Conductance-Based Structural Brain Connectivity in Aging and Dementia

Abstract

Background: Structural brain connectivity has been shown to be sensitive to the changes that the brain undergoes during Alzheimer's disease (AD) progression.

Methods: In this work, we used our recently proposed structural connectivity quantification measure derived from diffusion magnetic resonance imaging, which accounts for both direct and indirect pathways, to quantify brain connectivity in dementia. We analyzed data from the second phase of Alzheimer's Disease Neuroimaging Initiative and third release in the Open Access Series of Imaging Studies data sets to derive relevant information for the study of the changes that the brain undergoes in AD. We also compared these data sets to the Human Connectome Project data set, as a reference, and eventually validated externally on two cohorts of the European DTI Study in Dementia database.

Results: Our analysis shows expected trends of mean conductance with respect to age and cognitive scores, significant age prediction values in aging data, and regional effects centered among subcortical regions, and cingulate and temporal cortices.

Discussion: Results indicate that the conductance measure has prediction potential, especially for age, that age and cognitive scores largely overlap, and that this measure could be used to study effects such as anticorrelation in structural connections.

Impact statement

This work presents a methodology and a set of analyses that open new possibilities in the study of healthy and pathological aging. The methodology used here is sensitive to direct and indirect pathways in deriving brain connectivity measures from diffusion-weighted magnetic resonance imaging, and therefore provides information that many state-of-the-art methods do not account for. As a result, this technique may provide the research community with ways to detect subtle effects of healthy aging and Alzheimer's disease.

Introduction

Brain structural connectivity reflects the physical connections through white-matter (WM) axon bundles between different regions of interest (ROIs), and can be measured by diffusion-weighted magnetic resonance imaging (dMRI). Brain connectivity analysis has been proven to be useful in the study of many conditions, such as the effects of aging on the brain (Damoiseaux, 2017; Fjell et al., 2016; Wu et al., 2013), and in the study of disease.

Differences in brain connectivity patterns between healthy and diseased populations are potential indicators of changes in the brain “wiring” due to disease processes. In particular, Alzheimer's disease (AD) has been found to impact structural connectivity (Daianu et al., 2013; Prasad et al., 2015; Rose et al., 2000). The changes that the brain undergoes with aging and AD can be confounded, thereby contributing to a delay in AD diagnosis. Nonetheless, the spatial and temporal patterns of changes in connectivity are different in healthy aging and in AD. Accurate modeling of structural connectivity may therefore reveal the effects of aging and AD progression in WM degeneration, and help to differentiate the two.

Brain changes in aging and dementia

Healthy aging is associated with a moderate decline of some cognitive abilities. AD dementia causes severe deterioration of similar cognitive domains, and also additional cognitive functions, in such a way that it compromises independent living. The abnormal decline preceding AD, with noticeable alterations in cognition yet short of functional independence, is termed mild cognitive impairment (MCI) (Petersen et al., 1999). Both the gray matter (GM) and the WM undergo changes in volume and integrity in healthy and pathological aging, but the affected regions vary.

Healthy aging has been found to be related to decline in the frontal and temporal regions of the GM. An age-related volume decline has been localized in the prefrontal cortex (PFC), insula, anterior cingulate gyrus, superior temporal gyrus, inferior parietal lobule, and precuneus (Ghosh et al., 2011), as well as in the striatum, caudate, and medial temporal lobe (hippocampus and adjacent, anatomically related cortex, including entorhinal, perirhinal, and parahippocampal cortices) (Raz and Rodrigue, 2006). Other regions, such as the occipital cortex, are mostly unaffected by aging.

In healthy aging, degeneration in WM has been found to often follow an anterior/posterior gradient of fractional anisotropy (FA) reductions, indicating that frontal connections are especially vulnerable (Toepper, 2017). Small and less myelinated fibers are particularly vulnerable to age-related decline, such as fiber tracts whose myelinization is completed later in life (Salat et al., 2005).

In patients with AD dementia, changes in regional volume are not uniform. A significant volume reduction has been found in the hippocampal formation and the entorhinal cortex bilaterally very early in the disease (Thompson et al., 2004), followed by a degeneration in the PFC (Ghosh et al., 2011). In the WM, signal abnormalities (WMSA) have been found in AD in regions such as rostral frontal, inferior temporal, and inferior parietal WM, with a greater volume of WMSA in AD with respect to healthy aging consistently across different ages. In MCI, frontal and temporal regions have been found to have greater WMSA volume with decreasing time-to-AD-conversion (Lindemer et al., 2017).

Brain connectivity changes in aging and dementia

While enabling the segregation and integration of information processing, brain networks are also responsible for the widespread effects resulting from local disease-related disruptions, thereby complicating relationships between pathological processes and clinical phenotypes in AD (Tijms et al., 2013).

The disconnection model of AD has long been discussed (de LaCoste and White, 1993), with cumulative evidence associating plaques and tangles with local synaptic disruptions (Arendt, 2009; Takahashi et al., 2010), as well as linking the cognitive dysfunction in AD to dysconnectivity between highly interrelated brain regions (Brier et al., 2014; Delbeuck et al., 2003; Matthews et al., 2013). Identifying large-scale brain networks that are vulnerable or resilient in aging and AD (by studying the human connectome) can therefore reveal underlying disease propagation patterns in the brain and provide connectivity-based biomarkers (Gomez-Ramirez and Wu, 2014) in prodromal AD.

Network-based analysis of brain WM connections through dMRI (Basser and Özarslan, 2014; Goveas et al., 2015; Madden et al., 2012) has been proven promising in revealing the structural basis of cognitive changes in AD, MCI, and aging, and the discovery of diagnostically and therapeutically important biomarkers.

Structural networks have been used to predict the process of disease spread in AD (Raj et al., 2012, 2015), and to distinguish the groups of cognitively normal (CN), MCI, and AD (Aganj et al., 2014; Frau-Pascual et al., 2019b; Prasad et al., 2015; Shao et al., 2012), as well as AD from vascular dementia (Zarei et al., 2009). Relative to CN controls, AD patients have been shown to have significantly lower integrity of association fiber tracts (Rose et al., 2000), weaker cingulum connectivity (Huang et al., 2012; Mielke et al., 2009; Zhang et al., 2007), and structural brain networks with disruption in their rich club organization (Daianu et al., 2016; Lee et al., 2018), and reduced local efficiency (Lo et al., 2010; Reijmer et al., 2013).

The performance in memory and executive functioning of AD patients has been inversely correlated to the path length (Reijmer et al., 2013), and network small-worldness has been shown to predict brain atrophy in MCI (Nir et al., 2015). Structural brain networks are affected even in individuals without dementia with the APOE ϵ4 allele (Liu et al., 2013), a genetic AD risk factor.

Connectivity disruption within a brain network is also occasionally accompanied by hyperconnectivity in a reciprocal network. For instance, increased frontal connectivity may be observed alongside reduced temporal connectivity in AD (Supekar et al., 2008; Wang et al., 2007), and an inverse relationship has been reported between frontal activity and occipital activity in aging (Davis et al., 2008).

