Memory performance-related dynamic brain connectivity indicates pathological burden and genetic risk for Alzheimer’s disease

The study sample included 37 (13 females and 24 males) cognitively normal Swiss-German elderly adults (mean age (SD) 73 (6.6) years, range 62–89 years, mean education (SD) 14 (3) years, range 8–20) living in the canton Zurich from an ongoing study [51], who received ultra-high field strength MRI at 7T and neuropsychological follow-up after 2 years (Table 1). Genotyping of the APOE gene (rs429358 and rs7412) revealed 13 carriers with at least one ApoE-ε4 allele (two subjects had the genotype ε4/ε4). All study procedures were carried out in concordance with regulations issued by the local ethics authority (Kantonale Ethikkommission Zürich, www.​kek.​zh.​ch), as well as good clinical practice (GCP) guidelines and the declaration of Helsinki [52]. Written informed consent was obtained from all participants before inclusion in the study. Inclusion criteria were age between 55 and 80 years, no significant cognitive impairment as indicated by Mini-Mental State Examination (MMSE) <26, no acute medical or neurological comorbidities, and no present psychiatric disorder or current substance abuse. Exclusion criteria included evidence of infarction and focal or significant hemorrhagic lesions in the MRI, as indicated in detail previously for neuroimaging studies on nondemented elderly populations at our center [8, 41, 51, 53].

All subjects were evaluated by neuropsychological tests as described previously [51]. For the current study, neuropsychological follow-up was scheduled after 2 years and included German language versions of the MMSE as a general measure of cognitive performance [54], as well as domain-specific testing using the Digit Span Forward (DSF) and Digit Span Backward (DSB) for measuring working memory [55], the delayed recall Verbal Learning and Memory Test (VLMT) [56] for assessing episodic memory, the abbreviated CERAD Boston Naming Test (BNT) on confrontational word retrieval [57, 58], and the ratio between part B and A of the Trail Making Test (TMT) as a measure of executive function [59]. To characterize the performance of each measure over time, the percentage difference between follow-up value and baseline was used to obtain yearly variability ratios (relative change per 365 days) for each participant and each investigated cognitive domain. Subjects were allocated to the group “Decline” if the respective yearly variability ratios were negative and to the group “No Decline” if ratios were equal or better than at inclusion (Table 1).

Participants underwent one session of scanning using a Philips 7-Tesla Achieva whole-body scanner (Philips Healthcare, Best, The Netherlands) equipped with a Nova Medical quadrature transmit head coil and 32-channel receive coil array, located at the Swiss Federal Institute of Technology (ETH) in Zurich, as reported previously [8, 41, 53]. For the current study, this scanner was used to acquire T1-weighted MP2RAGE image [60] data (TR/TE = 4.8 ms/2.1 ms, voxel size = 0.6 × 0.6 × 0.6 mm³, SENSE-factor = 2 × 1 × 2, scan duration = 7:50 min) for referencing and automated image segmentation. Resting-state fMRI data was acquired using a 3D T2-prep gradient recalled echo (GRE) sequence, optimized for ultra-high field strength acquisition [36] (TR = 2 s, TR_GRE/TE_GRE = 3.08 ms/1.6 ms, voxel size = 1.5 × 1.5 × 1.5 mm³, scan duration = 7:03 min). Iron load was measured by QSM-MRI at ultra-high field strength using a multi-echo three-dimensional gradient recalled echo (GRE) sequence with three echoes (TR/TE/ΔTE = 23/6/6 ms, flip angle = 10°, voxel size = 0.5 × 0.5 × 0.5 mm³, SENSE-factor = 2.5 × 1 × 2, flow-compensated, scan duration = 13:48 min) for acquiring anatomical MR measures of magnitude and phase for calculation of QSM images using a previously described pipeline [8, 29].

Cerebral Aβ was measured with ¹¹C-PiB-PET [8, 41, 53]. Briefly, participants were administered a dose of 350 MBq of the ¹¹C-labeled tracer intravenously and cerebral amyloid deposition was estimated based on late frame signals representing 50–70 min. Measures of individual brain Aβ load were derived from the ratio of standardized uptake values (SUV) of regional PiB referenced to cerebellar SUV after coregistration using the PMOD brain tool (PNEURO) software, Version 3.4 (PMOD Technologies Ltd., Zurich, Switzerland).

