Aberrant Topological Organization and Age-Related Differences of Human Connectome in Subjective Cognitive Decline by Using Regional Morphology from Magnetic Resonance Imaging

Zhenrong Fu Beihang University Department of Biological Engineering: BeiHang University School of Biological Science and Medical Engineering Mingyan Zhao Xuanwu Hospital Yirong He Beihang University Department of Biological Engineering: BeiHang University School of Biological Science and Medical Engineering Xuetong Wang Beihang University Department of Biological Engineering: BeiHang University School of Biological Science and Medical Engineering Xin Li Yanshan University School of Electrical Engineering Guixia Kang Beijing University of Posts and Telecommunications Ying Han Xuanwu Hospital, department of neurology Shuyu Li (  shuyuli@buaa.edu.cn ) Beihang University https://orcid.org/0000-0002-3459-6821


Introduction
(2) feeling of worse performance than others of the same age group; (3) the MoCA score was in the normal range; (4) only one of the two memory tests(AVLT-D and AVLT-R) was abnormal(decline one Standard Deviation(SD) compared with NC); and (5) the CDR score was 0; (6) patients diagnosed with aMCI, AD, or other types of dementia were excluded. The inclusion criteria for NC were as follows: (1) NC had no reported memory decline; (2) the MoCA, AVLT and CDR scores were in the normal range; (3) without a history of diabetes. In addition, the normal range of neuropsychological tests were adjusted for age and education year. The exclusion criteria for all participants in this study were: (1) HAMD scores higher than 24, and the score of Center for Epidemiologic Studies depression scale higher than 21; (2) Hachinski ischemic scale in the abnormal range (higher than 4); (3) not right-handedness; (4) the executional, visual or auditory functions impaired; (5) cognitive function decline due to no-AD neurological diseases(e.g. brain tumor, brain injury, Parkinson disease, encephalitis, normal pressure hydrocephalus, multiple sclerosis or epilepsy); (6) individuals with a history of stroke; (7) subjects with a history of alcohol or drug abuse/addiction within two years(DSM-IV(Diagnostic and Statistical Manual of Mental Disorders)); (8)large-vessel disease (e.g. cortical and/or subcortical infarcts and watershed infarcts); (9) patients with any other systemic diseases or uncertainty prevents the completion of the project; (10) subjects with frequent head motion which may in uence the quality of MRI. Combined with the conceptual framework of SCD and the exclusion criteria in this study, and considering the completeness MRI data, 35 individuals with SCD and 42 demographically matched NC were included in this study. Three neurologists in XuanWu hospital with 8 to 28 years of experience completed this diagnostic procedure. The main demographical and neuropsychological characteristics of all subjects are summarized in Table 1. used. Then, the GM images were normalized to standard Montreal Neurological Institute (MNI) space. Thereafter, the voxel intensities of GM images were modulated to preserve regional volume information by Jacobian determinants derived from the normalization. Finally, the modulated GM images were isotropic smoothed with 8-mm full-width at half-maximum (FWHM) Gaussian kernel for further analyses.

Construction of morphometric networks
In order to construct the connection matrix, the two fundamental elements of network nodes and edges must be determined rst. In general, the nodes are de ned by different anatomical and/or functional atlases and the edges are de ned by the relationships between each pair of nodes.

Node de nition
In the study, the nodes represented by neuroanatomical structures. The automated anatomic labeling(AAL) atlas(Tzourio-Mazoyer et al. 2002) was used to parcel the brain into 90 cortical and subcortical structures de ned as nodes. Besides, the AAL atlas has been widely used in previous brain network analysis.

