Aberrant intra- and inter-network connectivity architectures in Alzheimer’s disease and mild cognitive impairment

Alzheimer’s disease (AD) patients and those with high-risk mild cognitive impairment are increasingly considered to have dysfunction syndromes. Large-scale network studies based on neuroimaging techniques may provide additional insight into AD pathophysiology. The aim of the present study is to evaluate the impaired network functional connectivity with the disease progression. For this purpose, we explored altered functional connectivities based on previously well-defined brain areas that comprise the five key functional systems [the default mode network (DMN), dorsal attention network (DAN), control network (CON), salience network (SAL), sensorimotor network (SMN)] in 35 with AD and 27 with mild cognitive impairment (MCI) subjects, compared with 27 normal cognitive subjects. Based on three levels of analysis, we found that intra- and inter-network connectivity were impaired in AD. Importantly, the interaction between the sensorimotor and attention functions was first attacked at the MCI stage and then extended to the key functional systems in the AD individuals. Lower cognitive ability (lower MMSE scores) was significantly associated with greater reductions in intra- and inter-network connectivity across all patient groups. These profiles indicate that aberrant intra- and inter-network dysfunctions might be potential biomarkers or predictors of AD progression and provide new insight into AD pathophysiology.

Subjects. The subjects were recruited from two sources: as outpatients from the Chinese PLA General Hospital or through a website advertisement (http://www.301ad.com.cn, Chinese version). All subjects met the identical methodological stringency criteria, and comprehensive clinical details were described in our previous studies; additional details regarding participant selection and exclusion for this data set can be found elsewhere [44][45][46][47] and Part I of the supplemental material. Briefly, the recruited AD patients were diagnosed using the National Institute of Neurological and Communicative Disorders and Stroke and the Alzheimer's Disease and Related Disorders Association criteria for probable AD. The enrolled MCI patients fulfilled the diagnostic criteria described by Petersen et al. 2 . At the same time, our AD and Scientific RepoRts | 5:14824 | DOi: 10.1038/srep14824 MCI patients also met the core clinical criteria of the new diagnostic criteria for probable AD dementia and MCI due to AD 49,50 . Exclusion criteria included significant neurological or psychiatric illness that can influence cognitive functions as well as significant unstable systemic illness or organ failure. In addition, patients with a metallic foreign body were also excluded from the study for security and imaging quality control reasons. No subjects were treated with any medication that could have influenced their cognition during the data collection. Each subject was right-handed and underwent a battery of neuropsychological tests: the Mini-Mental State Examination (MMSE), Auditory Verbal Learning Test (AVLT), Geriatric Depression Scale, Clinical Dementia Rating (CDR) and Activities of Daily Living (ADL) scale.
Briefly, after excluding subjects with large head motions (see the criteria in data preprocessing), 89 subjects-35 AD patients, 27 MCI subjects, and 27 age-and gender-matched NC subjects-were included for further analysis. Demographic and neuropsychological details for the subjects are shown in Table 1 and can be found in our previous studies 44-47 . fMRI data acquisition and preprocessing. As was previous noted, the MRI scans were performed at the Chinese PLA General Hospital, Beijing, China, with a 3.0 T GE MR system (GE Healthcare, USA) using a standard head coil. During the scanning, the subjects were instructed to keep their eyes closed and relax; comfortable foam padding was used to minimize head motion, and ear plugs were used to reduce the scanner noise. Before the resting fMRI data were collected, T2-weighted images Figure 1. Group effects by one-way ANOVA for and the relationship between integrity connectivity and MMSE scores. Bar graphs show the differences in the mean Z scores for the affected regions among the three groups. The changed regions were anchored in the right motor cortex (rMC) (A), the right posterior intraparietal sulcus (rpIPS) (B) and the bilateral primary visual (V1) (C,D). The Z scores for the bilateral V1 and rpIPS were higher in the NC group than in the AD and MCI groups, and the rMC of the NC group was greater than that of the MCI group only. The NC group is indicated by black rectangles, the MCI group by grey rectangles and the AD group by white rectangles. The error bars represent the standard error of each subgroup. The scatter plots show the relationship between the mean Z scores of rpIPS and MMSE scores in the MCI patients (r = − 0.418, p = 0.03) (E). were collected and evaluated by two senior radiologists. Resting-state fMRI data were acquired using an echo planar imaging (EPI) sequence with repetition time = 2000 ms, echo time = 30 ms, flip angle = 90°, matrix = 64 × 64, field of view = 220 mm × 220 mm, slice thickness = 3 mm and slice gap = 1 mm. Each volume was composed of 30 axial slices, and each functional run lasted for 6 minutes and 40 seconds.
The data were preprocessed using the same steps as those in our previous studies using the in-house Brainnetome fMRI toolkit (Brat, www.brainnetome.org/brat) based on statistical parametric mapping (SPM8, http://www.fil.ion.ucl.ac.uk/spm). These steps were: (1) slice-timing with reference slice = 2, (2) realignment to the first volume, (3) normalization to a standard EPI template and reslicing to 2 × 2 × 2 mm cubic voxels, (4) de-noising by regressing out multiple effects, i.e., the six motion parameters, the constant, the linear drift and the mean time series of all voxels within the white matter and cerebrospinal fluid, and (5) temporal filtering (0.01-0.08 Hz) to reduce noise. The data were not further smoothed for we intended to investigate the connectivity patterns of the prior defined seed regions. Group differences in head motion followed those in Van Dijk's study 51 , and no significant differences in head motion were found among the three groups (Table 1) 47 .