Structural and functional images were preprocessed with a standardized in-house-developed preprocessing pipeline [61] implemented in MATLAB scripts (MATLAB 2015b, Version 8.6; MathWorks Inc., Natick, MA, USA), which included functions from SPM8, SPM12 (http://​www.​fil.​ion.​ucl.​ac.​uk/​spm/​) and DPARSFA toolboxes [62]. Images were spatially realigned and smoothed (FWHM = 5 mm). Nuisance variables were regressed out from the regional time series (these included linear and quadratic trends, six head motion parameters, average cerebrospinal fluid signal from ventricular masks and white matter signal from white matter masks). All the functional data was postprocessed using the Artifact Detection Tool (ART; www.​nitrc.​org/​projects/​artifact_​detect/​), resulting in exclusion of three subjects from further analysis due to excess motion (>2 mm). The remaining 37 functional volumes were regionally parcellated using the automated anatomical labeling (AAL) atlas [63], resulting in the definition of 90 anatomical grey-matter regions of interest. Regional mean time series were then extracted by averaging the preprocessed BOLD signal over all voxels in each region and filtered with a band-pass filter cutoff (0.017–0.15 Hz), as performed previously for investigating dynamic functional networks at rest [45].

A whole brain dynamic functional connectivity (FC) matrix for each subject was calculated using a sliding time window approach (window size = 30 TRs = 60 s; step =1 TR = 2s) to calculate the time-varying correlations between BOLD fluctuations in distinct brain regions, as reported previously [45, 64]. Briefly, in order to account for spurious fluctuations while preserving meaningful fluctuations within the time windows, the window size was determined by calculating 1/smallest frequency in the data, represented by the high-pass filter cutoff [64]. The FC matrices obtained for all windows were then vectorized and temporally concatenated to obtain a dFC (connections × time) matrix, which was normalized by first removing the global mean and dividing for the global standard deviation and secondly applying a row-wise demeaning to focus on the exclusive contributions of dynamics. Finally, dFC data were concatenated into a (connections × time*subjects) matrix and principal component analysis (PCA) was used for reducing data dimensionality and to define subject specific eigenconnectivities, referring to distinct connectivity patterns [45] (i.e., building blocks of dFC, highlighting dominant patterns of FC increase/decrease that recur across time and in the population). For the current study, the following properties of dynamic network connectivity were investigated: 1) Time-contributed weights and percentage of positive weights were obtained for identified connectivity patterns as a measure of dynamic FC changes and compared between groups for inferring expression of connectivity patterns within the acquired fMRI data (“network expression”). Effects of cognitive decline and ApoE-ε4 status on changes in dynamic FC network expression were assessed using a multivariate Hotelling’s T2 test. Singular value decomposition with bootstrapping matched to a Procrustes transform [65] was conducted to test the reproducibility of the eigenconnectivity patterns [66]. Through this process, the stronger connections are attributed higher values and the weaker connections are gradually discarded. This eases interpretation without affecting group comparison tests and these data were therefore used for the visualization of the eigenconnectivity networks. 2.) Node strengths for significant connectivity patterns were calculated as a graph theoretical measure indicating involvement of a particular region in a network, as derived from the sum of all connection weights for all connections attached to a given node [67].

All data processing was performed in MATLAB (2015b, Version 8.6) and its Statistics and Machine Learning Toolbox (Version 10.0). A Hotelling’s T2 multivariate test in combination with sequentially rejective Holm-Bonferroni correction [68] for investigating variability indices pertaining to six neuropsychological tests was used to assess relationship with expression of dynamic connectivity patterns. To follow-up on significant associations, secondary analysis investigated: 1) interactive effects of test performance and ApoE-ε4 on network alterations using a multivariate Hotelling’s T2 test; and 2) relationships between node strength of significant connectivity patterns and local Aβ and iron load, respectively, by applying false discovery rate (FDR)-corrected permutation-based multivariate analysis of variance (MANOVA) [69]. Effect sizes of differences between the “Decline” and “No decline” groups were estimated using Cohen’s d [70].

