Data used in the preparation of this article were obtained from the Alzheimer’s Disease Neuroimaging Initiative (ADNI) database.¹ The ADNI was launched in 2003 as a public–private partnership, led by Principal Investigator Michael W. Weiner, M.D. The primary goal of ADNI has been to test whether serial magnetic resonance imaging (MRI), positron emission tomography (PET), other biological markers, and clinical and neuropsychological assessment can be combined to measure the progression of mild cognitive impairment (MCI) and early Alzheimer’s disease (AD).

In February 2017, we remotely accessed the Laboratory of NeuroImaging’s Image Data Archive at the University of Southern California and searched for suitable participants to include in our anatomical connectome cohort: ADNI participants whose brains were imaged with both structural magnetic resonance imaging (MRI) and diffusion-weighted imaging (DWI) at a single study visit. We sought age-matched groups across the four diagnoses of Cognitively Normal (CN), Early Mild Cognitive Impairment (EMCI), Late Mild Cognitive Impairment (LMCI), and probable AD (AD). This resulted in 42 participants per group. We downloaded the unprocessed structural MRI (3.0T, T1-weighted, non-accelerated IR-SPGR; GE Medical Systems) and DWI (Axial diffusion tensor imaging; GE Medical Systems) for the first available suitable visit of these 168 participants from the ADNI2 phase of ADNI (including: 5 rollovers from the first ADNI phase, ADNI1; and a single rollover from the ADNI Grand Opportunity phase, ADNIGO). Herein, we refer to our anatomical connectome cohort as The168. We also downloaded associated demographics data and metadata in CSV format. Demographics for The168 are summarized in Table 1, and for all ADNI2 (at ADNI2 baseline) in Table 2.

Our disease progression modeling requires both a control and a patient group, which we defined as amyloid-negative CN participants and amyloid-positive AD participants, respectively (see bold figures in Table 1). The threshold for amyloid positivity was chosen as an amyloid PET (Florbetapir ¹⁸F-AV-45, hereafter AV45) cut-point from the literature (28): AV45 Standardized Uptake Value Ratio SUVR ≥ 1.10, which was based on the upper 95% confidence interval for young healthy subjects. Using this criteria, in The168, we identified 26 controls and 38 patients out of a possible maximum of 42 each. For the wider ADNI2 cohort (including rollovers from ADNI1 and ADNIGO), we found 106 controls and 122 patients—see the bold figures in Table 2.

Structural connectomes were generated using tools provided in the MRtrix3 software package,² customized to work with the Geodesic Information Flows algorithm (29) for segmentation and parcelation. The pipeline included DWI denoising (30), preprocessing (31, 32), and bias field correction (33); inter-modal registration (34); T1 tissue segmentation (29); spherical deconvolution (35, 36); probabilistic tractography (37) utilizing anatomically constrained tractography (38) and dynamic seeding (39); spherical deconvolution informed filtering of tractograms (SIFT) (40); T1 parcelation (29); and robust structural connectome construction (41). We used the dwiintensitynorm script in MRtrix3 during DWI preprocessing, but found it necessary to modify the subsequent usage of the population_template script such that DWI masks were not used to create the template (the DWI were pre-masked). Note that we use SIFT to produce biologically plausible tractograms where the streamline density in each voxel matches the fiber orientation distributions estimated from the DWI, instead of simply thresholding the number of fibers connecting two regions.

Our anatomical connectome for each participant is a weighted adjacency matrix that includes only inter-node connections across 130 regions of interest consisting of cortical and subcortical gray-matter regions (including striatal), plus the cerebellum and brain stem. Weights, or connection strengths, are normalized to [0, 1], and so represent within-participant anatomical connection density. The image analysis pipeline is visualized in Figure 1. The 130 regions of interest are a subset of those in the labeling protocol used by Geodesic Information Flows (29, 42), which is a modified version of the Desikan–Killiany–Tourville protocol (43).

