Structural Network Topology Reveals Higher Brain Resilience in Individuals with Preclinical Alzheimer's Disease

Subjects

In this study, the selection criteria for the database were CN participants with availability of T1-weighted MRI, florbetapir (AV45) PET, flortaucipir (AV1451) PET, and diffusion tensor imaging at baseline. The diffusion imaging was further limited to acquisition using Siemens scanners with 48 diffusion directions to reduce the impact of scanner types on tractography and graph construction (Kurokawa et al., 2021). Demographic data such as age, sex, years of education, clinical dementia rating (CDR), mini-mental state examination (MMSE), and clinical diagnosis were recorded.

PET biomarkers and group classification

According to the National Institute on Aging and Alzheimer's Association 2018 Research Framework (Jack et al., 2018), the presence of Aβ determines an individual's placement on the Alzheimer's continuum. Specifically, the presence of Aβ without tau (A+T−) indicates Alzheimer's pathologic change, while the presence of both Aβ and tau (A+T+) is indicative of AD.

Following the recommendation, we categorized CN subjects into the following groups: (1) CN subjects with normal Aβ and tau (A−T−); (2) CN subjects with abnormal Aβ, but normal tau (A+T−); and (3) CN subjects with abnormal Aβ and tau (A+T+). The status of Aβ and tau was determined using florbetapir and flortaucipir PET values provided by the University of California, Berkeley. The florbetapir PET was used to assess global Aβ load.

The mean standardized uptake value ratio (SUVR) on florbetapir PET images was determined in anterior/posterior cingulate, lateral parietal, and lateral temporal regions using the whole cerebellum as the reference region. A cutoff value of 1.11 was applied to determine Aβ positivity (A+) (Landau et al., 2013). The flortaucipir PET SUVR values were used to quantify tau load.

A size-weighted mean flortaucipir SUVR was calculated in metatemporal regions, including the amygdala, entorhinal cortex, fusiform, and inferior temporal and middle temporal gyri, and normalized by inferior cerebellar gray matter. Tau positivity (T+) was thresholded at SUVR >1.23 (Jack et al., 2017).

MRI acquisition

All data were collected using a 3T Siemens scanner (Siemens Medical Solutions, Erlangen, Germany). Image protocols were standardized across ADNI study sites. Details of MRI parameters can be found at http://adni.loni.usc.edu/methods/documents/mri-protocols/. Three-dimensional, T1-weighted volumetric sequences were acquired with the following parameters: echo time = 2.98 ms, repetition time = 2300 ms, inversion time = 900 ms, flip angle = 10°, field of view = 208 × 240 × 256 mm, and acquired resolution = 1 × 1 × 1 mm.

Axial diffusion MRI data were acquired with 48 diffusion-encoding directions at b = 1000 s/mm² and 6 nondiffusion-weighted b₀ images. Other scanning parameters include resolution = 2 × 2 × 2 mm; time repetition = 7200–9600 ms; time echo = 56–82 ms; flip angle = 90°; and field of view = 232 × 232 × 160 mm.

Structural MRI preprocessing

T1-weighted images were corrected for head motion and intensity inhomogeneity, followed by nonbrain tissue removal using the pipeline implemented in FreeSurfer (version 7.1.1, http://surfer.nmr.mgh.harvard.edu/). The brain was parcellated into 68 cortical regions based on the Desikan–Killiany atlas (Desikan et al., 2006) as well as the following 7 subcortical regions in each hemisphere (Fischl et al., 2002): thalamus, caudate, putamen, pallidum, hippocampus, amygdala, and accumbens.

Considering that white matter hyperintensities (WMHs) shown on T2 FLAIR images are possible factors associated with vascular disease and may confound the fiber tracking algorithm (Reginold et al., 2018), we controlled for WMH volumes in the statistical analysis. WMH volumes were obtained from ADNI, which were estimated with a modified Bayesian probability structure through histogram fitting. Briefly, the likelihood estimates of each image were calculated through histogram segmentation and thresholding. Then, along with WMH priors (which were estimated from more than 700 individuals) and tissue constraints, the probabilities were thresholded at 3.5 standard deviations above the mean to create a binary WMH mask. Finally, the mask was back-transformed to native space for volume calculation. WMH volumes were available for all included subjects.