Edge de nition
sMRI can accurately show the anatomical structure of the brain, and the intensity of voxels within brain regions in modulated GM images can mediately represents the content of brain tissue, the distribution and density of neuronal cells. Therefore, the distribution of intensity values of voxel within a speci ed brain structure from MRI images can represent the morphological features of the brain structure ( Where P(x) and Q(x) are two intensity distributions from different brain regions.
Then, the KL divergence was converted as a similarity measure using the following formula: The KLS was normalized from 0 to 1, where 0 represents the two completely different distributions, in contrast is 1. Then, the morphological distributions from GM images were de ned as estimated probability density functions (PDFs) for each region through kernel density estimation (KDE) (Botev et al. 2010), which included all voxels in each region. In particular, the Gaussian kernel is assumed and the bandwidth is chosen automatically. In order to obtain unbiased PDFs, the number of the voxel in each region were used to correct the effect of volume. Besides, the PDFs were truncated distributions, which were restricted by the range of the intensity of each region. Finally, we calculated the KLS of all each pair of the brain regions and a 90 × 90 matrix was used to quantify the morphometric connection ( Fig. 1). Figure  divergence of their morphological distributions from GM images and converting the KL to similarity measure KLS. (5) The network metrics were calculated by GRETNA toolbox and the between-group differences were computed. (6) The age-related differences of network metrics in the two groups and the relationships between cognitive test and network metrics were computed.

Threshold selection
Because of the continuous weights contain more information between nodes ( Barrat et al. 2004) and the structural alterations in individuals with SCD were subtle, the weighted networks were used in this study to quantify the alterations of topological organization in individuals with SCD. Sparsity was de ned as the ratio of the number of existing edges divided by the maximum possible number of edges in a network (Kong et al. 2015). The sparsity approach can ensure there are same number of nodes and edges in all individual networks. Sparsity of 15% meant that only the 15% strongest edges were remained and 85% weaker edges were removed.Because there is no de nitive way to select a single sparsity value (He et al. 2008;He et al. 2007), a sparsity approach with the sparsity range from 15-45% and an interval of 1%, was used to reserve meaningful connections and remove redundant invalid connections. All sparsity threshold of differences of global topological properties between the two groups were compared. To preserve more than 90% nodes in all networks and ensure all individual networks exhibited small-world properties (He et al. 2007), the sparsity threshold was set as 30% (referred to median value of the sparsity range ) in further analyses, including the topological properties at nodal level, the rich-club organization, the hierarchical organization, the connections divided by anatomical distance and the age-related alterations of topological properties.

Graph Theoretical Characterization
The metric of integration in a network represents the capacity of the brain to integrate the information from different brain structures, such as the global e ciency (E glob ) and Characteristic Path Length (Lp) and the modularity are measures of segregation, which represents the biologically meaningful feature of the brain to enable highly specialized processing through densely interconnected communities of regions (Ferreira et al. 2019;Watts and Strogatz 1998;Sporns 2013). In addition, the measures of centrality degree centrality (DC) and betweenness centrality (BC) were combined to identify the hub regions.

Rich-club organization
We used a multivariate approach to identify the hub regions of the individual morphometric networks. First, we normalized the values of DC and BC of the networks ranges from 0 to 1, and averaging them across all participants; Then, the average of normalized DC and the average of normalized DC were added; Finally, the hub regions were de ned as the combine values at least 1 standard deviation higher than the mean. As the nodes were divided into hub regions and peripheral regions, the connections were classi ed into rich-club connections (the hub nodes link with each other), feeder connections (linking the hub nodes and peripheral nodes) and local connections (linking between peripheral nodes)(van den

Functional organization and Anatomical distance
The structural co-variance and fMRI functional connectivity showed a signi cant statistical relationship(Alexander-Bloch et al. 2013). To prove that the individual GM networks we built are biological meaningful, we have investigated the connectivity strength of the networks at functional organization level. To further characterize alterations in the global and local functional organization of brain networks in SCD individuals, a parcellation scheme including ve key functional modules (primary sensory, subcortical, limbic, paralimbic, and association areas.) was applied in this study (Supekar et al. 2009). In previous study, MCI targeted more middle-and long-distance functional connections, which the connections were de ned by the anatomical distance . In this study, the anatomical distance (de ned as Euclidean distance between stereotaxic coordinates of the centroids for two regions) of brain structures parceled by AAL atlas were computed, and the connections were classi ed into shortrange (lower than 45mm), middle-range (higher than 45mm and lower than 80 mm) and long-range (higher than 80 mm) , to investigate whether the GM networks are in uenced by the anatomical distance. Then, the inter-and intra-connections of the ve modules and the connections in each distance range were calculated.
All network analyses were performed on the GRETNA toolbox  and the results were visualized by BrainNet Viewer toolbox ).
2.6. Statistical analysis 2.6.1 Between-group differences Group differences of age, years of education, estimated Total Intracranial Volume (eTIV) and neuropsychological test scores (Moca and AVLT) were evaluated by analyses of variance (ANOVA) (p < 0.05). Then, the Mann-Whitney U test was used to assess the differences of sex distribution (p < 0.05). A general linear model was used to determine between-group differences of global network properties as well as the differences of connections classi ed by function organization, anatomical distance and hub regions, where the group factor was used as xed effect and the factors of age, sex and years of education were covariates (p < 0.05). Moreover, the between-group differences of nodal network properties were using bonferroni corrections for multiple comparisons, which means p < 0.05/90.