Furthermore, AD has been shown to reduce connectivity in the default mode network (DMN) but intensify it at the early stages in the salience network—a collection of regions active in response to emotionally significant stimuli (Seeley et al., 2007; Uddin, 2016)—whereas behavioral variant frontotemporal dementia has been shown to attenuate the salience network connectivity but enhance DMN connectivity (Brier et al., 2012; Zhou et al., 2010). Even so, most existing studies monitor such hyperconnectivity with respect to the progression of dementia, but not with respect to the deterioration of other networks.

Contributions of this work

We have previously introduced a method for inferring structural brain connectivity from dMRI using an electrical conductance model (Frau-Pascual et al., 2019b), which accounts for all possible WM pathways, and is solved globally. This method was shown to produce structural connectivity measures that were more strongly correlated with resting-state functional connectivity and more sensitive to AD-related WM degeneration than standard streamline tractography methods did.

In this work, we extend our analysis to demonstrate the impact that this new measure of structural brain connectivity could have in the study of aging and AD dementia. To that end, we investigate the relationship of structural connectivity with age and cognitive and volumetric measures, attempt to predict age and cognitive scores from dMRI data, and identify some anticorrelated connections that might help to study compensation. This article extends our preliminary conference publications (Aganj et al., 2020; Frau-Pascual et al., 2019a); in particular, we have added more data analysis and experiments.

Methods

Conductance model

In our previous work (Frau-Pascual et al., 2019b)¹, we extended the heat equation method proposed by O'Donnell et al. (2002) with a combination of differential circuit laws. Our method assigns to each image voxel a local anisotropic conductivity value D, which is the diffusion tensor computed from dMRI (Basser et al., 1994). By solving the partial differential equation (PDE), $-\nabla \cdot \left(D\nabla {\varphi }_{i,j}\right)={\gamma }_{i,j},$ for a given current configuration ${\gamma }_{i,j}$ between a pair of source (i) and sink (j) voxels, we find the potential map ${\varphi }_{i,j}$ for that specific configuration. $\nabla$ and $\nabla \cdot$ are the gradient and divergence operators, respectively.

Our algorithm solves the PDE for a 1-ampere current (without loss of generality) between a pair of voxels i and j: ${\gamma }_{i,j}={\delta }_i-{\delta }_j$, where ${\delta }_k\left(x\right):=\delta \left(x-x_k\right)$, with x_k the position of voxel k and $\delta \left(\cdot \right)$ the Dirac delta. To compute ROI-wise conductance, we distribute the currents among the sets of voxels I and J (the two ROIs) as follows: ${\gamma }_{I,J}=\frac{1}{\left|I\right|}{\sum }_{i\in I}{\delta }_i-\frac{1}{\left|J\right|}{\sum }_{j\in J}{\delta }_j$.

The conductance between two points is then computed from Ohm's law as the ratio of the current to the potential difference. In our case, the potential difference between two voxels (or ROIs) i and j is ${\varphi }_{i,j}\left(x_i\right)-{\varphi }_{i,j}\left(x_j\right)$. The conductance is therefore computed, for ROI-wise connectivity, as follows:

$$C_{I,J}=\frac{1}{\frac{1}{\left|I\right|}{\sum }_{i\in I}{\varphi }_{I,J}\left(x_i\right)-\frac{1}{\left|J\right|}{\sum }_{j\in J}{\varphi }_{I,J}\left(x_j\right)}.$$

The conductance among all N regions can be computed efficiently in O(N) using the superposition principle (Frau-Pascual et al., 2019b). High conductance (i.e., low resistance) between two points represents a high degree of connectivity in our model. Since the ROIs are all at least weakly connected, these maps can then be thresholded.

Brain connectivity matrix generation

With the conductance method, we model and quantify diffusion data in a nonconventional way. As mentioned in the Conductance Model section, a 1-ampere current is split across voxels and the PDE is solved once per ROI to compute a conductance measure between each pair of ROIs using the superposition principle. This eventually results in a connectivity matrix per subject that reflects the ease with which this small current goes from one region to another, following the diffusion tensors. This measure also embeds geometrical information, such as volumes (number of voxels in each ROI) and distances between ROIs (as implied by the spatial derivatives in the PDE).

Study of conductance matrices

In this study, we considered the relationship between the mean conductance and other variables, such as age and cognitive scores (and cortical/subcortical volumes in the Supplementary Materials). The cognitive scores considered here were the Clinical Dementia Rating (CDR) scale (Morris, 1991) and the Mini–Mental State Examination (MMSE) score (Pangman et al., 2000), explained in more detail in the Analysis of AD Population section. We fit a linear function to our data points and report the correlation (r) and significance (p) values of the fit, revealing whether conductance was significantly correlated with these variables.

We also attempted to predict variables such as age and cognitive scores via linear regression. We initially discarded outliers in every cohort, while considering an outlier a subject with mean conductance higher or lower than the average by two standard deviations. We report r and p values of the correlation between the predicted and the true variable when fitting with a cohort and predicting with a different cohort. The p-values below 0.05 were considered significant when r was positive (as a negative r would have indicated the opposite of the desired effect). We also report the Bonferroni-corrected p-values (p_b), that is, the original p-values multiplied by the number of comparisons. Furthermore, we tried fitting and predicting within the same cohort (which involved fewer comparisons), using cross-validation with 20 folds.

We then measured the correlation of the conductance values between every pair of regions with age and cognitive scores (CDR and MMSE). For each pair of regions, we corrected the p-value for multiple comparisons using Bonferroni correction. The results would reveal which regions correlate more strongly with age and/or cognition.

We further used a general linear model to regress out the effects of age and sex before correlating the conductance with CDR or MMSE. This would disentangle the overlapping contributions of age/sex and cognition, and reveal the residual effects of CDR and MMSE unexplained by age/sex.

Identification of anticorrelated connections

Next, we attempted to identify negative (cross-subject) interrelationships among brain connections. As opposed to focusing only on the relationship between connectivity and the clinical data, we identified pairs of connections that are significantly negatively correlated with each other, and validated them on external data sets. Such a connection-wise correlation approach might help to reveal pathways that are potentially compensatory and define the resilience mechanism of brain networks against AD.

We first vectorized the lower triangular part of each $N\times N$ connectivity matrix to a vector of length $N\left(N-1\right)∕2$, and reduced this vector to keep M corticocortical and cortico/subcortical connections. Next, we computed the cross-subject linear correlation coefficient between each pair of connections, resulting in two symmetric $M\times M$ connection-wise matrices of correlations, R, and p-values, P.

We then kept only the elements of R with a correlation value smaller than a negative threshold, for example, $\tau =-0.1$, as ${ℛ}^{-}=\left\{\left(i,j\right)|R_{i,j}<\tau \right\}$. From that set, we considered the connection pairs whose p-values survived a cutoff threshold, namely $\alpha =0.05$, as $ℒ=\left\{\left(i,j\right)\in {ℛ}^{-}|P_{i,j}^{\ast }<\alpha \right\}$. $P^{\ast }$ was the set of p-values corrected for multiple comparisons among the elements of ${ℛ}^{-}$ with the Holm/Bonferroni method. We regarded the surviving set $ℒ$ as the pairs of connections with significant cross-subject anticorrelation. We kept either the entire $ℒ$ or a most significant subset of it.