For each participant and region of interest, we calculated 12 brain connectivity metrics from the anatomical connectomes, using the Brain Connectivity Toolbox (24) in MATLAB, after appropriate normalization using the weight_conversion function. The local network metrics fall into the following broad categories:

Hubs (basic centrality): degree, strength, degree z-score;

For each region of interest, the medians of the controls and patients distributions for each network measure were statistically compared using a Mann–Whitney–Wilcoxon rank-sum test. Only measures where p < 0.05/12 (Bonferroni-corrected within region) were retained for further analysis within the event-based model of disease progression—we refer to such biomarkers as having “disease signal.”

Note that local efficiency and local clustering coefficient return similar information to each other, as do eigenvector centrality and PageRank centrality, and so one may ask whether we are including redundant information in our models. We argue that we are not, as seen in our results where: (1) local efficiency contained disease signal, whereas local clustering coefficient did not and (2) the centrality measures appear in different positions within the model sequences.

The event-based model (EBM) (8, 10) is a data-driven approach for probabilistically sequencing a cross-sectional set of scalar measurements (“biomarkers”) in the order in which they become observably abnormal. In this context, an “event” constitutes a biomarker appearing more abnormal/diseased than normal/healthy. The EBM is able to estimate an average sequence of disease progression events from cross-sectional data because the proportion of abnormal measurements within a cohort will decline in concert with the average ordering. That is, the biomarker that changes earliest (the first disease event) will contain the highest proportion of abnormal measurements (from affected and presymptomatic individuals), and so on. The EBM fuses multiple biomarker measurements across individuals, with the simplest versions assuming a single disease progression sequence for all individuals, as done here. Determining biomarker abnormality in a data-driven manner necessitates mixture modeling within biomarkers, for which we use a new method (25), described below in Section 2.4.1. For convenience, we normalize all biomarkers to “c-scores” (standardized z-scores relative to controls) that increase with abnormality. Otherwise, we used the same fitting procedures as in Ref. (10). Cross-validation of our EBMs was estimated by refitting the sequence (but not the event measures) to 100 separate bootstrap samples from the data.

We planned to build four EBMs of network connectivity changes, corresponding to the three broad categories of connectivity in Section 2.3, plus all biomarkers together:

“EBM0”—non-network biomarkers.

EBM0 acts as a reference point and for investigating consistency with previous EBMs of AD (10, 45) and includes only non-network biomarkers: average cortical level of amyloid (from AV45 PET) and hypometabolism (from fludeoxyglucose (FDG) PET), test score on the Mini-Mental State Examination (MMSE) (46), and structural MRI volumes of the hippocampus, entorhinal area, ventricles, and whole brain. Biomarkers were adjusted for healthy age, education, and gender using regression (residuals method, controls only). Brain volumes were also adjusted for intracranial volume. We did not fit EBM3 because there was no disease signal (see Section 2.3) in regional clustering coefficients nor eccentricity.

Figure 2 shows that global network connectivity in health and disease did not differ significantly in our cohort, The168. That is, we found no significant group-level difference between the 26 controls (AV45-positive CN) and 38 patients (AV45-negative AD) across four global brain network metrics: density (connectedness), efficiency (segregation), transitivity (segregation), and assortativity (resilience). The null hypothesis in each Mann–Whitney–Wilcoxon rank-sum test was accepted using a Bonferroni-corrected significance level of p = 0.05/4 = 0.0125: density p = 0.0147; transitivity p = 1; efficiency p = 0.145; assortativity p = 0.363.

Brain network hubs are thought to experience increased susceptibility to AD pathology due to higher activity and metabolism, such as regions within the DMN, e.g., Ref. (50), and references within. Our experimental results are consistent with this idea, suggesting that the earliest abnormality in AD occurs in hubs (Figure 6) and other centrally important regions (Figure 7) in the left temporal lobe, and bilaterally in the hippocampus—see Figure 8. Specific left temporal lobe regions include the transverse temporal gyrus (part of the auditory cortex), the temporal pole (which may be involved in social and emotional cognition, e.g., because it has been shown to be affected in frontotemporal dementia (51)), and the triangular part of the inferior frontal gyrus (which may be involved in semantic memory (52)).

