probability, the aging effect on the DMN did not reach the severe criteria. No other RSN showed a significant correlation with age.

Furthermore, we investigated whether the internetwork correlations are modulated by age. This analysis was performed for the network pairs that showed a significant connection (see Figure 3). Figure 5 shows that the internetwork correlations of the salience-visual, salience-auditory, and default mode-visual networks were significantly decreased with age (rs < −.37, ps < .002). There is no effect of aging on other internetwork correlations (Tables 3 and 4).

The correlation analyses between neuropsychological scores and resting state functional connectivity were performed with age and gray matter probability map as nuisance covariates (Figure 6). The functional connectivity of the left insula (xyz = [−36 26 1], t = 3.72, cluster size = 11) and dorsal ACC (xyz = [−3 11 40], t = 4.37, cluster size = 12) in the salience network was significantly correlated with the KBDT and FAB scores, respectively (ps < .001). There was no significant correlation between the inter-network connectivity and the cognitive scores.

We also performed ROI-based analyses. Twenty-four regions that displayed robust contribution to each network (t > 20 in Table 2) were defined as ROIs. On the basis of the correlation between each time series of the ROIs, the functional connectivity among the main regions that consisted of eight networks is shown in Figure 7. Of the 276 possible connections, 71 connections showed significant differences from the baseline. Moreover, regions within the same independent network had higher connectivity with each other, and regions consisting of the DMN had moderate connectivity with other network regions. We conducted a partial correlation analyses between ROI-based functional connectivity, age, and cognitive indices (Table 5). The connectivity of the MPFC–left IPL, left IPL–left MTG, left INS–left INS, and SMA–right primary somatosensory area was negatively correlated with age. The MMSE score was positively correlated with the connectivity between the PCC and left DLPFC, whereas it was negatively correlated with the connectivity between the bilateral frontoparietal networks. The connectivity of the left INS–right TP and right MTG–right TP increased with higher FAB scores. The KBDT score showed a positive correlation with the connectivity between the PCC and ACC and a negative correlation with the connectivity between the right INS and SMA. The categorical Verbal Fluency Test score showed a negative correlation with the connectivity of the PCC–left IPL and a positive correlation with the connectivity between bilateral INS. The connectivity of the PCC–right PHG and right INS–right TP increased with higher phonemic Verbal Fluency Test scores.

Human cognitive decline with aging is likely to reflect age-related changes in the structure and in the neurotransmitter system of the brain. The current study investigated the relationship between aging and RSNs and its implication for age-related cognitive decline in a fairly large sample of healthy adults. Aging altered functional connectivity of the salience, visual, and DMNs under a resting state, and some internetwork connectivity also decreased with aging. Specifically, the disruption of the salience network connectivity was associated with the decline of cognitive function.

According to previous research on the effect of aging on the DMN (Tomasi & Volkow, 2011; Biswal et al., 2010; Koch et al., 2010; Esposito et al., 2008), the connectivity of the PCC, which is the central region of the DMN, was decreased by aging. It is well known that aging is associated with atrophy of the gray matter (Good et al., 2001). Our finding is important in that the altered DMN connectivity was observed after controlling for the individual degree of atrophy, suggesting the effect of aging on connectivity is not simply because of brain atrophy. The current study showed that not only the connectivity within the DMN changed but also the internetwork connection (DMN–visual network) was also decreased by aging. The DMN act as a hub of brain networks (Guye, Bettus, Bartolomei, & Cozzone, 2010). Therefore, aging might attenuate the function of the DMN as the network hub.