Diffusion MRI preprocessing

Diffusion-weighted imaging (DWI) volumes were processed using the FMRIB software library (FSL) toolbox and MRtrix3 (Tournier et al., 2019) to correct for noise, Gibbs ringing artifacts, head motion, and eddy current distortions. b vectors were rotated accordingly (Leemans and Jones, 2009). All nonbrain tissues were removed from the diffusion-weighted images using the FSL Brain Extraction Tool.

A field map was utilized to correct for echo planar imaging (EPI)-induced susceptibility artifacts, which usually caused geometric distortions at tissue–fluid interfaces. The field map was calculated from a double-echo gradient echo sequence by taking the difference between two phase images and dividing that by the echo time difference. Skull-stripped b₀ images were registered to their respective T1 image based on white matter boundaries with the FSLs epi_reg tool to coregister DWI and T1 images. The resulting 3D transformation matrices and deformation fields were back-transformed and applied to the T1 image, so all images were coregistered in the diffusion space.

All processed images were visually inspected for quality control. Fifteen images were excluded from further analysis due to excessive distortion artifacts, and one image was excluded due to lack of a field map for EPI artifact correction.

White matter tractography

We performed probabilistic tractography with MRtrix3. The signal expected for a single-fiber white matter population (so-called response functions) was calculated for constrained spherical deconvolution to estimate fiber orientation distributions (FODs). The FOD maps were passed to the iFOD2 algorithm for fiber tracking. Seeding points were determined dynamically using the spherical-deconvolution informed filtering of the tractogram model (Smith et al., 2015). Anatomical constraints were applied to ensure minimal contamination from spurious streamline trajectories through gray matter (Smith et al., 2012). A total of 1 million streamlines were generated for better reproducibility (Roine et al., 2019). Other default tracking parameters included a step size of 1 mm, minimum fiber length of 4 mm, and FOD cutoff of 0.1. Finally, the SIFT2 algorithm was applied on all tractograms to filter implausible streamlines by assigning weights to each streamline. This step can improve the biological plausibility of the final connectome reconstruction (Smith et al., 2015).

Network construction and graph theoretical analysis

Graph theory was used to analyze topological changes in the brain network. In such analysis, the brain network is represented by a graph with nodes and edges. Nodes refer to brain regions, while edges refer to the connection between them. In this study, we defined nodes as the 82 regions segmented by FreeSurfer (68 cortical +14 subcortical regions) and edges as white matter links between these regions. Therefore, an 82 × 82 connectivity matrix was built for each subject. An overview of the methodology for building structural networks is shown in Figure 1.

To compare the network topology across different subjects and minimize the number of spurious connections in each network, we thresholded and binarized the connectivity matrices to ensure the same number of edges in all subjects. Fixed-density thresholding can yield a more reproducible and consistent network architecture while maintaining sensitivity to disease effects (De Brito Robalo et al., 2022). The network density is defined by the fraction of the number of edges over the maximal possible number of edges (i.e., 82 × 81/2 = 3321 possible edges in our case). All self-connections were excluded.

Considering the lack of a definitive method to select a single threshold, we chose to binarize the graph over a range of density threshold levels (5–9%, in the step of 0.2%) and estimated the global network properties at each threshold value. For global network characteristics, we analyzed global efficiency and assortativity. Global efficiency is a measure of network integration and it is calculated by the average inverse of the shortest path in the network. Assortativity is a measure of brain resilience and refers to the tendency of nodes to link with other similar nodes (Franciotti et al., 2019; Miraglia et al., 2022). A positive value indicates that nodes tend to link to other nodes with the same or similar degree.

For the local network analysis, we constructed structural networks at a specific threshold of 7.8% (Fig. 3A). This threshold ensures that all regions are included in the network while minimizing the number of false-positive connections. For nodal characteristics, we focused on the hippocampus and assessed various nodal topological metrics such as degree, local efficiency, and clustering coefficient.