Age-related differences
To determine the interactive effect of age and group, a stepwise regression model (p < 0.05) and partial correlation analysis (p < 0.05) were adopted in this study (Zhao et al. 2017;Brown et al. 2011). A stepwise regression procedure began with age, group, age × group interaction, sex, years of education, and returned the subset of terms producing the most accurate model was selected. Thereafter, interpretability of the selected model was con rmed by performing a partial correlation analysis between network measures with signi cant interaction effect and age, separately for NC and SCD, with removement of the effects of sex and years of education. No multiple comparisons adjustment was performed on partial correlation analysis here, because the nodes were selected by the prior stepwise regression model and we consider this is an exploratory study.

Associations between network metrics and cognitive scores
We used partial Pearson's correlations controlled for age, sex and years of education to evaluate how clinical performance related to the altered network measures in individuals with SCD. The signi cance threshold was set at p < 0.05. All the statistical analyses were performed on MATLAB and SPSS.

Reproducibility analysis
Herein, we have computed the global network measures with a wide range of sparsity (15%-45%), and the between-group differences were assessed in all sparsity thresholds. In addition, we employed different parcellation schemes to de ne network nodes. The AAL atlas contain 12 subcortical structures, and we removed the subcortical structures to construct a 78 × 78 matrix to quantify the cortical network connections. Then, we repeated the same network analyses.

Background characteristics of the participants
The differences were not signi cant in age, sex, years of education, eTIV, MoCA scores, ALVT-immediate recall and AVLT-recognition scores between the NC and SCD group. In particular, the SCD group performed worse than the NC group in delayed recall of the AVLT (p < 0.05). The results of this section were summarized in Table 1. 3.2 Between-group differences of network metrics

Global network properties
The small-world networks characterized with higher clustering coe cients and similar characteristic path length compared with random networks. In this study, all individual morphometric networks exhibited small-world properties, where the Lp were similar with the matched random networks (lambda ≈ 1) and the Cp were higher than the matched random networks (gamma > 1) in all sparsity thresholds. In addition, that the sigma higher than 1 in all networks demonstrates a small-world organization in each network.
Compared with NC group, the sigma was slightly lower in SCD group(p < 0.05), which the sparsity located in 17-33%. Besides, another small world parameter gamma (at the sparsity of 15-32% with exception of 16%) was lower in individuals with SCD than NC (p < 0.05). Compared with NC group, the changes of E glob (lower in SCD) and Lp (higher in SCD) were signi cant in SCD group at the sparsity from 22-45%. Signi cant decreases of Cp (except the sparsity of 17% and 37%) and E loc (except the sparsity of 21%) had been detected in SCD compared with NC group (Fig. 2). In addition, all the network properties at the sparsity of 30% in individuals with SCD were deterioration compared with NC. Figure 2 Between-group differences in global network metrics as a function of sparsity. (1) The middle column is the schematic representation and calculation formula of network metrics. In the order from top to bottom, they are clustering coe cient, shortest path length or characteristic path length, local e ciency, global e ciency and the schematic representation of small-world network and random network. (2) The left and right column are the between-group differences of the network metrics. The arrows point from the representation of the network parameters to the results of the between-group differences. One asterisk means p < 0.05, two asterisk means p < 0.01, three asterisk means p < 0.001.