Some studies have emphasized the altered connectivity of the DMN in elderly people. However, our findings indicate that other networks are also affected by aging. One such network was the salience network that mainly consists of the insula and ACC first described by Seeley et al. (2007). Since then, a number of studies have confirmed the salience network using seed-based analyses (Li, Qin, Zhang, Jiang, & Yu, 2012; Woodward, Rogers, & Heckers, 2011) and ICA (Thomason, Hamilton, & Gotlib, 2011; Weissman-Fogel, Moayedi, Taylor, Pope, & Davis, 2010; Habas et al., 2009). We revealed that the functional connectivity of the salience network showed negative correlations with age, and the connections with the visual and with the auditory networks decreased. Furthermore, the functional connection of the salience network was strongly correlated with the cognitive decline even after correction for atrophy and age. These results were replicated in the seed-based analyses. Our evidences indicate that the salience network is one of important components for understanding aging effects on brain function as well as the DMN. Seeley and his colleagues (Seeley et al., 2007) proposed that the salience network is well suited for identifying the most homeostatistically relevant event among myriad inputs and integrating highly processed sensory data with visceral, autonomic, and hedonic markers so that the organism can decide what to do next. The insula and ACC are specialized modules for sympathetic efference and interoceptive feedback (Critchley, 2005) and coactive in response to varied forms of salience, including the emotional dimension of pain (Price, 2000), empathy (Singer & Lamm, 2009), and social rejection (Onoda et al., 2009, 2010; Eisenberger, Lieberman, & Williams, 2003). Our evidence that the salience network was positively correlated with the multimodal sensory networks strengthens the integrating function of the salience network.

Furthermore, it is important that FAB and KBDT showed strong correlations with functional connectivity of the salience network. FAB is an assessment battery of frontal lobe functions, and KBDT is an intelligence test that measures visuospatial skills, reflective of the functioning of parietal and frontal lobes. Similar to the findings of the current study, it has been reported that aging decreases FAB (Iavarone et al., 2011) and KBDT (Rozencwajg et al., 2005) scores. However, the network that showed strong correlations with these indices was not the bilateral frontoparietal network, but the salience network. Dosenbach and his colleagues have proposed an interesting hypothesis that is known as the “dual-network” hypothesis (Dosenbach et al., 2007). They called the salience network the cinguloopercular network, and its distribution corresponded to the salience network in our study. On the basis of differences in connectivity and activation profiles, they argued that frontoparietal and salience networks support distinct functions; adaptive control in the case of the frontoparietal network and stable set maintenance in the case of the salience network. According to the dual-network hypothesis, the lower FAB/KBDT scores observed in the current study might reflect the disrupted salience network through decreased motivation. This interpretation is supported by neuropsychological theories (Jones, Ward, & Critchley, 2010; Devinsky, Morrell, & Vogt,1995).

Our ICA analysis indicated that the visual network included the parahippocampal gyrus. An anatomical connection between the visual cortex and hippocampal gyrus has been reported (Yeterian & Pandya, 2010). Our data also suggested that these regions might be functionally connected with each other. The visual network in elderly people showed a decreased connectivity within the network and between the networks. There are two possible explanations for the decreased connectivity of the visual network. One is a deficit of sensory processing. Recent studies examined aging effects on brain activation measured with task-related fMRI and found increased frontal and decreased occipital activation (Cabeza et al., 2004). These results are interpreted such that the former is attributed to functional compensation and the latter to insufficient sensory processing. The visual network disruption with aging in the current study might account for the insufficient sensory processing in an aged individual. The other possible explanation is a memory deficit. The decreased connectivity in the visual network was also observed in the parahippocampus, which might contribute to memory dysfunction in elderly people (Henson, 2005). Furthermore, in the ROI-based analyses, the decreased connectivity between the right parahipocampal gyrus and PCC was correlated with poor performance in the verbal fluency task. These results suggest that the connectivity of the parahipocampal gyrus with default and visual networks decrease with aging and that the decreased connectivity might cause aging-related memory deficits.

In the current study, we observed decreased functional connectivity within and between brain regions in elderly participants. Recently, remarkable evidence was reported (Allen et al., 2011). Allen et al. introduced a new approach that optimizes sensitivity, with which they identified the effects of aging on the functional connectivity. The rs-fMRI data were divided into 75 components by ICA; the RSNs were evaluated in terms of intra- and internetwork connectivity. They showed decreased functional connectivity for both intra- and internetworks in elderly people. Such decreased functional connectivity because of aging is associated with the geriatric change of gray matter (Galluzzi, Beltramello, Filippi, & Frisoni, 2008), corresponding to the idea that structural substrates underlie brain function. However, as suggested in the current study, the cognitive decline was related to the functional connectivity even after controlling for gray matter probability. As some articles reported similar evidence (Allen et al., 2011), it seems that gray matter alone cannot explain functional connectivity change.