Next, to externally test if the surviving set $ℒ$ was anticorrelated, we computed $R_{test}$ and $P_{test}$ for the connection pairs in $ℒ$ in a different population, and verified both $R_{test}<0$ and $P_{test}^{\ast }<\alpha$ for that set, with $P_{test}^{\ast }$ being $P_{test}$ corrected for multiple comparisons among the pairs in $ℒ$. We also tested the hypothesis that the surviving pairs of connections were left/right symmetric; that is, whether a significant anticorrelation was also a significant anticorrelation in the mirrored hemisphere.

Lastly, we correlated the identified connections with cognitive performance measures.

MR data processing

The common pipeline for brain connectivity computation is segmentation of brain ROIs, quantification of brain connections from dMRI, and aggregation of connectivity values in a matrix. The constructed brain connectivity matrix describes how strongly different regions are connected to each other according to the dMRI acquisition of WM connections. We processed the MRI data similarly for all data sets.

Structural MRI processing

We performed tissue segmentation and parcellation of the cortex into ROIs using FreeSurfer² (Fischl, 2012). The parcellation used in this work was the Desikan/Killiany atlas (Desikan et al., 2006), which has 86 regions, among which 68 were cortical and 18 were subcortical or brainstem. The atlas used here has a moderate number of ROIs, which helps to preserve statistical power after Bonferroni correction. For stability and replicability, we decided to use the segmentation results already provided by the database staff (except for European DTI Study in Dementia [EDSD]), rather than to reprocess all the structural images with the latest version of FreeSurfer.

Diffusion MRI processing

Diffusion preprocessing was performed using the FSL software³ (Jenkinson et al., 2012) and included BET for brain extraction and EDDY for eddy current and subject motion correction.⁴ From the preprocessed dMR images, we reconstructed the diffusion tensors using the diffusion tensor imaging (Basser et al., 1994) reconstruction module of DSI Studio⁵, which we then used as input to our conductance computation algorithm.

Results

Analysis of AD population

We retrospectively analyzed four anonymized and publicly available databases to evaluate how our conductance method could help in studying AD stage.

For this, we used two publicly available data sets that included subjects across the AD dementia spectrum (Fig. 1): the second phase of Alzheimer's Disease Neuroimaging Initiative (ADNI-2)⁶ (Beckett et al., 2015; Jack et al., 2008), and the third release in the Open Access Series of Imaging Studies (OASIS-3) (Fotenos et al., 2005), which is a longitudinal neuroimaging, clinical, and cognitive data set for normal aging and AD. To avoid data heterogeneity, we divided the OASIS-3 data set into four cohorts, each of which included more than 100 subjects with similar description in the “Scans” field of the data sheet. These two data sets, which we used for training and internal validation, enabled us to compare structural brain connectivity in different stages of the disease and correlate neuroimaging data to clinical cognitive scores.

FIG. 1.

Demographics of the AD data sets used here. (a) ADNI-2 cohort of 213 subjects (77 CN, 89 MCI, 47 AD), (b) OASIS-3 group of 652 subjects, consisting of 4 cohorts, each of which had more than 100 subjects with similar description in the “Scans” field of the data sheet: (c) 272-subject cohort (187 CN, 38 AD, 47 other dementias), (d) 139-subject cohort (86 CN, 33 AD, 20 other dementias), (e) 125-subject cohort (112 CN, 4 AD, 9 other dementias), and (f) 116-subject cohort (103 CN, 6 AD, 7 other dementias). Other dementias included vascular dementia, or AD dementia with depression or additional symptoms (refer to Fig. 2). (g) EDSD Freiburg cohort (16 CN, 11 MCI, 10 AD). (h) EDSD Rostock-3T cohort (20 CN, 22 MCI, 15 AD). AD, Alzheimer's disease; ADNI-2, second phase of Alzheimer's Disease Neuroimaging Initiative; CN, cognitively normal; EDSD, European DTI Study in Dementia; MCI, mild cognitive impairment; OASIS-3, third release in the Open Access Series of Imaging Studies.

We also compared these two data sets with 100 subjects of the publicly available Human Connectome Project (HCP) (Van Essen et al., 2013), which contains data of younger healthy subjects and provides a reference, helping us to interpret our results in the targeted population.

Lastly, we evaluated our predictive models on held-out subjects from the EDSD (Brueggen et al., 2017), including two cohorts imaged in the cities of Freiburg and Rostock, which were chosen due to their within-cohort consistency of image acquisition parameters.

The data used in this work, as shown in Table 1, are quite heterogeneous in terms of acquisition, even though all MR images were acquired at 3T. Nevertheless, the conductance method uses only the simple diffusion tensor (as opposed to a higher order) model, whose performance has been shown to stabilize after the asymptotic limit of 30 gradient orientations (Jones, 2004). As seen in Table 1, the number of gradient orientations is at least 23 for all our subjects and we expect the derived diffusion tensors to be robust enough for a fair comparison.

Table 1.

Diffusion-Weighted Magnetic Resonance Imaging Acquisition Parameters for the Data Sets Used

Data set	No. gradient orientations	b-value (sec/mm${}^2$ )	Voxel size (mm${}^3$ )	Field strength (Tesla)
ADNI-2	41	1000	1.37 $\times$ 1.37 $\times$ 0.7	3
OASIS-3	23–64	$\le$1000–$\le$1400	1.22 $\times$ 1.22–2.5 $\times$ 2.5	3
			$\times$2–$\times$4	3
HCP	270	1000 + 2000 + 3000	1.25 $\times$ 1.25 $\times$ 1.25	3
EDSD-Freiburg	61	1000	2 $\times$ 2 $\times$ 2	3
EDSD-Rostock	60	1000	0.98 $\times$ 0.98 $\times$ 2.4	3

ADNI-2, second phase of Alzheimer's Disease Neuroimaging Initiative; EDSD, European DTI Study in Dementia; HCP, Human Connectome Project; OASIS-3, third release in the Open Access Series of Imaging Studies.

Demographic and clinical data from these populations were also available: age, sex, diagnosis, cerebral cortical and subcortical volumes, and cognitive scores such as the CDR scale (Morris, 1991) and the MMSE score (Pangman et al., 2000). CDR measures from 0 to 3 the cognitive capabilities of each subject, with 0 being CN and a higher number reflecting higher cognitive impairment. MMSE rates cognitive capabilities from 0 to 30, with 30 being CN and a lower value reflecting higher cognitive impairment.

Figure 2 shows the relationship between these scores and the subject diagnosis in ADNI-2 and OASIS-3. It is to be noted that the ratings differ across diagnoses and data sets. ADNI-2 rates people with AD diagnosis with CDR scales of 0.5 and 1, MCI with 0.5, and CN with 0, but the MMSE scores are overlapping for the three diagnoses. OASIS-3 rates are variable, with MMSE values overlapping across diagnoses and CDR scores. Therefore, the ratings across the data sets are slightly different, and the cognitive scores of MMSE and CDR are used differently.