Our finding of early connectivity changes involving the auditory cortex is potentially of particular clinical relevance to Alzheimer’s disease: hearing impairment has recently emerged as a major risk factor for cognitive decline and a focus of intense epidemiological interest (53), while functional alterations in central auditory processing have been identified in both presymptomatic and established Alzheimer’s disease (54, 55). Our findings suggest how such alterations may fit within the pathogenic cascade of Alzheimer’s evolution.

Our results also suggest that this deterioration of anatomical connectivity in the brain network may be detectable prior to bulk estimates of amyloidosis, such as the mean SUVR across the cortex in AV45 PET. However, prior to claiming that network-based biomarkers may form sensitive early biomarkers of AD, we would seek confirmation using larger anatomical connectivity cohorts (see future work in Section 4.3).

Our experiments also identified that the anatomical connectivity of some DMN regions is not affected until later in the pathological cascade of AD. These included hubs in the angular gyrus, and hubs and centrality in the medial cingulate gyrus—see Figure 8.

We note that the ordering of brain regions involved in our sequence of anatomical connectivity changes to the Alzheimer’s disease brain (Figure 8) does not follow the ordering of regions in the neuropathological sequence identified by Braak and Braak (56). This suggests that changes in anatomical connectivity may not occur in sync with deposition of abnormal protein in brain tissue. Indeed, our results suggest that connectivity changes may appear before bulk deposition of amyloid across the cortex is detectable in vivo.

The brain’s anatomical connectivity network is altered by AD. These disease-related connectivity changes may be due to neurodegeneration in gray matter, alterations in deep white matter, or alterations in superficial white matter (57) located near the gray-matter/white-matter boundary. We designed our pipeline to be sensitive to these disease-related alterations by employing anatomically constrained tractography (38) and the SIFT method (40), which should improve the biological accuracy of the tractography-based connectivity estimates.

We have studied the late onset, typical variant of Alzheimer’s disease. It will be of interest to apply our technique to explore genetic factors that might influence the progression of Alzheimer’s disease; and to young onset Alzheimer’s disease where different phenotypic presentations are more commonly seen.

Our approach differs from previous network spreading models in two very important ways. We investigated dynamic pathological disruption of the elderly brain’s anatomical connectivity network as a function of AD progression. Previous network models of AD (14, 19–22) used static connectivity patterns estimated from young and healthy individuals, then employed only end-stage patterns of pathology, such as atrophy, to study network spreading mechanisms. Thus, the previous approaches ignore both the effects of normal aging and the effect of disease progression on the network substrate being used to predict pathology.

Our approach provides utility beyond contributing to our understanding of the AD pathological cascade. These include the ability to estimate a longitudinal disease progression signature that is useful for patient staging (10), and for identification of candidate biomarkers for early diagnosis.

In this respect, we believe that our work may be the first data-driven quantification of the widely held idea that AD progression reflects specific anatomical network disruption, not just neurodegeneration.

Future work will include seeking a separate, larger anatomical connectivity cohort upon which to validate our results. While our own cross-validation experiments reported above provide some level of confidence in our conclusions, external validation on a larger cohort would be all the more convincing. Validation on animal models of AD is a related idea that might be worth pursuing.

Our analyses could be applied to other neurodegenerative diseases. We will actively seek suitable cohorts from other diseases upon which to examine pathological changes in the connectivity of the human brain. This would subsequently extend quite naturally to include application of the models to differential diagnosis, and to the related problem of within-disease subtyping (7, 58).

On the highest level, our work involved comparison of graphs (59). There are subtle challenges to comparing graphs and metrics derived from them, even when the number of nodes is constant and the average number of connections is the same (60), as was the case in our study. In the future, we will consider normalizing our connectivity metrics with those derived from random graphs or by looking at distances between graphs (60).

We have sequenced the progressive deterioration of anatomical connectivity in the Alzheimer’s disease brain. We believe that this is the first attempt to do so in a data-driven manner that incorporates the effects of both aging and disease progression. Aging is accounted for by using healthy elderly brains as controls. Disease progression is estimated using event-based modeling.

Our experimental results reveal new insights into Alzheimer’s disease progression: that degeneration of anatomical connectivity in the brain may be a viable, even early, biomarker and should be considered when studying such neurodegenerative diseases.