Functional neuroimaging can be used to detect distinct patterns of network disruption depending not only on aging but also on the type of dementia. In Alzheimer's disease and mild cognitive impairment, early DMN functional disruption involves the medial-temporal lobe and the posterior cingulate cortex/precuneus (Qi et al., 2010; Sorg et al., 2007; Greicius, Srivastava, Reiss, & Menon, 2004), subsequently worsening and extending to the lateral parietal and the medial frontal regions with increasing disease severity (Zhang et al., 2010). On the other hand, frontotemporal dementia showed the disruption of the salience network, and the combination of the DMN and the salience network allowed for clear discrimination between Alzheimer's disease and frontotemporal dementia (Zhou et al., 2010), although the reliability remains to be tested in independent samples of patients. Furthermore, it is reported that functional connectivity patterns in dementia with Lewy bodies were distinct from those seen in Alzheimer's disease (Galvin, Price, Yan, Morris, & Sheline, 2011). Functional connectivity may improve the discrimination of various dementias and cognitively healthy individuals.

Certain limitations of the current study are described below. First, although previous studies have consistently reported aging effects on the DMN (Allen et al., 2011; Tomasi & Volkow, 2011; Koch et al., 2010; Damoiseaux et al., 2008; Esposito et al., 2008), we found this effect only when a more liberal threshold was adopted, and moreover, no significant correlation between the DMN and cognitive performance was observed in this study. Some studies (Allen et al., 2011; Damoiseaux et al., 2008) demonstrated aging effects, even after controlling for the gray matter effects. However, in the current study, the aging effects on the DMN were weak, regardless of the correction for the gray matter. The absence of a robust aging effect on the DMN may not be because of the correction for atrophy. The discrepancy may have been caused by differences in the age range of the subjects. Our subjects were between 30 and 80 years, whereas previous studies have included subjects in the 20s. According to Tomasi and Volkow (2011), the functional connectivity of the DMN is notably higher in people aged about 20 years compared with older people. The absence of subjects in the 20s in the current sample might explain the weak aging effects on the DMN. Second, it is not easy to interpret the difference in connectivity within and between networks. Our ICA results showed a decrease in intranetwork connectivity with age including a decrease in either the DMN or the salience network that was disrupted as a result of aging. However, it is not clear whether the decreased internetwork connectivity was associated with the decreased intranetwork connectivity. We also performed seed-based analyses and found an aging related decrease in connectivity between the ROIs within the DMN and salience network. These results corresponded with the results of the ICA. Conversely, there was an aging related decrease in internetwork connectivity between the left inferior parietal lobule (belong to DMN) and the left MTG (belong to temporal network). This result did not coincident with that of the ICA. Therefore, ICA-based data on internetwork connectivity did not necessarily match those of the seed-based analysis. Therefore, it is suggested that, although internetwork connectivity is a useful marker of aging, its interpretation should be carefully considered. Moreover, temporal correlations among components of ICA were not as large as those between regions within a component, nevertheless, internetwork connectivity in the current study showed robust aging effects, suggesting that not only ROI-based measures and within-network connectivity measures but also internetworks connectivity measures were important when considering the whole brain network. Third, ICA could not differentiate intraregional from interregional connectivity, because result maps reflected the contributions of each component. For instance, because the connectivity between the bilateral insula was negatively correlated with age in the ROI-based analysis, we could partially confirm the aging effect on the interregional connectivity of the salience network. However, we also need to investigate the intraregional connectivity using such measures as integrated local correlation (Deshpande, LaConte, Peltier, & Hu, 2009).

In summary, our evidence suggests that aging disrupts not only the DMN but also the salience network. The disrupted salience network seems to be involved in the cognitive decline of elderly people. Notably, these relationships are independent from the gray matter change. Our evidence indicates that the functional connectivity is a useful tool for exploring cognitive decline with aging.

Reprint requests should be sent to Shuhei Yamaguchi, Department of Neurology, Shimane University, 89-1, Enya-cho, Izumo, Shimane, 693-8501, Japan, or via e-mail: yamagu3n@med.shimane-u.ac.jp.