Dynamic network model reveals distinct tau spreading patterns in early- and late-onset Alzheimer disease

We enrolled 142 participants who completed baseline and follow-up tau PET examinations at Gangnam Severance Hospital from January 2015 to March 2019. All participants underwent two PET (¹⁸F-flortaucipir for tau and ¹⁸F-florbetaben for amyloid-beta (Aβ)) and magnetic resonance imaging (MRI) scans, and neuropsychological tests [36] at both baseline and follow-up. Based on baseline Aβ-positivity as determined by the agreement of two nuclear medicine specialists, validated visual assessments [37, 38], and neuropsychological tests, baseline Aβ-positive cognitively impaired (CI) individuals with amnestic presentation were identified using diagnostic criteria supplied by the National Institute on Aging and Alzheimer’s Association (“mild cognitive impairment due to AD with intermediate or high likelihood” for prodromal AD and “probable AD dementia with evidence of the AD pathophysiological process” for AD dementia) [39, 40]. We referred to symptomatic patients included in the AD spectrum as CI individuals. Cognitively unimpaired (CU) individuals were healthy volunteers who achieved normal cognition on neuropsychological tests and for whom no abnormality was evident in MRI at baseline. According to the age at onset, the CI group was divided into early-onset (EOCI: onset age < 65 years) and late-onset (LOCI: onset age ≥ 65 years) groups. Onset age was determined through an interview with family members or caregivers of each CI individual. Similarly, the CU group was also divided into young (YCU: baseline age < 65) and old (OCU: baseline age ≥ 65 years) groups. Ultimately, 30 YCU, 44 OCU, 15 EOCI, and 53 LOCI individuals were enrolled in this study.

Images from PET scans were acquired in a Biograph mCT PET/CT scanner (Siemens Medical Solutions, Malvern, PA, USA) for 20 min at 80 min after injection of ¹⁸F-flortaucipir and 90 min after injection of ¹⁸F-florbetaben. After correcting for attenuation with computed tomography images, three-dimensional (3D) PET images were reconstructed using the ordered-subsets expectation maximization algorithm in a 256 × 256 × 223 matrix with 1.591 × 1.591 × 1 mm voxel size. A 3.0 Tesla MRI scanner (Discovery MR750; GE Medical Systems, Milwaukee, WI, USA) was used to produce axial T1-weighted brain scans with 3D-spoiled gradient-recalled sequences (512 × 512 matrix with voxel spacing of 0.43 × 0.43 × 1 mm).

Using FreeSurfer 5.3 software (Massachusetts General Hospital, Harvard Medical School; http://​surfer.​nmr.​mgh.​harvard.​edu), participant-specific volumes-of-interest (VOIs) were created with T1-weighted MRI scans as described in our previous study [41]. In brief, MRI scans were resliced to FreeSurfer space (a 256 × 256 × 256 matrix with 1 mm isovoxels) and then corrected for inhomogeneity. After segmentation of gray and white matter, 3D surfaces were created with trigons. Finally, participant-specific composite VOIs were created with the cortical areas parcellated using curvature information under the guidance of the Desikan–Killiany atlas [42], and subcortical regions were segmented using probabilistic registration [43].

Statistical parametric mapping 12 (Wellcome Trust Centre for Neuroimaging, London, UK) and in-house software implemented in MATLAB 2017b (MathWorks, Natick, MA, USA) were used for integrative processing of ¹⁸F-flortaucipir PET images. These images were first co-registered to MRI counterparts in FreeSurfer space, and then corrected for partial volume effect (PVE) using the region-based voxel-wise method [44]. Finally, we created PVE-corrected standardized uptake value ratio (SUVR) images with the cerebellar crus median obtained from spatially normalized PET images as a reference, and regional SUVR values for the regions defined by the Desikan–Killiany cortical atlas [42].

Regional ¹⁸F-flortaucipir SUVR values were then converted to W-scores representing regional tau burdens and compared with controls after adjusting for covariates [45‐48]. Multiple linear regression models were created for each region with the regional SUVR values as the outcome and baseline age, sex, and years of education as predictor variables in the Aβ-negative CU group. Residuals were then calculated for each participant accompanied by the individual outcome and predictor variables and divided by the standard deviation (SD) of the residuals obtained from the Aβ-negative CU individuals. The same regression models were applied to calculate W-scores for the follow-up data, and annual changes in W-score (ΔW/year) were calculated.