Connectivity analysis.
The five RSNs used in this study have been well investigated in previous studies [36][37][38][39]52 . To maintain consistency with these prior studies, the seed regions in the present study were defined a prior based on a study by Dr. Ances; specifically, thirty-six spherical (6 mm radius) regions of interest (ROIs) that represented the five RSNs(DMN, DAN, SAL, CON, and SMN) were obtained using the Brainnetome fMRI toolkit. As in previous studies [36][37][38][39]52 , the SMN included the primary auditory, primary visual, and somatomotor cortices. The Montreal Neurological Institute coordinates of the 36 ROIs are presented in Table S1, and the individual ROIs that were displayed on the brain surfaces are shown in Figure S1 in the supplemental material.
The representing mean time series was estimated by averaging the time series of all voxels in this ROI. The Pearson's correlation coefficients were computed between each pair of ROIs for each subject. Fisher's r-to-z transformation was applied to obtain Z scores and improve the normality of the correlation coefficients. Then, the connectivity pattern of these ROIs was investigated at the following three levels: (1) Integrity: for each node, the integrity Z score was defined as the sum of connection strength, that is, , where , Z i j is the Z score between the i th and j th ROI. This measure is equivalent to "degree centrality" in the graph theory.
(2) Network: for each of the five RSNs, the intra-network strength was defined as the mean connection strength of the ROIs in the same network, i.e., where n X is the number of ROIs within a specific subnetwork X and k is a range from 1 to 5 that represents the five subnetworks. For each pair of subnetworks, the inter-network connectivity strength was defined as the mean strength of all of the possible connections, that is, , where X and Y represent the subnetworks of the five selected RSNs.  Statistical analysis. To assess the statistical significance, the Z scores at each level were entered into one-way analyses of variance (ANOVAs) with group as a factor (3 groups: NC, MCI, AD) after age and gender effects were regressed out using a general linear model; the statistical significance was P < 0.05. Considering that we have performed many times comparisons among the three groups, we performed a 10,000 times random permutations to test if the identified altered connections are really significant. For all the identified impaired connectivity, the significant effects were assessed by post hoc two-sample, two-sided t-tests of NC versus MCI, NC versus AD and MCI versus AD (P < 0.05).
To investigate the relationship between functional connectivity and cognitive ability, we also explored the Pearson's correlations between the MMSE scores and the Z scores at each level in the MCI, AD and MCI plus AD groups. Because these relationships were exploratory in nature, we used a statistical significance level of P < 0.05 (uncorrected).