Nodal network properties and Rich-club organization
Compared with NC group, the nodal local and global e ciency were signi cant decreased only in left paracentral lobule (p < 0.05, Bonferroni corrected). Based on the combined measure, similar hub distributions were observed in the two group, mainly located in left thalamus, Pre-frontal lobe, occipital lobe and parietal lobe, which 17 regions in NC and 16 regions in SCD (Fig. 3A and Table S2). In addition, 13 hub regions in the two groups were overlapped. For different categories of connections classi ed by hub regions (Fig. 3B), the feeder connections showed lower strength in SCD subjects compared with NC (F = 11.515, p = 0.001), which consistent with the previous WM network studies (Shu et al. 2018;Yan et al. 2018). All comparisons in this section were performed at the sparsity of 30%.

Functional organization and Anatomical distance
For intra-connections within the ve functional organizations, the lower connectivity strength of paralimbic system was observed in individuals with SCD (F = 5.216, p = 0.025) (Fig. 3E). The inter-module connections between paralimbic and association area (F = 4.375, p = 0.04) as well as the interconnections between paralimbic and subcortical (F = 4.291, p = 0.042) were signi cantly lower in SCD subjects compared with NC. In addition, the strength of long connectivity (anatomical distance larger than 80mm) was signi cantly decreased in individuals with SCD compared with NC (F = 4.22, p = 0.044) (Fig. 3B).
Analysis of the network-based statistic resulted with one subnetwork with 24 nodes and 24 connections (edge-p < 0.01 and component-p = 0.002) (Fig. 3C). The NBS connectivity strength was the total strength within the disconnected network. The receiver operating characteristic curve (ROC) analysis revealed the connectivity strength of the subnetwork identi ed by NBS, showing high area under curve (AUC) value of 0.959 for classifying the two groups (Fig. 3D). In addition, the network connectivity strength in SCD group was lower than in NC group (F = 4.538, p = 0.037) (Fig. 3B). The results in this section were based on the networks with sparsity of 30%.     . 5B). Besides, the right paracentral lobule (r=-0.302, p = 0.029) showed signi cant decrease with age in nodal local e ciency in NC group, but not within SCD group (Fig. 5B). × × For the connectivity at divisional level (association area, limbic, paralimbic, primary sensory, and subcortical), the stepwise regression model revealed that intra-connectivity of paralimbic system, interconnectivity between association area and paralimbic system, as well as inter-connectivity between paralimbic system and subcortical showed signi cant age group interaction effects across all participants(p < 0.05). For partial correlation analyses, the inter-connectivity between association area and paralimbic system (r=-0.321, p = 0.034) showed signi cant correlation with age within SCD group, while nonsigni cant correlation was found in NC group. Additionally, the intra-connectivity of paralimbic system in NC group showed signi cant correlation with age but not in SCD group. The connectivity strength of the subnetwork calculated by NBS exhibited signi cant age group interaction effects in the regression model (t=-11.514, p < 0.001, beta=-0.799), and the followed partial correlation analyses revealed signi cant correlation between age and the calculated connectivity strength within SCD group (r=-0.412, p = 0.009), while nonsigni cant correlation was found in NC group (Fig. 3F). All comparisons in this section were performed at the sparsity of 30%.