Degree is calculated as the number of links that are connected to a particular node. Local efficiency reflects the extent of integration between immediate neighbors of a given node (Fornito et al., 2016) and it is calculated as the average shortest path length within the node's neighborhood. The clustering coefficient is calculated as the fraction of a node's neighbors that are neighbors with each other. All network topological metrics were calculated by the Brain Connectivity Toolbox in MATLAB (Rubinov and Sporns, 2010).

Statistical analyses

The normality of continuous variables was first examined using the Shapiro–Wilk test, in which the assumption of normality was violated. WMH was log-transformed for subsequent analysis due to its skewness. For continuous variables with normal distribution, an analysis of variance, followed by Scheffe's tests, was conducted for group comparison. For continuous variables with skewed distribution, the Kruskal–Wallis test along with the Mann–Whitney U test for post hoc comparisons was performed. The chi-square test was employed to analyze categorical variables.

An analysis of covariance was performed with control of age, sex, years of education, and WMH volume to evaluate group differences in network metrics. A false discovery rate was used to adjust false-positive results in multiple comparisons. If the results were significant, then post hoc comparisons were performed with the Bonferroni method.

Pearson's partial correlation coefficient was utilized by controlling for age, sex, years of education, and WMH volume to assess the associations of network topological metrics with global Aβ load and tau burden in the temporal metaregion of interest (ROI). Network metrics were calculated under the density threshold of 7.8% to reduce the number of comparisons. Correlation was examined in the whole sample as well as within each group.

All statistical analyses were performed using SPSS 26.0 (Armonk, NY). Statistical tests were two-sided, with the significance level set to p < 0.05.

Demographics

After selection and image quality control, a total of 68 CN participants were included and classified as A−T− in 19 cases, A+T− in 28 cases, and A+T+ in 21 cases according to amyloid and tau PET imaging. The demographic details are presented in Table 1. The groups were comparable in terms of sex (p = 0.13), age (p = 0.35), years of education (p = 0.42), CDR (p = 0.35), MMSE (p = 0.34), WMH volumes (p = 0.65), and hippocampal volume (p = 0.88). However, they differed significantly in terms of global amyloid-PET SUVR and tau-PET SUVR (all with p < 0.001).

Global network properties

Results from the comparison of global network metrics among groups are presented in Figure 2 and Supplementary Table S1. Overall, no significant differences were observed in global efficiency across groups at different densities, while after controlling for age, sex, years of education, and WMH volume, a significant difference in assortativity was found.

Specifically, the A+T+ group exhibited increased assortativity compared with the A−T− group and A+T− group at the density range of 5.4–8.4% and 6.0–9.0%, respectively. Additionally, the A+T− group showed higher assortativity than the A−T− group at the density range of 5.6–6.0%.

Local network properties in the hippocampus

In Figure 3, comparisons among different groups of topological properties in the hippocampus are presented, such as degree, local efficiency, and clustering coefficient calculated under the density threshold of 7.8%, in which all 82 nodes were included in the largest component of the 3 groups.

In the left hippocampus, significant differences in network properties were observed across groups for local efficiency [F (2, 61) = 4.845, p = 0.011] and clustering coefficient [F (2, 61) = 4.513, p = 0.015], but results were comparable in terms of degree [F (2, 61) = 1.815, p = 0.171]. The A+T+ group showed lower local efficiency (p = 0.009) and a lower clustering coefficient (p = 0.014) than the A−T− group. Conversely, no significant differences were found across groups in the right hippocampus (all p values >0.05).

Association of network measures with Aβ and tau burden

As significant differences were observed in the assortativity, local efficiency, and clustering coefficient of the left hippocampus across the three groups, we conducted further analysis to explore their relationship with global Aβ load and tau burden in the temporal meta-ROI under the density threshold of 7.8%.

The results showed that in the entire sample, an association was observed between assortativity and tau load (Fig. 4), with higher assortativity correlated with a greater tau-PET SUVR (r = 0.29, p = 0.02). However, this correlation was not observed in the subgroups (Table 2). No correlation was found between assortativity and amyloid load (r = 0.13, p = 0.31). Nodal metrics, including local efficiency and clustering coefficient, did not show any association with either amyloid or tau load (Supplementary Table S2).