Results
Group differences at the three levels. For each group, a 36 × 36 functional connectivity matrix was computed. In the NC group, the majority of strong positive functional connectivities were within each network, and most negative correlations were between different networks ( Figure S2A-C). A similar pattern was observed in the MCI and AD groups, except that the intra-network and inter-network correlations were decreased ( Figure S2D-F).
At the integrity level, significant group differences were anchored in the right motor cortex (rMC), the bilateral primary visual (V1), and the right posterior intraparietal sulcus (rpIPS) ( Table 2, Fig. 1). Post hoc analysis showed that the integrity connectivity of these regions was reduced in the AD and MCI groups compared with the NC, but there was no significant alteration between the AD and MCI groups ( Fig. 1).
At the network level, the composite Z scores of the DMN, CON and SMN showed significant differences in the severity of the cognitive impairments among the groups (Table 3, Fig. 2A). Compared with those for the NC group, the Z scores within the DMN and SMN showed significant decreases in the AD and MCI groups ( Fig. 2A). Particularly, the Z scores for CON showed a slight increase in the MCI group but a sharp decrease in the AD group ( Fig. 2A). For the inter-network pairs, only the DAN-SMN connectivity showed a significant decrease in the AD and MCI groups ( Table 4, Fig. 3A).
As Fig. 4 shows, a large number of changed intra-network and inter-network functional connectivities were identified by one-way ANOVA after age and gender effects were controlled using linear regress ( Fig. 4A; details of the statistical values can be found in Table S2). The most significantly affected pairs  were mainly distributed between the DMN and other RSNs and between the DAN and SMN (Fig. 4A).
Post hoc analysis strengthened these manifestations (Fig. 4C-E). Specifically, the most significant alterations were mainly distributed between the DAN and the SMN in the MCI subjects in comparison with the NC individuals (Fig. 4C). For all the identified impaired connectivities, random permutation tests indicate that these findings are significant with P < 0.05 (Figs 1-4. Tables 2-4, S2).

Correlations between altered connectivity and MMSE scores.
At the integrity level, the connectivities of the laPFC and the rpIPS demonstrated significant correlations with the MMSE scores in the MCI patients (Table 2, Fig. 1). At the network level, the affected intra-network interactions between  the DMN and CON and the DAN and CON were significantly correlated with the MMSE scores in the AD and MCI patients (Tables 3-4, Figs 2B and 3B). Our results also demonstrated that 9 connectivity pairs showed positive correlations with MMSE scores and that 4 connectivity pairs showed negative correlations with MMSE scores in the identified impaired connectivity pairs ( Fig. 4B; for details, please refer to Table S2).

Discussion
In the present study, widespread impaired functional connectivity patterns including intra-network and inter-network disconnections that could have been the basis for cognitive impairment were identified in the MCI and AD patients. Based on the integrity-level analysis, the differences were anchored in the right motor cortex (rMC), the bilateral primary visual (V1) and the right posterior intraparietal sulcus (rpIPS) in the AD and MCI subjects ( Table 2, Fig. 1). The connectivities within the DMN and SMN were significantly decreased in the MCI and AD groups, and the connectivities within the CON increased slightly in the MCI group whereas it showed a sharp decrease in the AD group ( Fig. 2A). Connectivity analysis of both the inter-network and the ROI pairs showed that the DAN-SMN connectivity was impaired significantly in the patient groups, especially in the MCI individuals. Among the identified impaired connectivities, some demonstrated significant correlations with cognitive ability as assessed by the MMSE scores ( Fig. 4B and Table S2).