Reproducibility ndings
The effects of the sparsity threshold for global network metrics were validated, and the results were summarized above. All 12 subcortical regions were excluded from the AAL atlas and the cortical network were constructed by using the same method. All cortical networks we have constructed were characterized with small world properties, which the Lp were similar with the matched random networks (lambda ≈ 1) and the Cp were higher than the matched random networks (gamma > 1) in all sparsity thresholds (sigma > 1) (Fig. S1). The clustering coe cients in individuals with SCD showed higher signi cance compared with NC group, at sparsity range of 15%-44%, except 19%-20% (all p < 0.05). Then, the local e ciency exhibited decrease in SCD group compared with NC at all sparsity thresholds excepted 21% (all p < 0.05). Thereafter, the characterized path length was longer in SCD group compared with NC at the sparsity of 23%-45% (all p < 0.05). In addition, the global e ciency was higher in NC group compared × × with SCD at the sparsity of 23%-45% (all p < 0.05). For the small world properties, sigma (at the sparsity of 23%-39% and 41%-45%) and gamma (at the sparsity of 15%-16%, 21%-34% and 43%-45%) were higher in NC group compared with SCD group (all p < 0.05). The global network metrics in this section were consistent with the ndings before (Fig. S1). For rich-club organization, 15 hub regions were found in each group (Table S2) and the strength of feeder connections were lower in SCD group compared with NC. Also, the subnetwork based on NBS can divide the two groups accurately (edge-p < 0.01 and component-p = 0.005).  (Table S3). For regional global e ciency, similar distributions of region were found signi cantly correlated with age, which mainly located in bilateral inferior frontal opercular gyrus, right superior frontal medial orbital gyrus left anterior cingulate and paracingulate gyri and right parahippocampal gyrus (Table S3). In general, the ndings showed slight effects of subcortical structures for global metrics, but changed local network properties.

DISSCUSSION
In this study, we used individual morphometric networks and graph theory analysis to assess the altered topological properties in individuals with SCD and age-related differences compared with NC. Main ndings were as follows: (1) the global network metrics such as global/local e ciency, clustering coe cients and small world properties decreased in SCD compared with NC; (2) the altered nodal network metrics in SCD mainly located in prefrontal lobe, parietal lobe and subcortical system; (3) compared with NC, signi cant decreases of global/local e ciency with increasing age were found in SCD subjects; (4) the signi cant age-related differences of nodal network metrics between two groups mainly located in prefrontal lobe; (5) disrupted strength of paralimbic system and feeder connectivity were found in SCD group. Finally, the robustness of the results was validated by using different sparsity thresholds as well as the effects of subcortical structures. The functional network based on fMRI coordinated brain activity correlations of the uctuate magnetic properties of oxygenated blood between regions, which can re ect the synchronized activity between brain regions(Alexander-Bloch et al. 2013). Disrupted topological organization of functional network in AD-related patients meant that functional integration and segregation of the brain activity was deterioration. While, WM ber bundles across the entire brain traced in diffusion MRI was labelled 'WM anatomical network', which can re ect the WM ber connections between brain regions. Altered topological organization of WM structure network in AD-related patients meant that the WM ber connections were impaired. Then, the GM network coordinated the morphological features between GM regions of brain, which can re ect the synchronized anatomical changes of GM in brain(Alexander-Bloch et al. 2013). Altered topological organization of GM network in AD-related patients may be suggestive of GM loss in correlated regions or localized degeneration in one region. The similarities in results across imaging modalities meant that the synchronized anatomical change indeed results from brain connectivity of some kind, such as synchronized brain activity change and/or WM ber connections.

Aberrant topological organization in individuals with SCD
Our results partial consistent with previous studies across imaging modalities above, which meant that the individuals GM networks can accurately explore the structural alterations of brain at network level in individuals with SCD independently. For functional organization, the intra-and inter-connections including paralimbic system decreased in SCD group compared with NC in this study. The paralimbic system is one of the transmodal areas with highest synaptic levels of sensory-fugal processing (Mesulam 1998), which plays a causal role in activating attentional and memory systems within association areas to facilitate controlled processing of stimuli during cognitively demanding tasks (Supekar et al. 2009;Sridharan et al. 2008). In our results, the decreased integration of paralimbic system and association areas in individuals with SCD may induce the decline of the cognitive ability. In addition, the altered intra-connections in paralimbic system and inter-connections between paralimbic system and subcortical in SCD group may enhance that individuals with SCD are at higher risk to cognitive decline in the future compared with NC.
The GM degeneration in individuals with SCD has been detected in previous studies (Zhao et al. 2019;Rabin et al. 2017). However, the underlying neuropathological mechanism of GM degeneration in SCD still remains barely unknown. We have revealed the GM degeneration at a large system level, and studies combined with multimodal imaging techniques should be considerate by researchers in the future.