Global network properties

At the global scale, efficiency did not differ significantly across groups of CN individuals with varying levels of AD pathology, while assortativity was higher in the A+T+ group than in the A−T− and A+T− groups. Assortativity describes how nodes in a complex network tend to link with other similar nodes. A positive assortativity coefficient implies that nodes in a network are more likely to connect to nodes with similar degrees, indicating enhanced information processing efficiency and greater resistance to random attacks or deliberate removal of central nodes (Sun et al., 2020). Therefore, higher assortativity in the A+T+ group suggests that their brain networks are more robust and resilient.

Brain resilience refers to an individual's ability to cope with pathology while remaining CN (Arenaza-Urquijo and Vemuri, 2018). Our findings support this notion by showing increased assortativity in groups with greater pathology, while maintaining comparable global efficiency. Thus, the network measure could complement identification of pre-clinical AD.

Assortativity measured from structural connectivity was rarely reported. In a single study by Coninck et al. (2020), increased assortativity was found in subjects with AD. Our results further demonstrate that the increased assortativity can be detected at the pre-clinical stage. In other AD studies, assortativity has usually been measured from functional connectivity using functional MRI (Bahrami and Hossein-Zadeh, 2015; Luo et al., 2021) and electroencephalography (Fathian et al., 2022; Franciotti et al., 2019; Kim et al., 2022). However, comparisons with these studies are not valid because structural connectivity tends to form an assortative network, while functional connectivity tends to form a disassortative network (Lim et al., 2019).

In the present study, there was no significant difference in global efficiency between CN participants with different levels of amyloid and tau pathology. However, previous studies have reported lower global efficiency in individuals with mild cognitive impairment (Bergamino et al., 2022; Jacquemont et al., 2017), subjective cognitive decline (Shu et al., 2018), and pre-clinical AD (Fischer et al., 2015; Pereira et al., 2018) and CN individuals with an autosomal dominant AD mutation (Prescott et al., 2022).

Some studies suggested that amyloid alone may not be sufficient to disrupt the structural network topology during the pre-clinical stage (Pereira et al., 2018). The lack of decreased global efficiency in the A+T+ group in this study suggests that deposition of both amyloid and tau may not be sufficient to alter global efficiency.

Local network properties in the hippocampus

At the nodal scale, the A+T+ group showed a lower clustering coefficient and local efficiency than the A−T− group. The clustering coefficient measures the proportion of neighbors of a node that are connected to each other, and a decrease in this coefficient could indicate reduced connectivity within the local neighborhood or increased connectivity between the node and remote nodes (Jacquemont et al., 2017). In the left hippocampus, the degree, or the number of connected links, remained the same, so the lower clustering coefficient suggests reduced connectivity within the neighborhood.

The decrease resulted specifically for the left hippocampus, which is more susceptible to AD pathogenesis (Shi et al., 2009). The reduction in local efficiency indicates a loss of local integration around the left hippocampus, and this has also been observed in the amnestic MCI converter (Sun et al., 2019). Therefore, changes in the topology of the hippocampus could potentially serve as biomarkers for monitoring the early accumulation of AD pathology substrates.

Limitation

This study has several limitations. First, the size of the cohort is relatively small, leading to less statistical power to distinguish between groups. While multicenter diffusion images were available in ADNI, the reliability of the graph theory metrics could be reduced by including different scanners (Kurokawa et al., 2021). We limited the data to Siemens scanners with 48 diffusion directions to minimize such impact, resulting in the small sample size.

Second, we did not further classify subjects based on the neurodegeneration profile due to the small sample size. Neurodegeneration factors such as decrease of fluorodeoxyglucose and temporal atrophy may correlate with the disrupted structural network more than amyloid in pre-clinical AD (Pereira et al., 2018) and MCI (Jacquemont et al., 2017). We controlled the neurodegeneration effect in the complementary analysis and arrived at the same conclusion. Further studies may explore the complete spectrum of pre-clinical AD, assessing additional neurodegeneration factors.

Third, probabilistic tractography was utilized to construct connectomes, and this method suffers from severe false-positive connections (Sarwar et al., 2019). We applied the stringent thresholding method to maximize the likelihood of identifying true white matter connections (Buchanan et al., 2020).