FIG. 2.

Distribution of the MMSE score for each diagnostic category in each dementia database, color-coded by the CDR values of 0 (blue), 0.5 (green), 1 (orange), 2 (red), and 3 (brown) (refer to Fig. 3, right, for the color map). Diagnostic labels are quoted from the databases. CDR was not available for HCP and EDSD. CDR, Clinical Dementia Rating; HCP, Human Connectome Project; MMSE, Mini–Mental State Examination.

Correlation of conductance values with clinical data

We computed the correlation between mean conductance and other variables, such as age, CDR, and MMSE (shown in Fig. 3), and cortical/subcortical volumes (provided in the Supplementary Materials). Mean conductance consistently exhibited a decreasing trend (negative r) with respect to age and CDR and mostly an increasing trend (positive r) with respect to MMSE.

FIG. 3.

Correlation of mean conductance with age and cognitive scores of CDR and MMSE, color-coded with respect to the CDR: 0 (blue), 0.5 (green), 1 (orange), 2 (red), and 3 (brown). CDR was not available for HCP.

Predictive value of conductance matrices

We then assessed the predictive value of our conductance matrices. We used linear regression to fit on one cohort and predict from a different one, or fit and predict on the same cohort using cross-validation (see the Study of Conductance Matrices section). We first discarded outliers—subjects with mean conductance higher or lower than the average by two standard deviations—in every cohort. We removed three subjects of the ADNI-2 cohort, three, six, two, and six subjects of the different OASIS-3 cohorts, and no subject from the HCP cohort.

In Tables 2 and 3, we show the r and p-values when fitting and predicting age with the same cohort (within-cohort prediction), and when fitting age with a cohort and predicting with a different cohort (cross-prediction), respectively. Table 3 shows significant correlation between predicted and true values of age across ADNI-2 and OASIS-3, but not when fitting using HCP.

Table 2.

Prediction of Age Within ADNI-2, OASIS-3, and HCP

ADNI-2	OASIS-3				HCP
213 subj	272 subj	139 subj	125 subj	116 subj	100 subj
r = 0.33	r = 0.14	r = 0.34	r = 0.56	r = 0.33	r = −0.06
p = 8 × 10 ${}^{-7}$	p = 0.02	p = 6 × 10 ${}^{-5}$	p = 10 ${}^{-11}$	p = 5 × 10 ${}^{-4}$	p = 0.55
p b  = 5 × 10 ${}^{-6}$	pb = 0.14	p b  = 4 × 10 ${}^{-4}$	p b  = 7 × 10 ${}^{-11}$	p b  = 3 × 10−3	pb = 1

p_b stands for Bonferroni-corrected p-value (for six comparisons). p-values under 0.05 with a positive r are highlighted.

subj, subjects.

Table 3.

Prediction of Age in ADNI-2, OASIS-3, and HCP

Predict	ADNI-2	OASIS-3				HCP
Fit	213 subj	272 subj	139 subj	125 subj	116 subj	100 subj
ADNI-2, 213 subj		r = 0.25	r = 0.22	r = 0.29	r = 0.23	r = 0.07
		p = 4 × 10 ${}^{-5}$	p = 0.01	p = 10 ${}^{-3}$	p = 0.02	p = 0.51
		p b  = 10 ${}^{-3}$	pb = 0.33	p b  = 0.03	pb = 0.51	pb = 1
OASIS-3, 272 subj	r = 0.34		r = 0.25	r = 0.26	r = 0.29	r = −0.18
	p = 3 × 10 ${}^{-7}$		p = 4 × 10 ${}^{-3}$	p = 4 × 10 ${}^{-3}$	p = 3 × 10 ${}^{-3}$	p = 0.08
	p b  = 10 ${}^{-5}$		pb = 0.11	pb = 0.11	pb = 0.08	pb = 1
OASIS-3, 139 subj	r = 0.36	r = 0.26		r = 0.29	r = 0.26	r = 0.12
	p = 6 × 10 ${}^{-8}$	p = 2 × 10 ${}^{-5}$		p = 10 ${}^{-3}$	p = 7 × 10 ${}^{-3}$	p = 0.24
	p b  = 2 × 10 ${}^{-6}$	p b  = 5 × 10 ${}^{-4}$		p b  = 0.04	pb = 0.2	pb = 1
OASIS-3, 125 subj	r = 0.37	r = 0.2	r = 0.33		r = 0.35	r = −0.02
	p = 2 × 10 ${}^{-8}$	p = 10 ${}^{-3}$	p = 9 × 10 ${}^{-5}$		p = 2 × 10 ${}^{-4}$	p = 0.8
	p b  = 6 × 10 ${}^{-7}$	p b  = 0.03	p b  = 3 × 10 ${}^{-3}$		p b  = 5 × 10 ${}^{-3}$	pb = 1
OASIS-3, 116 subj	r = 0.49	r = 0.41	r = 0.31	r = 0.34		r = 0.28
	p = 2 × 10 ${}^{-14}$	p = 3 × 10 ${}^{-12}$	p = 2 × 10 ${}^{-4}$	p = 10 ${}^{-4}$		p = 4 × 10 ${}^{-3}$
	p b  = 6 × 10 ${}^{-13}$	p b  = 8 × 10 ${}^{-11}$	p b  = 7 × 10 ${}^{-3}$	p b  = 3 × 10 ${}^{-3}$		pb = 0.12
HCP, 100 subj	r = −0.33	r = −0.16	r = −0.1	r = −0.19	r = −0.3
	p = 5 × 10${}^{-7}$	p = 8 × 10${}^{-3}$	p = 0.24	p = 0.03	p = 10${}^{-3}$
	pb = 10${}^{-5}$	pb = 0.25	pb = 1	pb = 0.93	pb = 0.04

p_b stands for Bonferroni-corrected p-value (for 30 comparisons). p-values under 0.05 with a positive r are highlighted.

When we combined all OASIS-3 data, we got values of $r=0.212$, $p=7\times 10^{-8}$, $p_b=2\times 10^{-7}$ for within-cohort prediction, and values of $r=0.383$, $p=7\times 10^{-9}$, $p_b=4\times 10^{-8}$ when training on OASIS-3 and testing on ADNI-2, and $r=0.265$, $p=10^{-11}$, $p_b=6\times 10^{-11}$ when training on ADNI-2 and testing on OASIS-3, and a negative r when training with OASIS-3 and testing on HCP and vice versa. In this comparison (with all of OASIS-3 in a single cohort), we Bonferroni-corrected (p_b) with a factor 3 in the within-cohort prediction and a factor 6 in the cross-prediction case.

Tables 4–7 show results when similarly predicting the MMSE and CDR cognitive scores. Prediction in these cases yielded less significant p-values, both in within- and cross-cohort predictions.

Table 4.