Age-related differences of network metrics
For global network metrics, the global e ciency and local e ciency in SCD showed signi cant age group interaction effects, which was consistent with previous results from MCI studies (Zhao et al. 2017). Then, the partial correlation analyses revealed that global e ciency and local e ciency were signi cant correlated with age in individuals with SCD, while nonsigni cant correlations were found in NC subjects.
Moreover, similar with global/local e ciency, the shortest path length and clustering coe cients showed signi cant age group interaction effects, and there were signi cant relationships between those properties and age in individuals with SCD. Age is the main risk factor of AD, and accelerated decrease of global network metrics with age in SCD indicates that SCD subjects with the future risk of cognitive decline.
For nodal e ciency metrics, some brain regions showed decrease with age in individuals with SCD, including bilateral inferior frontal opercular gyrus, right superior frontal medial orbital gyrus, left anterior cingulate and paracingulate gyri, right insula, right inferior occipital gyrus and left putamen. These regions mainly located in prefrontal lobe and subcortical system. While, the lateral prefrontal cortex plays a critical role in working memory-executive function network (Mesulam 1998), and the anterior medial prefrontal cortex belong to midline core of default mode network (DMN) ( longitudinal study revealed that MCI patients showed accelerated GM decrease compared with normal controls in whole brain volume, temporal gray matter, and orbitofrontal and temporal association cortices, including the hippocampus (Driscoll et al. 2009). Our study revealed that individuals with SCD showed accelerated GM degeneration with age at a macroscopic scale. However, the underlying mechanism of the age-related changes, AD-related pathology changes and their co-pathologies were still need to explore.

Correlation between network metrics and clinical scores
Clinical neuropsychological testing is a conventional method for memory examinations and diseaseassisted diagnoses. Then, two test scores were used for correlation analysis, including AVLT and MoCA.
We have observed that the delayed recall score of AVLT exhibited signi cant correlation with sigma and gamma in individuals with SCD. Moreover, signi cant correlation between sigma and recognition score of AVLT was found in individuals with SCD but not in NC subjects. Sigma and gamma were classical small world indices, and if the indices were higher, the property of small world was stronger. In the sense, stronger the small world organization, the higher and more e cient information segregation and  Zhao et al. 2017) revealed that the connectivity of rich-club, feeder and local were correlated with the delayed recall score of AVLT, but not showed in our study. In this study, the strength of feeder connectivity showed signi cant correlation with MoCA scores. In our opinion, the difference between our study and previous studies is mainly due to the different neurobiological mechanisms behind GM networks and WM networks. The correlative variation in GM network means the changing of regional morphology, and the correlations in WM network represents the strength of WM ber connectivity between two regions. The GM network and WM network characterize the network in human brain in two ways. In addition, majority of correlations between the network measures and cognitive test scores were negative, a more randomly organized GM network would be a reason (Verfaillie et al. 2018). In the end, the coordination of GM network, WM network and functional network to investigate the alterations in individuals with SCD should be considered in future studies.

Limitations
Methodological issues in our research should be addressed. First, the sample size is small. Although we have constructed the individual network in this study, a large sample size will be better. Second, crosssectional samples were used in this study, so a longitudinal MRI data will be collected for future study.
Third, only distribution of GM density was used to construct the networks, and thus more morphological indices will be used to de ne the network connections. Fourth, very limited neuropsychological battery adopted was always a limitation. As the modi ed research framework for SCD was published (Jessen et al. 2020), more comprehensive neuropsychological tests should be addressed.

Conclusion
In summary, this study revealed that aberrant topological organization was showed in individuals with SCD, including the decreased local/global e ciency, clustering coe cients, sigma, gamma and the increased shortest path length. Compared with NC group, the rich-club connections in individuals with SCD were persevered but the feeder connections decreased. Moreover, the connectivity of paralimbic system was disrupted in individuals with SCD compared with NC subjects. In addition, the age-related decreases in nodal global e ciency in individuals with SCD mainly distributed in prefrontal lobe. The ndings in this study may enhance the understanding of the underlying pathological mechanisms in individuals with SCD. The network metrics were calculated by GRETNA toolbox and the between-group differences were computed.

Declarations
(6) The age-related differences of network metrics in the two groups and the relationships between cognitive test and network metrics were computed.