Impaired intra-network connectivity within the RSNs. Consistent with previous studies, the
present study demonstrated that the default mode network (DMN) was one of the most affected network in the AD 21 and MCI 34 subjects ( Fig. 2A,B). The DMN plays a key role in cognitive processes, especially in episodic memory processing 53,54 , and the impaired performance of episodic memory is one of the core features of a diagnosis of AD 6 ; this could be one of the main reasons that impaired functional connectivity within the DMN has been frequently identified in AD and MCI patients using multiple imaging techniques 6,19,36 . In addition, amyloid plaques, which have been considered the main pathophysiological process of AD to date, were found to be preferentially deposited in regions of the DMN 55 and to cause the impaired resting-state fMRI connectivity 40,[56][57][58] . Based on the DMN's core role in modulating daily cognition, functional deficits in the DMN might contribute to AD pathology and might be potential biomarkers for distinguishing AD from MCI.
In the present study, the functional connectivity within the CON increased slightly in the MCI group, whereas it decreased sharply in the AD group ( Fig. 2A). As was described, the CON is crucial for active maintenance of and manipulating information in working memory and for rule-based problem solving and decision making, and it is related to executive function [59][60][61] . MCI patients showed subtle executive dysfunctions in higher-order activities, such as financial capacity 62 , but everyday abilities were preserved; hence the increased connectivity within the CON may reflect a coherent compensatory recruitment in MCI patients. Consistent with this, neuroimaging studies have suggested that the increased prefrontal activity (the main region of CON) reflects compensatory strategies used in performing cognitive tasks 63,64 . Clinically, AD patients lose the ability to manage their daily life activities, which could be because activities such as cooking, shopping, and driving are target-based problem-solving tasks that require the participation of the executive control brain regions 65 and executive dysfunctions are considered pervasive 66 . Hence, the decreased connectivity within the CON in AD patients might reflect the foundation of executive dysfunction.
An interesting finding was that our results demonstrated that the functional connectivity within the sensorimotor network (SMN) was impaired in the MCI and AD subjects (Figs 1 and 2A). The SMN, composed of the primary visual, auditory and somatomotor cortices, plays a role in receiving external signals, which could aid in perceiving the world, selecting the relevant information and determining the target; it then conveys the signals to the attention or control systems to induce reasonable responses. For many years, the SMN was usually thought to be relatively stable, and was seen as a reference network in studying AD and MCI 39,67 . Although a number of studies have indicated that the functional changes in the olfaction, hearing, visual, and motor systems (the major components of the SMN) might precede the onset of cognitive impairments, worsen as the disease progresses, and be strong risk factors for AD [68][69][70][71] , this impaired functional connectivity, together with the recognition that AD pathology will develop over many years, raises the exciting possibility that declines in specific primary daily functions may be early noninvasive biomarkers for AD. Even more provocatively, treating these daily symptoms may help to All affected ROI pairs for all 5 RSNs (blue for DMN, dark green for DAN, yellow for CON, green for SAL and pink for SMN), except for intra-network ROI pairs, were mainly distributed between the DMN and other RSNs and between the DAN and the SMN. (B) The correlations between the functional strength of the affected ROI pairs and the MMSE scores. The blue color represents the functional connectivity that shows positive correlations with the MMSE scores, and the red color represents negative correlations. (C) The differences in connectivity between the NC and MCI groups. (D) The differences in connectivity between the NC and AD groups. (E) The differences in connectivity between the MCI and AD groups. In the subfigure (C-E), the blue color indicates that the functional connectivity of the former group is stronger than that for the latter, and the red color indicates the reverse. For details, please refer to Table S2 in the supplemental material. delay or treat MCI/AD 72 . Our results provided additional evidence that the affected SMN should arouse attention not only as a reference stable system in studying AD and MCI.