Participants were neuropsychologically tested at baseline and after 2 years (mean (SD) follow-up time between testing, 719 (277) days). All participants had normal levels of general cognitive performance both at inclusion (mean (SD) MMSE 29.03 (1.17)), as well as at follow-up (mean (SD) MMSE 29.03 (1.19), mean average change per year 0.39%) and for the entire population (n = 37) no significant difference between baseline and follow-up could be observed for the investigated neuropsychological tests. For stratifying the study population by cognitive performance over time, a subgroup of “decliners” was defined that included subjects with negative yearly variability ratios for each cognitive domain. Decliners thus included four subjects for the Boston Naming Test (average yearly change –5.53%), 13 subjects for the Digit Span Forward Test (average yearly change –14.09%), 17 subjects for the Digit Span Backward Test (average yearly change –11.69%), 18 subjects for Trail-Making Test ratio (average yearly change –18.84%), and 18 subjects for the VLMT delayed recall test (average yearly change –9.31%) (Table 1). By generating a Venn diagram, a small degree of overlap between decline in the respective domains could be visualized (Fig. 1), indicating no participant with concurrent decline in all five tests and only two subjects with declined performance in four out of five tests. Moreover, the null hypothesis that the performed tests for domain performance assessed one common factor was dismissed by means of a factor analysis (χ²(5) = 11.995, p < 0.04).

Individuals characterized as having memory decline showed reduced expression of an anterior-posterior connectivity pattern, which remained significant (alpha = 5%) after applying Holm-Bonferroni correction for multiple testing (Roy’s maximum root F(9,27) = 3.23, p < 0.0078; factor loading –0.521; Fig. 2a and b). No significant effects (Roy’s maximum root test, indicated are uncorrected p values without adjustment for multiple testing) on dynamic FC were found for groups defined by the cognitive tests Boston Naming (F(9,27) = 0.53, p = 0.85), Digit Span Forward (F(9,27) = 1.33, p = 0.27), Digit Span Backward (F(9,27) = 0.97, p = 0.48) and Trail Making Test B/A (F(9,27) = 0.84, p = 0.6).

The observed relationship between memory decline and anterior-posterior network expression was followed up for potential relationships between node strength of significant connectivity patterns and neuropathology, by permutation MANOVA. This approach yielded a set of five nodes where nodal contributions to the network were associated with significantly higher susceptibility, as an indicator of local iron, in subjects with lower memory performance at follow-up, compared to those with equal or better performance (F(1,35) = 13.1, p < 0.022, FDR-corrected; Fig. 3c). The most distinct increases for local iron were observable for the left Precuneus (Cohen’s d = 0.493), right Caudate (d = 0.410) and right anterior Cingulate (d = 0.160) (Fig. 4). While no significant relationship between regional network contribution and local Aβ load could be observed for any of the identified nodes (p = 0.128, uncorrected), the association between node strength in the anterior-posterior network and memory decline was not significant when correcting iron levels for Aβ (t(27) = 12.49, p = 0.1).

A multivariate Hotelling's T2 test was used to investigate whether the observed change of dynamic connectivity in a context of memory decline relates to the individual risk of developing AD. Participants with both ApoE-ε4, as well as declined memory over 2 years (n = 6), were tested against the rest of the sample (n = 31), resulting in significant differences in the percentage of positive weights between groups (Roy’s maximum root F(9,27) = 2.22, p < 0.005). Secondary testing of ApoE-ε4 carriers against noncarriers within the group of subjects with declined memory indicated a nominally significant difference in the percentage of positive weights between groups (Roy’s maximum root F(9,8) = 5.29, p < 0.03). Here, the factor loadings of each eigenconnectivity point to three connectivity patterns that drive this group difference: the global network expression (factor loading –0.5), a fronto-temporal network (factor loading –0.44), and a fronto-occipital network (factor loading –0.48). These findings indicate a significant difference in specific network expression associated with ApoE-ε4 carrier status and memory decline, concerning: 1) a pattern showing the alteration of global connectivity (first eigenconnectivity, positive pattern); 2) the alteration between anterior-posterior network connectivity (negative pattern) and interhemispheric fronto-temporal (positive pattern) connectivity (second eigenconnectivity); and 3) the alteration between parieto-temporal connections (negative pattern) and fronto-occipital connectivity (positive pattern) (Fig. 5). The negative sign of the factor loadings indicate that these networks show an increase in the negative patterns and decrease in the positive patterns.