Prediction of Mini–Mental State Examination Within ADNI-2, OASIS-3, and HCP

ADNI-2	OASIS-3				HCP
213 subj	272 subj	139 subj	125 subj	116 subj	100 subj
r = 0.32	r = 0.11	r = 0.16	r = −0.05	r = 0.28	r = −0.03
p = 3 × 10 ${}^{-6}$	p = 0.08	p = 0.07	p = 0.56	p = 4 × 10 ${}^{-3}$	p = 0.73
p b  = 2 × 10 ${}^{-5}$	pb = 0.46	pb = 0.41	pb = 1	p b  = 0.02	pb = 1

p_b stands for Bonferroni-corrected p-value (for six comparisons). p-values under 0.05 with a positive r are highlighted.

Table 7.

Prediction of Clinical Dementia Rating in ADNI-2 and OASIS-3

Predict	ADNI-2	OASIS-3
Fit	213 subj	272 subj	139 subj	125 subj	116 subj
ADNI-2, 213 subj		r = −0.01	r = 0.03	r = 0.15	r = 0.02
		p = 0.83	p = 0.75	p = 0.1	p = 0.85
		pb = 1	pb = 1	pb = 1	pb = 1
OASIS-3, 272 subj	r = 0.25		r = 0.04	r = 0.08	r = 0.1
	p = 3 × 10 ${}^{-4}$		p = 0.62	p = 0.41	p = 0.32
	p b  = 5 × 10 ${}^{-3}$		pb = 1	pb = 0.82	pb = 1
OASIS-3, 139 subj	r = 0.2	r = 0.18		r = 0.07	r = 0.11
	p = 4 × 10 ${}^{-3}$	p = 4 × 10 ${}^{-3}$		p = 0.45	p = 0.28
	pb = 0.08	pb = 0.08		pb = 1	pb = 1
OASIS-3, 125 subj	r = 0.28	r = 0.27	r = 0.08		r = 0.16
	p = 4 × 10 ${}^{-5}$	p = 6 × 10 ${}^{-6}$	p = 0.36		p = 0.09
	p b  = 8 × 10 ${}^{-4}$	p b  = 10 ${}^{-4}$	pb = 1		pb = 1
OASIS-3, 116 subj	r = 0.33	r = 0.19	r = 0.14	r = 0.13
	p = 9 × 10 ${}^{-7}$	p = 10 ${}^{-3}$	p = 0.1	p = 0.16
	p b  = 2 × 10 ${}^{-5}$	p b  = 0.03	pb = 1	pb = 1

p_b stands for Bonferroni-corrected p-value (for 20 comparisons). p-values under 0.05 with a positive r are highlighted.

Table 5.

Prediction of Mini–Mental State Examination in ADNI-2, OASIS-3, and HCP

Predict	ADNI-2	OASIS-3				HCP
Fit	213 subj	272 subj	139 subj	125 subj	116 subj	100 subj
ADNI-2, 213 subj		r = −0.03	r = −0.02	r = 0.02	r = 0.13	r = −0.01
		p = 0.6	p = 0.79	p = 0.84	p = 0.17	p = 0.89
		pb = 1	pb = 1	pb = 1	pb = 1	pb = 1
OASIS-3, 272 subj	r = 0.29		r = 0.21	r = −0.04	r = −0.02	r = −0.14
	p = 2 × 10 ${}^{-5}$		p = 0.02	p = 0.63	p = 0.84	p = 0.16
	p b  = 6 × 10 ${}^{-4}$		pb = 0.45	pb = 1	pb = 1	pb = 1
OASIS-3, 139 subj	r = 0.11	r = 0.18		r = −0.04	r = 0.07	r = −0.05
	p = 0.11	p = 3 × 10 ${}^{-3}$		p = 0.67	p = 0.46	p = 0.63
	pb = 1	pb = 0.09		pb = 1	pb = 1	pb = 1
OASIS-3, 125 subj	r = 0.22	r = 0.1	r = 0.05		r = 0.13	r = 0.1
	p = 10 ${}^{-3}$	p = 0.1	p = 0.54		p = 0.2	p = 0.3
	p b  = 0.045	pb = 1	pb = 1		pb = 1	pb = 1
OASIS-3, 116 subj	r = 0.19	r = 0.04	r = 0.05	r = 0.1		r = −0.07
	p = 5 × 10 ${}^{-3}$	p = 0.56	p = 0.59	p = 0.3		p = 0.48
	pb = 0.16	pb = 1	pb = 1	pb = 1		pb = 1
HCP, 100 subj	r = −0.06	r = 0.05	r = −0.12	r = 0.02	r = 0.0
	p = 0.39	p = 0.42	p = 0.17	p = 0.86	p = 0.96
	pb = 1	pb = 1	pb = 1	pb = 1	pb = 1

p_b stands for Bonferroni-corrected p-value (for 30 comparisons). p-values under 0.05 with a positive r are highlighted.

Table 6.

Prediction of Clinical Dementia Rating Within ADNI-2 and OASIS-3

ADNI-2	OASIS-3
213 subj	272 subj	139 subj	125 subj	116 subj
r = 0.2	r = 0.04	r = 0.18	r = 0.1	r = 0.25
p = 4 × 10 ${}^{-3}$	p = 0.54	p = 0.04	p = 0.29	p = 8 × 10 ${}^{-3}$
p b  = 0.02	pb = 1	pb = 0.18	pb = 1	p b  = 0.04

p_b stands for Bonferroni-corrected p-value (for five comparisons). p-values under 0.05 with a positive r are highlighted.

When we put all the OASIS-3 data together, we got significant values for within-cohort prediction: $r=0.132$, $p=9\times 10^{-4}$, and $p_b=2\times 10^{-3}$ for CDR, $r=0.091$, $p=0.02$, and $p_b=0.06$ for MMSE. For cross-prediction with ADNI-2: $r=0.321$, $p=2\times 10^{-6}$, $p_b=4\times 10^{-6}$ when training on OASIS-3 and testing on ADNI-2, and $r=0.054$, $p=0.17$, $p_b=0.34$ when training on ADNI-2 and testing on OASIS-3, when using CDR; $r=0.317$, $p=2\times 10^{-6}$, $p_b=10^{-5}$ when training on OASIS-3 and testing on ADNI-2, and $r=0.01$, $p=0.83$, $p_b=1$ when training on ADNI-2 and testing on OASIS-3, when using MMSE. p-values were significant only in one direction: when we trained with all OASIS-3 data together.

CDR was not available for the HCP subjects, and a factor 2 was used in the Bonferroni correction in both the within-cohort prediction and cross-prediction. In the case of MMSE, we had the values for HCP too, and when comparing with all of OASIS-3 in a single cohort, we Bonferroni-corrected (p_b) with a factor 3 in the within-cohort prediction and a factor 6 in the cross-prediction case. We observed no significant values in all cases with positive r (as negative r would not indicate prediction).