An illustrative figure of the proposed methods is represented in Fig. 1. All participants were sorted in an ascending order by the number of regions with a baseline tau-PET W-score greater than 2.5 [41, 49], creating a pseudo-longitudinal order of disease progression. Participants with the same number of supra-threshold regions were sorted additionally by the median value of W-scores across all regions. To investigate stage-dependent tau propagation networks, we selected a subgroup of subjects using the sliding window method, in which we moved a window for a fixed number of ordered subjects. We applied the sliding window to two age groups: YCU/EOCI (n = 45) and OCU/LOCI (n = 97). The sliding window was designed to include 40 subjects and moved by one subject from the left (earlier in the pseudo-longitudinal order) to the right (later in the pseudo-longitudinal order). In the case of YCU/EOCI, the window was designed to include 20 subjects due to the smaller number of subjects in the group.

Directional graph theory regression (DTGR) was applied to all subjects in the pseudo-longitudinal order. A longitudinal model of DTGR was used to infer inherent tau spreading networks between two regions taking into account network temporal directionality [50]. For each sliding window, we calculated Spearman correlation coefficients between the baseline W-scores in a seed region (i) and the annual W-score change rates in another region (j). Because those two regional values were ordered temporally with each other, the correlation coefficient worked as a weight of an edge from region i to j. We called this directed network a tau-spreading network for each sliding window. For the tau-spreading networks of all sliding windows, we selected the edges with a P-value < 0.005, but a range of different P-value thresholds were also tested to evaluate reliability. We considered only positive coefficients implying that current accumulation of tau in one region may affect future accumulation in the connected region, which is referred to as tau spreading. Furthermore, the connecting edges were excluded if the target region has a negative mean annual change indicating a decrease in tau over time. The sliding windows were divided into five segments, and mean tau-spreading networks for each segment were constructed using only consistent edges that more than half the windows have within the segment to prevent analytical disturbances from unstable edges along disease progression.

Using the resulting tau-spreading networks, the brain regions were grouped into sets such that regions were more densely connected to each other than by chance (Fig. 1b). A modular structure (community) was identified based on modularity maximization [51]. To consider temporal dependency across the sliding windows, we defined a multilayer modularity for the directed graph [52, 53]:where μ is the total edge weight in the network, A_ijs is the adjacency between node i and j in layer s, γ_s is the weight of intralayer connections (structural resolution parameter), P_ijs is the component of the corresponding null model matrix, δ_ij is the Kronecker delta symbol, C_jsr is the connection strength between node j in slice s and slice r (interlayer coupling parameter), and g_is is the community assignment of node i in layer s. We set parameters γ and ω to 1, which is a frequently used default value [54‐56].

A Louvain-like greedy community-detection algorithm [52, 57, 58] was used to determine the optimal modularity function, Q. This optimizing method was iterated until the resulting community structure did not change from one iteration to the next, and a post-processor function was applied to ensure convergence [59]. Due to the heuristic nature of the algorithm, we repeated the optimization process 1000 times for each group. We then constructed a representative community structure based on a comparison with null models to deal with the degeneracy [60]. We first constructed a regional association matrix (a frequency matrix in which any two regions are assigned to the same community across the repetitions) from original assignments and randomly permuted assignments. We obtained a thresholded regional association matrix by subtracting the maximum value of the random association matrix from the original association matrix. A representative assignment was determined by conducting a Louvain-like algorithm with the thresholded association matrix.

Each community was then characterized by its hub regions, which were expected to lead overall spreading of tau pathology within the community. We performed a seed-based analysis for a community of each tau-spreading network. Assuming that a higher number of paths departing from a seed region indicates a greater ability to provide pathology to the connected regions, the regional out-degree was calculated as the number of edges that originated from the seed region. We considered a region with a relatively higher out-degree (> mean + 1.5 SD of out-degrees from all intracommunity seed regions) as a tau-providing hub. Because a region with only a few influential connections can be identified as a hub due to a low number of pathways overall within the community, we excluded regions with fewer than three connections or those appearing as hubs only in a single window.

We used MATLAB 2019a for statistical analysis of demographic data. For between-group comparisons, a Wilcoxon rank sum test was used for continuous variables, and chi-square tests were used for categorical variables.

Although the OC group exhibited slightly lower mini-mental state examination (MMSE) scores compared with the YC group (p = 0.0036), no differences in sex ratio, years of education, clinical dementia rating sum-of-boxes (CDR-SB), frequencies of the ApoE ε4 genotype, or follow-up intervals between the older groups (OCU or LOCI) and their corresponding younger groups (YCU or EOCI) were evident. The ranges of cognitive decline (MMSE and CDR-SB) in both CI groups are detailed in Fig. S1. Sex and years of education did not differ between the CI and corresponding CU groups, but the CI groups were older (younger group: p = 0.0052, older group: p = 5e−4) and had worse MMSE (younger group: p = 6e−6, older group: p = 3e−10) and CDR-SB (younger group: p = 1e−10, older group: p = 7e−19) scores, higher frequencies of the ApoE ε4 genotype (younger group: p [statistics] = 8e−4 [11.250], older group: p [statistics] = 0.0015 [10.067]), and shorter follow-up intervals (younger group: p = 0.0038, older group: p = 6e−5) compared with their corresponding CU groups. Detailed demographic characteristics are provided in Table 1.