Impaired inter-network connectivity between RSNs in AD and MCI.
Resting-state functional connectivity and network analysis provides a new tool for mapping large-scale function and dysfunction in the brain system 12,73 .Inter-network connectivity, especially between the DAN and the SMN, was impaired in the AD and MCI groups (Fig. 3A, Fig. 4A,C,D and Table S2), and this phenomenon prompted us to rethink the role of the interactions between the RSNs in understanding AD pathology and clinical performance. The dorsal attention network (DAN) is one of the networks that is associated with cognitive functions, and the multiple somatosensory integrations of sensory, motor and cognitive systems provide the signals for the organism to perceive and respond to its environment; that is, deficits in any of these components will lead to impaired function at the clinical level 66,72,74 . Often, decreased attention and sensory or motor declines are seen as signs of aging; clinically, the performance of these functions is poor in MCI individuals and even worse in AD subjects 75,76 . Our research prompted us to infer that the impaired connectivity within the SMN or reduced connectivity between the SMN and the DAN might be one of the reasons for the attention deficits in the early stages of AD and MCI. Additionally, convergence evidence has suggested that AD pathology develops over many years, raising the exciting possibility that daily cognitive impairments, even very slight (such as reduced sensory or motor ability) may be early, noninvasive markers for AD 72 .
In addition to the above-discussed SMN and DAN, the interactions among the CON, SAL and DMN were also impaired in the AD and MCI groups (Fig. 4A,D,E). Previous resting state fMRI studies have demonstrated aberrant functional connectivity of DMN, SAL and CON in AD 36,77 , aMCI 78 and even in normal cognition with amyloid burden 40,79 . Convergent evidence has suggested that these three brain systems were closely correlated and play particularly crucial roles in higher cognitive function 15,80,81 . Functional connectivity between the SAL and DMN is important for cognitive control 80,82 , and the SAL also plays a central role in switching between the CON and the DMN 80,82-84 . Our findings provide additional evidence to support the viewpoint that the CON, SAL and DMN are intrinsically well-organized in normal healthy subjects and that aberrations in these networks are the prominent features of functional deficits in AD 16,83,85 . By contrast, the DMN plays a core role in brain activity and other networks that are involved in succession, which suggests that AD pathology might spread from the DMN to the nearby networks, including those involved in visuospatial and executive function, and in other peripheral networks 86 . Therefore, we speculated that AD patients who show clinical attention deficits, memory impairment, executive dysfunction and other disabilities might be reflecting the integration dysfunction in the different brain networks 31,87 . Further discussion, limitation and future directions. It should be noted that for multiple metrics but the CON (Figs 2-4), MCI subjects tend to be similar to AD patients (and different from controls), with little differences between MCI and AD patients. This might because the severity of AD individuals were still mild (35 subjects, 25 with CDR = 1, 10 with CDR = 2) and the variability among patients was relatively high 46 . This might be the reason that we only found some significant correlations between the used measures and cognitive ability in the MCI and AD patients groups (Figs 1-3 and Table S2). Another possible reason is that the MCI subjects might include different subgroups, for not all the MCI individuals will convert to AD. In fact, for all the used data in the present study, 14 of the 27 MCI subjects were recruited to return for an examination and data collection after around one year (range: 10-16 months), one subject converted to AD 47 . Hence, a long time longitudinal study is needed for further investigation.
Notably, the impaired connectivity at all three levels (integrity, network and connectivity) were significantly correlated with performances on the Mini Mental State Examination (MMSE) neuropsychological tests (Figs 1, 2B, 3B and 4B). The MMSE involves multiple cognitive domains and reflects global brain function; it can be used as a tool for quantitatively assessing the severity of cognitive impairment and reflects the cognitive changes that occur during the progression of AD 88 . As discussed in our previous work, the correlations between MMSE and connectivity and network markers are relatively scant because the MMSE is a brief general screening tool 34 . However, based on the robust findings from previous studies 21,36,89 , including our own, we speculated that the correlations between MMSE scores and network functional connectivity indicate that abnormal brain function might be a feature that represents disease severity and could potentially be used as an early marker to distinguish patients from healthy subjects.
As is widely known, the whole brain network is complex, varied, and interrelated; the five networks in this study are merely a small part of this complex system, and thus, a whole-brain network analysis with finely defined regions is needed in the future. Second, cross-sectional research cannot dynamically observe changes in network patterns with disease progression 36 . Also many studies have found neuronal dysfunction and disconnection of brain network in normal cognition with amyloid burden 40,41,42,79 or APOE e4 carrier 90,91 . Hence, longitudinal studies combining multiple imaging measures (such as, fMRI, structure MRI, PET etc.) and genetic genotype are needed in the future to follow individuals from healthy to disease states and to different severity levels, exploring network-vulnerability interactions, and to study different subtypes of AD and MCI.

Conclusion
The novel finding of the present study is the relationship between disease severity and impaired intra-network and inter-network connectivity from global to fine-pair connectivity in AD and MCI patients. The results demonstrated a progressive alteration of network connectivity; at the early stage of the disease (MCI), sensorimotor and attention functions were involved, but with disease progression, the whole-brain function degeneration induced a wider range of inter-network impairments. All of these findings provide new insight into AD pathophysiology and suggest that altered network connectivity patterns may be useful for the preclinical determination of AD.