Region-specific conductance results

We then correlated the elements of the conductance matrix individually with age, the MMSE score, and CDR, while correcting for multiple comparisons by multiplying the p-values by the number of connections. Figure 4 shows many of the connections being significant (i.e., $p_b<0.05$) for OASIS-3 and ADNI-2. In the case of age, 1342 and 1981 connections out of 3655 (Table 8) had p-values under $0.05$, respectively. Regarding the cognitive scores: for CDR, 683 and 807 connections out of 3655 and for MMSE, 653 and 474 connections out of 3655 had p-values under $0.05$, in OASIS-3 and ADNI-2, respectively. It is worth noting that in general we found more significant values in OASIS-3, possibly due to the larger sample size.

FIG. 4.

Sig-values for the correlation of conductance with age, CDR, and MMSE. We depict the negative logarithm of the Bonferroni-corrected p-value ($sig=-{log}_{10}\left(p_b\right)$), and consider significant values above 1.3 (i.e., $p_b<0.05$). In this figure, all OASIS-3 cohorts were used together.

Table 8.

Number of Significant (Surviving) Connections (of 3655)

	OASIS-3	ADNI-2
Age	1342	1981
MMSE	653	474
CDR	683	807
MMSE after regressing out age and sex	282	91
CDR after regressing out age and sex	417	518

CDR, Clinical Dementia Rating; MMSE, Mini–Mental State Examination.

In OASIS-3, age correlated with connectivity with higher significance among all subcortical regions, especially thalamus and hippocampus, bilaterally, and with regions in the cortex, mostly transverse temporal, cingulate regions, insula, and precuneus cortex, bilaterally. In ADNI-2, age also correlated with higher significance in connections involving thalamus and hippocampus, bilaterally, and with regions in the cortex such as middle and superior temporal, lateral orbitofrontal, posterior cingulate, entorhinal, fusiform, and some other regions in the occipital and parietal cortices.

In OASIS-3, connections found to most significantly correlate with cognitive scores were similar to those correlating with age the most, namely connecting hippocampus, amygdala, insula, and transverse temporal. In ADNI-2, the correlation with age and cognitive scores was different: age was significantly correlated to connections between many regions, whereas the correlation of cognitive scores was significant with connections from hippocampus and amygdala to temporal and cingulum lobes, as well as orbitofrontal.

We then regressed out the age and sex effects before correlating the conductance with the CDR and MMSE scores. By doing so, the general significance levels decreased. The number of significant connections out of 3655 went down, in OASIS-3, from 683 to 417 for CDR and from 653 to 282 for MMSE and, in ADNI-2, from 807 to 518 for CDR and from 474 to 91 for MMSE. The connectivity plots for this experiment are depicted in Figure 5, showing the most significant connections to be similar to those in Figure 4, but with a lower significance level in general.

FIG. 5.

Sig-values for the correlation of conductance with CDR and MMSE, once the age and sex effects have been removed. We show the negative logarithm of the Bonferroni-corrected p-value ($sig=-{log}_{10}\left(p_b\right)$), and consider significant values above 1.3 (i.e., $p_b<0.05$). In this figure, all OASIS-3 cohorts were used together.

External validation on held-out data

For a true external validation on unseen data (without fine-tuning the pipeline), we applied the predictive models that we trained using ADNI-2, OASIS-3, and HCP on the two held-out EDSD cohorts. We removed four subjects from each of the Freiburg and Rostock cohorts due to failure in structural image processing. The results of the prediction of age and MMSE are reported in Table 9. CDR was not available for EDSD.

Table 9.

Prediction of Age and Mini–Mental State Examination in the Held-Out European DTI Study in Dementia Database

	EDSD-Freiburg, 37 subj		EDSD-Rostock, 57 subj
Predict Fit	Age	MMSE	Age	MMSE
ADNI-2, 213 subj	r = 0.49	r = 0.16	r = 0.27	r = −0.21
	p = 4 × 10 ${}^{-3}$	p = 0.36	p = 0.05	p = 0.13
	pb = 0.05	pb = 1	pb = 0.72	pb = 1
OASIS-3, 272 subj	r = 0.17	r = 0.45	r = 0.18	r = 0.06
	p = 0.35	p = 9 × 10 ${}^{-3}$	p = 0.19	p = 0.64
	pb = 1	pb = 0.12	pb = 1	pb = 1
OASIS-3, 139 subj	r = 0.51	r = 0.35	r = 0.2	r = 0.07
	p = 3 × 10 ${}^{-3}$	p = 0.04	p = 0.15	p = 0.61
	p b  = 0.04	pb = 0.6	pb = 1	pb = 1
OASIS-3, 125 subj	r = 0.44	r = −0.3	r = 0.22	r = 0.15
	p = 0.01	p = 0.09	p = 0.12	p = 0.27
	pb = 0.15	pb = 1	pb = 1	pb = 1
OASIS-3, 116 subj	r = 0.47	r = 0.14	r = 0.19	r = −0.17
	p = 5 × 10 ${}^{-3}$	p = 0.45	p = 0.18	p = 0.24
	pb = 0.08	pb = 1	pb = 1	pb = 1
OASIS-3, all 635 subj	r = 0.25	r = 0.4	r = 0.19	r = −0.01
	p = 0.16	p = 0.03	p = 0.17	p = 0.94
	pb = 1	pb = 0.35	pb = 1	pb = 1
HCP, 100 subj	r = 0.13	r = −0.08	r = 0.1	r = −0.02
	p = 0.47	p = 0.64	p = 0.48	p = 0.89
	pb = 1	pb = 1	pb = 1	pb = 1

p_b stands for Bonferroni-corrected p-value (for 14 comparisons for each variable). p-values under 0.05 with a positive r are highlighted.

Anticorrelated connectivity results

For ADNI-2, we computed the cross-subject linear correlation coefficient between all pairs of structural connections, keeping $\left|R^{-}\right|=1978$ pairs for which $r:=R_{i,j}<-0.1$. From those, the correlation between the left cortico/subcortical insula/caudate connection and the left corticocortical precentral/entorhinal connection (Fig. 6, top, left) was most significant ($p=3\times 10^{-6}$, $p_b=6\times 10^{-3}$) with the robust (bisquare) fit slope m = −0.40. (The top 20 significant pairs all involved the insula/caudate connection.)

FIG. 6.

Negative correlation between the insula/caudate and the precentral/entorhinal structural connections in the left and right hemispheres, across the ADNI-2 (top) and 272-subject cohort of OASIS-3 (bottom) populations. CDR values are encoded in the colors of the dots: 0 (blue), 0.5 (green), 1 (orange), 2 (red), and 3 (brown) (refer to Fig. 3).

We then tested whether the same two connections were inversely correlated also in the right hemisphere, which was true with high significance (r = −0.15, $p=0.03$; m = −0.24; Fig. 6, top, right). Since here we tested a specific pair of connections in the right hemisphere, correction for multiple comparisons was not necessary.

Next, for external validation and replication, we tested the hypothesis that the pair of insula/caudate and precentral/entorhinal connections are negatively correlated in the first (largest) OASIS-3 database cohort. This hypothesis was validated on this new data set in both the left (r = −0.26, $p=2\times 10^{-5}$, m = −0.48) and the right (r = −0.23, $p=2\times 10^{-4}$, m = −0.28) hemispheres (Fig. 6, bottom).

