Early functional network alterations in asymptomatic elders at risk for Alzheimer’s disease

Amyloid-β (Aβ) deposition is known to starts decades before the onset of clinical symptoms of Alzheimer’s disease (AD), however, the detailed pathophysiological processes underlying this preclinical period are not well understood. This study aimed to investigate functional network alterations in cognitively intact elderly individuals at risk for AD, and assessed the association between these network alterations and changes in Aβ deposition, glucose metabolism, and brain structure. Forty-five cognitively normal elderly subjects, who were classified into Aβ-positive (CN+) and Aβ-negative (CN−) groups using 11C-Pittsburgh compound B PET, underwent resting state magnetoencephalography measurements, 18F-fluorodeoxyglucose PET (FDG-PET) and structural MRI. Results demonstrated that in the CN+ group, functional connectivity (FC) within the precuneus was significantly decreased, whereas it was significantly enhanced between the precuneus and the bilateral inferior parietal lobules in the low-frequency bands (theta and delta). These changes were suggested to be associated with local cerebral Aβ deposition. Most of Aβ+ individuals in this study did not show any metabolic or anatomical changes, and there were no significant correlations between FC values and FDG-PET or MRI volumetry data. These results demonstrate that functional network alterations, which occur in association with Aβ deposition, are detectable using magnetoencephalography before metabolic and anatomical changes are seen.


Participants' inclusion and exclusion criteria
The major inclusion criteria were: 1) Mini-Mental State Examination (MMSE) score >= 24; 2) Clinical Dementia Rating (CDR) = 0; and 3) a normal Logical Memory II score from the Wechsler Memory Scale-Revised (LM2), after adjustment for individual education levels. Because of the different education systems between the United States and Japan, the individual education levels adjustments were slightly different from those of ADNI2 as follows: a) LM2 (paragraph A, which states the maximum score is 25) >= 9 for 16 or more years of education, b) >= 5 for 10 to 15 years of education, and c) >= 3 for 0 to 9 years of education. Individuals under treatment for any significant medical, neurologic, or psychiatric disease, as well as with any history of a major psychiatric disorder, alcohol dependence, or substance dependence, were excluded. Based on MRI findings, individuals with any clinically significant brain focal legions were also excluded.
In addition, we used blood tests to confirm that no participant had abnormal thyroid function or vitamin B1 or B12 deficiency.

PiB-PET: Visual rating and classification
The classification of participants into PiB-positive (CN+) and PiB-negative (CN-) groups was determined by the visual interpretation. This was because we considered that visual interpretation is sensitive enough to detect the very localized amyloid deposition that can occur in the early stage of the amyloid pathology. The method was slightly modified from that reported by Rabinovici et al. 1 as follows: PiB-PET images were visually read by two experienced nuclear medicine physicians (K.I. and T.K.) who were blind to the clinical data. The obtained static images were displayed with a rainbow scale and an inverse gray scale. PiB images were rated as "positive" when the tracer binding in the cortical gray matter was deemed equal to or greater than that in the white matter, and as "negative" when only nonspecific tracer binding in the white matter was observed. If the visual interpretations by the two raters did not match after the independent readings, the cases were discussed and a consensus was reached. In this study, the two raters' judgments were matched in 43 of 45 (95.6%) cases, and only two cases needed to be discussed. In addition, the consensus decisions for these cases were obtained easily.

PiB-PET: Quantitative image analysis
The reconstructed static PET images (168 x 168 x 111 matrices, 2.036 x 2.036 x 2.036mm voxel size) were spatially normalized in Montreal Neurological Institute (MNI) stereotactic space with parameters obtained from individual 3D-T1 MR images coregistered to PiB-PET images by Diffeomorphic Anatomical Registration Through Exponentiated Lie Algebra (DARTEL) 2 . The normalized PiB-PET images were masked with the grey-matter-segmented MR images to exclude the white matter and regions outside the brain. Then the standardized uptake value ratio (SUVR) images were generated by dividing the masked PiB images by the average value in both cerebellar hemispheres on a pixel-by-pixel basis using the regions of interest (ROIs) of the Harvard-Oxford probabilistic atlas 3 . Mean cortical SUVR (PiB-mcSUVR) was obtained by averaging the SUVRs of the frontal, parietal, and temporal Harvard-Oxford ROIs, except for the primary motor and sensory areas. The PiB-SUVR images were spatially smoothed using a Gaussian kernel filter of 8mm at full width at half maximum. With these smoothed SUVR images, voxelwise regression analysis for MEG FC data with the PiB-SUVR images was performed using SPM8.

FDG-PET: Image acquisition
Prior to the FDG-PET examination, all participants fasted for at least 4 hours. After intravenous administration of 18 F-FDG (185 ± 37 MBq), participants were instructed to lie on a bed keeping their eyes opened for a resting period of 30 min in a dimly lit and quiet room. Then a dynamic scan was performed in the 3D mode for 30 min (300 sec frame x 6 times).