Figure 2 depicts the differential accumulation pattern for tau between the YCU/EOCI and OCU/LOCI groups. When descriptively compared with the baseline W-score and its annual change rate maps within the first segment of the YCU/EOCI group, tau first accumulated in the medial and lateral temporal and the inferior parietal cortex, and extended to the precuneus and posterior cingulate cortices in segment 3 to 4. The annual change rate in the parietal cortex was similar to that of the temporal cortex. Overall, YCU/EOCI group experienced more dramatic accumulation of tau in the diffuse cortical regions (Fig. 2a). However, tau first appeared in the medial temporal lobes of members of the OCU/LOCI group, followed by the inferior temporal and fusiform cortex in segment 2 and 3. It then expanded to the posterior cingulate and inferior parietal cortex in segment 4 and finally reached the remaining cortical regions in segment 5. The maps for the annual change rate in W-scores exhibited patterns similar to those of the baseline maps, but prominent changes were restricted to the temporal regions (Fig. 2b).

In the identified tau-spreading networks of the YCU/EOCI group, edges originated primarily in the temporal cortex, followed by the limbic and parietal cortices, but were seldom found in the frontal and occipital cortices. These appearances were generally robust across a range of different P-value thresholds for edge selection (Fig. S2). In segment 4, tau extended from the temporal cortex to the parahippocampal and insula cortices, and weakly to some regions in the parietal cortex, including the inferior parietal cortex, precuneus, and supramarginal gyrus. Conversely, some of the edges that departed from the entorhinal and isthmus cingulate cortex of the limbic cortex reached the temporal cortex. Several pathways originating in the parietal cortex were also remarkable. In segment 5, tau spreading became much more active between the widespread brain regions, even in connections to the frontal or occipital cortices, while keeping the source regions relatively active in the earlier segments (Fig. 3a).

Scans of the OCU/LOCI group revealed different patterns. Most of the out-edges appeared first in the limbic and temporal cortices and predominated in the temporal cortex in the latter part of the windows, which was similarly reproduced at other edge thresholds (Fig. S2). Tau began to spread from the entorhinal and parahippocampal cortices to the inferior temporal and fusiform cortices. Edges between those regions were most active even in segment 4. In segment 5, tau spreading was widespread, predominantly from the temporal cortex, and from the inferior temporal gyrus in particular. Tau burden in the temporal cortex affected the amount of accumulation in the frontal regions, inferior parietal, supramarginal, and cingulate cortices (Fig. 3b). To rule out the possible influence of the different proportion of CI participants and sliding window size, we reproduced these findings using randomly chosen participants from the OCU/LOCI group with the same number and proportion with those of the YCU/EOCI group (see Fig. S3).

Across the sliding windows in each group, the tau-spreading networks presented as a converged community structure. The YCU/EOCI group was characterized by three major communities. The regional composition for each community changed slightly across the windows, but presented as a mostly exclusive collection that provided hubs (Figs. 4 and 5). For the first community, the hubs were found mainly in the temporal cortex, including the fusiform and inferior temporal gyri in the windows for segment 4 (“Fu-IT driven”). Left banks of the superior temporal sulcus (BSTS), entorhinal (En), and temporal pole cortices were selected as hubs for the second community (“En-BSTS driven”), and left inferior parietal, isthmus cingulate, middle temporal, precuneus, and supramarginal cortices were identified for the third community (“parietal driven”). As shown in Fig. 4, tau spreading increased first within the Fu-IT driven community along with the earliest emergence of its hubs. The En-BSTS and parietal driven communities followed the Fu-IT driven community in the latter part of the segment 5.

On the other hand, three major communities were identified in the OCU/LOCI group. Left and right entorhinal cortices explained other regions’ tau accumulation changes within the first community in the middle part of segments 2 and 3 (“En-PH(L) driven”) but did not persist in the latter part. By contrast, left/right parahippocampal (PH) cortices, which were identified for the first and second communities, remained active in the middle to last windows (“En-PH[L]/PH[R] driven”). In the latter windows, providing hubs were found for the third community, including both the inferior and middle temporal and right temporal pole cortices (“lateral temporal driven”). Each community in the OCU/LOCI group appeared at different periods, while three major communities of the YCU/EOCI group exhibited nearly identical rising and fading patterns throughout the course of disease severity. These findings were broadly replicated in a subset of participants for the OCU/LOCI group adjusted for the proportion of CI participants and sliding window size in the YCU/EOCI group (see Fig. S3).