We then computed the correlation of the caudate/insula connection with the CDR and the MMSE score in the OASIS-3 database. While the CDR was negatively correlated with mean connectivity as reported in the previous subsection, it was positively correlated with the caudate/insula connection in the left ($r=0.19$, $p=10^{-3}$) and right ($r=0.22$, $p=2\times 10^{-4}$) hemispheres. Likewise, whereas the MMSE score was positively correlated with mean connectivity, it was negatively correlated with the caudate/insula connection in the left (r = −0.12, $p=0.046$) and right (r = −0.12, $p=0.04$) hemispheres.

Null results

In contrast, we did not observe any negative correlation between the insula/caudate and precentral/entorhinal connections across the young-adult HCP subjects⁷. By reversing the order of ADNI-2 and OASIS-3 databases in this experiment, the most significantly anticorrelated pair found in OASIS-3 was not negatively correlated in ADNI-2. In addition, the anticorrelation between the insula/caudate and precentral/entorhinal connections was not observed in OASIS-3 when we included most (652) OASIS-3 subjects, which had heterogeneous scan descriptions (as opposed to the 272-subject cohort).

Discussion

In this work, we used our previously proposed approach (Frau-Pascual et al., 2019b) to compute and analyze structural brain connectivity in dementia populations. This method models structural connectivity as electric conductance, computing it as a weighted sum of all possible paths between two areas, following the information given by the diffusion tensors. We previously showed (Frau-Pascual et al., 2019b) that this method outperformed deterministic tractography in producing structural connectivity that was more correlated with functional connectivity, possibly due to the fact that all paths—including direct and indirect⁸—were considered.

In this work, we used our conductance method and investigated its potential in the study of aging and AD. The conductance model was sensitive to AD-related changes in not only diffusion but also geometric properties of the brain WM. For instance, given that this method takes into account distances and paths, changes in subcortical volumes and cortical thickness could also affect the measured connectivity. The conductance might be affected by the GM and WM volume, as they affect ROI sizes and pathways between a pair of ROIs, respectively. Shrinkage in volume could also draw ROIs closer to each other, producing shorter pathways.

Our results were based on the HCP data for healthy young population, and ADNI-2, OASIS-3, and EDSD data for elderly and AD populations. In total, we analyzed 100 young healthy subjects, and 959 elderly subjects, from whom 153 had been diagnosed with AD dementia, 122 had MCI, and 83 had other types of dementias or pathologies. This was a heterogeneous pool of subjects with various diagnoses scanned at different sites with different acquisition parameters. Such heterogeneity could make the results more robust and allow for replicability analysis, but also introduce variability that could reduce statistical power.

We used independent variables such as the subject's age as well as the cognitive scores of CDR and MMSE, which quantify the progression of dementia for a subject, even though they do not clarify whether dementia is due to AD, aging, or other causes. It is worth noting that, as shown in Figure 2, the ways CDR is defined and used in ADNI-2 and OASIS-3 are not identical, and neither are the relationships between CDR and MMSE in these two data sets, as seen when comparing with diagnosis. We did not consider diagnosis when analyzing our data and focused only on these scores (see however, Fig. 7, left, in the Supplementary Materials for a diagnosis-specific histogram of mean conductance). Previous works have modeled the association between dMRI measures and changes in executive and memory function scores (Scott et al., 2017).

We first summarized the conductance values by averaging them across all region pairs. We considered the independent variables of age, CDR, and MMSE (as well as cortical and subcortical volumes in Fig. 8 of the Supplementary Materials). As illustrated in Figure 3, we found expected trends already in the mean conductance: it correlated negatively with age, negatively with CDR, positively with MMSE, and positively with the volumes, except for subcortical regions in ADNI-2. Interestingly, HCP data with healthy young adults followed the same trends in age and volumes. Similar trends have been reported for diffusion measures such as FA and mean diffusivity (Zavaliangos-Petropulu et al., 2019).

Next, we investigated the predictive power of mean conductance when training a model and predicting from it, either using the same data set or different ones. As seen in Tables 2 and 3, prediction values were significant for age using most cohorts except for HCP, although in some cases we lost significance after the Bonferroni correction. The generally lower accuracy in cross-prediction compared with within-cohort prediction could be due to the difference in scanners and/or inconsistency of imaging protocols used at different imaging sites. Putting all the OASIS-3 data together, in either training or testing, produced significant results. Regarding the HCP, however, most values were not significant, which may have been because the age range of the HCP subjects is not only narrow but also very different from those of the other (dementia) data sets.

Predicting cognitive scores of CDR and MMSE from conductance values, as seen in Tables 4–7, produced significant results mostly only in cohorts with more than 200 subjects. This is probably due to unbalanced values/categories, and also the fact that these scores are variable across data sets and diagnoses. When we considered all the OASIS-3 data together, we achieved significant results when training with OASIS-3. Nonetheless, the fact that the prediction was generally significant in data sets with over 200 subjects suggests that sample size may be playing an important role in the prediction power.

We then considered the region-pairwise connections to see how correlated individual connections were with changes in age, and cognitive scores. Previous works have shown correlations of dMRI-derived measures and cognitive scores in the corpus callosum (Moseley et al., 2002), cingulum (Mito et al., 2018), and temporal lobe (Nir et al., 2013). As described in the Region-Specific Conductance Results section, in OASIS-3 and ADNI-2 the correlation with age was significant for more than a third of the connections, whereas the correlation with cognition was significant in about a fifth of the connections. However, the affected regions were similar in OASIS-3, possibly due to an overlap between cognition and age effects on brain connectivity, and the cognitive decline that accompanies aging.

An interesting question here is how much of the correlation of conductance with CDR and MMSE overlaps with the correlation with age and sex. To clarify this nuance, in a different experiment, we regressed out the effects of age and sex and found residual connections that were significantly correlated with MMSE and CDR. In ADNI-2, we observed different patterns of correlation between age and cognitive scores. However, CDR and MMSE correlation patterns did not change when we regressed out the effects of age and sex. It is worth noting that subject head motion has been found to bias results when comparing groups (Yendiki et al., 2014), and that it is likely that in this case it is correlated with cognitive impairment and age.

Once we were done fine-tuning our analysis pipeline, we applied our trained models to two cohorts of the held-out EDSD database for external evaluation. Table 9 shows that the prediction was fairly successful on the Freiburg cohort, but not on the Rostock cohort. We noticed a general trend that the prediction was more accurate among data sets where age and MMSE strongly negatively correlated, such as ADNI-2 (r = −0.2, p = 4 × 10⁻³), OASIS-3 (r = −0.22, p = 2 × 10⁻⁸), and EDSD-Freiburg (r = −0.64, p = 5 × 10⁻⁵), than in other data sets, that is, HCP (r = 0.04, p = 0.67) and EDSD-Rostock (r = 0.14, p = 0.31). This might mean that for the best prediction results, models must be trained on samples that represent the target population demographically and clinically.

Lastly, we considered the interrelationship among connections.⁹ We found anticorrelation between certain connections, which was consistent across data sets and in both hemispheres. AD patients are known to suffer from connectivity attenuation, and also characterized by brain reorganization and plasticity (Dillen et al., 2016; Kim et al., 2015a). Early in the disease, connectivity within some (i.e., frontal) brain regions increases—possibly due to a compensatory reallocation of cognitive resources—but eventually declines as the disease progresses (Brier et al., 2012; Schultz et al., 2017; Sohn et al., 2014).

The connections that we found to be anticorrelated were the precentral/entorhinal cortex and the insula/caudate connections; that is, the stronger one connection, the weaker the other, across the population. Notably, we observed an increasing trend in the caudate/insula connection strength with respect to CDR.

In fact, there is evidence of increased FA in the left caudate in presymptomatic familial AD subjects (Ryan et al., 2013), increased structural connectivity in the right insula (Ye et al., 2019), and increased functional connectivity between the frontal lobe and the corpus striatum (Supekar et al., 2008) in AD.

In addition, because the conductance method accounts for indirect paths, a possible enhancement in the thalamus and putamen structural connectivity in AD (Ryan et al., 2013; Ye et al., 2019) might also have contributed to the increase in the caudate/insula connectivity. Furthermore, the fact that this negative correlation was observed consistently in the older adults and in those on the dementia spectrum (ADNI-2 and OASIS-3), but not in young healthy adults (HCP), suggests that this significant anticorrelation might be due to progression of dementia or aging.

Although these results do not necessarily imply a compensatory effect at this stage, our approach may prove useful in a study to discover compensatory connections. Including all OASIS-3 subjects (as opposed to only a subset with homogeneous scans) did not externally validate the anticorrelation hypothesis generated from ADNI-2, possibly because the various acquisition parameters created a large variance in the data that dominated the putative effects.

It is important to note that an increase in the measured structural connectivity could stem from factors other than an actual strengthening of the tract. WM atrophy, volume reduction (Ye et al., 2019), and other geometrical variabilities could make ROIs closer to each other, leading to elevated measured structural connectivity. To mitigate this effect, we excluded subcortico/subcortical regions in our study of anticorrelated connections.

In addition, in regions with fiber crossing, selective axonal loss can increase the FA and subsequently measured structural connectivity (Douaud et al., 2011; Kim et al., 2015b; Ryan et al., 2013).

This article focuses on findings from existing dementia populations using our conductance-based connectivity computation method. For an evaluation of this method, we refer the interested reader to our previous comparisons with existing approaches (Frau-Pascual et al., 2018, 2019a, b; Mohammadi et al., 2020).

Conclusion

In this work, we focused on the study of the aging and AD populations through connectomics. We applied our conductance method to several databases to detect brain changes related to aging and AD. Results indicated the predictive potential of the conductance measure, especially for age. Although mean conductance values exhibited the expected trends, the prediction of cognitive scores varied across data sets. An important but not surprising finding was that age and cognitive scores of CDR and MMSE largely overlapped. We also correlated brain connections with each other across populations and discovered significantly anticorrelated structural connections. Future work consists of using longitudinal data to further explore the prediction of cognitive scores, and test the hypothesis that anticorrelated connections are indeed compensatory.

Authors' Contributions

A.F.P. and I.A. did the project design, writing, and analysis. J.A., D.V., A.Y., D.H.S., and B.F. provided substantial feedback that considerably improved the article.

Author Disclosure Statement

B.F. has a financial interest in CorticoMetrics, a company whose medical pursuits focus on brain imaging and measurement technologies. D.H.S. has a financial interest in Niji, a company whose medical pursuits focus on brain health technologies. B.F.'s and D.H.S.'s interests were reviewed and are managed by the Massachusetts General Hospital and Mass General Brigham in accordance with their conflict of interest policies. A.F.P., I.A., J.A., D.V., and A.Y. have no conflicts to disclose.

Funding Information

Support for this research was provided by the BrightFocus Foundation (A2016172S). Additional support was provided by the National Institutes of Health (NIH), specifically the National Institute on Aging (NIA; R56AG068261, AG022381, 5R01AG008122-22, R01AG016495-11, R01AG016495, 1R56AG064027), the BRAIN Initiative Cell Census Network (U01MH117023), the National Institute of Diabetes and Digestive and Kidney Diseases (K01DK101631, R21DK108277), the National Institute for Biomedical Imaging and Bioengineering (NIBIB; P41EB015896, R01EB006758, R21EB018907, R01EB019956), the National Center for Alternative Medicine (RC1AT005728-01), the National Institute for Neurological Disorders and Stroke (R01NS052585, R21NS072652, R01NS070963, R01NS083534, U01NS086625, R01NS105820), and the NIH Blueprint for Neuroscience Research (U01MH093765), part of the multi-institutional HCP.

Computational resources were provided through NIH Shared Instrumentation Grants (S10RR023401, S10RR019307, S10RR023043, S10RR028832) and the O2 High Performance Compute Cluster at Harvard Medical School.

HCP data were provided by the HCP, WU-Minn Consortium (Principal Investigators: David Van Essen and Kamil Ugurbil; 1U54MH091657) funded by the 16 NIH Institutes and Centers that support the NIH Blueprint for Neuroscience Research; and by the McDonnell Center for Systems Neuroscience at Washington University.

ADNI data collection and sharing were funded by the AD Neuroimaging Initiative (NIH grant U01 AG024904) and DOD ADNI (Department of Defense award no. W81XWH-12-2-0012). ADNI is funded by the NIA, the NIBIB, and through generous contributions from the following: AbbVie, Alzheimer's Association; Alzheimer's Drug Discovery Foundation; Araclon Biotech; BioClinica, Inc.; Biogen; Bristol-Myers Squibb Company; CereSpir, Inc.; Cogstate; Eisai, Inc.; Élan Pharmaceuticals, Inc.; Eli Lilly and Company; EuroImmun; F. Hoffmann-La Roche Ltd. and its affiliated company Genentech, Inc.; Fujirebio; GE Healthcare; IXICO Ltd.; Janssen Alzheimer Immunotherapy Research & Development, LLC.; Johnson & Johnson Pharmaceutical Research & Development LLC.; Lumosity; Lundbeck; Merck & Co., Inc.; Meso Scale Diagnostics, LLC.; NeuroRx Research; Neurotrack Technologies; Novartis Pharmaceuticals Corporation; Pfizer, Inc.; Piramal Imaging; Servier; Takeda Pharmaceutical Company; and Transition Therapeutics. The Canadian Institutes of Health Research is providing funds to support ADNI clinical sites in Canada. Private sector contributions are facilitated by the Foundation for the NIH. The grantee organization is the Northern California Institute for Research and Education, and the study is coordinated by the Alzheimer's Therapeutic Research Institute at the University of Southern California (USC). ADNI data are disseminated by the Laboratory of NeuroImaging at the USC.

Supplementary Material

Supplementary Data