FDG-PET: Image interpretation
For image interpretation, the hypo-metabolism in the specific regions was rated as 2+ (definitely present), 1+ (probably present), 0 (equivocal), -1 (probably not present), or -2 (definitely not present). With a rating score greater than 1+, it was deemed to have the Alzheimer's disease-like regional hypo-metabolism. The raters could also refer to corresponding topographical images of FDG-PET and MRIs. If evaluations of the two raters did not match, the cases were discussed, and a consensus reading was reached.
There were only two cases that needed to be discussed.
Because the T1 images were also used for the MEG source modeling, the landmarks for the MEG coordinate systems nasion and bilateral preauricular points were marked with 5.6 mm-diameter vitamin D capsules (ALFAROL, Chugai Pharmaceutical, Tokyo, Japan) to match the MEG and MRI coordinate systems accurately.

MEG: Measurements
The illuminance of the room was set to about 11 lux. The position of four head-position indicator (HPI) coils attached to the scalp, and each subject's headshape relative to three anatomical locations (nasion and both preauricular points) were defined using a 3D digitizer (Fastrak, Polhemus, VT, USA). Subjects' head movements were monitored by these HPI coils, and eye movements were monitored by the vertical and horizontal EOG with two pairs of bipolar electrodes.

MEG: Data preprocessing
The raw recording data were at first submitted to Maxfilter software (v 2.2, correlation threshold = 0.9, time window = 10 seconds) to remove external noise with the temporal extension of the signal space separation method with movement compensation 4 . The 306channel system has 102 channel locations, each of which consists of two orthogonal planar gradiometers and one magnetometer. In this study, we used data only measured at 102 magnetometers for the subsequent analysis. Accordingly, all of the magnetometers' resting state signals were automatically scanned for ocular, muscle, and jump artifacts using Fieldtrip software 5 , and were visually confirmed by an MEG expert (P.C.). The artifact-free data were segmented in continuous 4-second fragments (trials). At least 20 clean trials (80 seconds of brain activity) were obtained from all participants and preserved for further analyses. The number of artifact-free trials for CN+ and CN-groups were 40 ± 17 and 47 ± 13, respectively, and there were no significant group differences (Mann-Whitney p = 0.20). To calculate the source reconstruction, the time series was filtered in the following frequency bands: delta (2-3.9 Hz), theta (4.1-7.9 Hz), alpha (8.1-11.9 Hz), beta (12.1-29.9 Hz), and gamma (30.1-55.0 Hz). The filtering was performed with a Finite Impulse Response filter of order 1500 designed with a Hamming window.
This filter was applied using a 2-pass procedure over the whole 5-minute registers to avoid phase distortion and edge effects.

MEG: Headmodels and source reconstruction
A regular grid of 2455 nodes, with 1 cm spacing, was created in the template MNI brain.
This set of nodes was transformed to each participant's space using a non-linear normalization between the native T1 image (whose coordinate system was previously converted to match the MEG coordinate system) and a standard T1 in MNI space. The forward model was solved with the realistic single-shell model introduced by Nolte 6 .
Source reconstruction was performed with a Linearly Constrained Minimum Variance Beamformer 7 . For each subject, the covariance matrix was first averaged over all trials to compute the spatial filter's coefficients, and then these coefficients were applied to individual trials, obtaining a time series per segment and source location.

MEG: Atlas-based analysis of functional connectivity (FC)
In this study, the FC was measured by means of phase-locking value (PLV), in each frequency band. The 6 DMN ROIs included in this study contained 156 nodes. Thus, the starting data set per each subject consisted of matrices with the following dimensions: 156 nodes x 4000 samples x 5 frequency bands x trials. Then, for each frequency band and trial, we calculated the PLV 8 via the following procedure: first, for each node j = 1…156, the phase of the signal ( ) was extracted by means of a Hilbert transform:  be causing these differences, we calculated the correlation between beamformer weights in both groups to produce an estimate of volume conduction 9 . Beamformer weights did not differ between groups in any frequency band, which makes it unlikely that the functional connectivity differences were caused by volume conduction.

MEG: Statistical analysis for FC
The analytic methodology relied on the cluster based permutation test introduced by Maris and Oostenveld 10 and was carried out independently for each frequency band. The methodology consisted of two steps: (1) an intra-ROI FC analysis that computed the local connectivity within each ROI, and (2) an inter-ROI FC analysis that evaluated the interregional connectivity from the PCu ROI, which is known as one of the DMN hubs, to each ROI. In both cases, the procedure was essentially the same. In the intra-ROI analysis, we analyzed the FC of all the nodes contained within a ROI, whereas in the inter-ROI analysis, we focused on the FC between the nodes in the corresponding two ROIs. The procedure started by assessing the FC difference between groups for each pair of nodes using the Mann-Whitney test. The significance of the links was assessed using nonparametric randomization (5000 permutations) testing 11 . Only those links with p-values < 0.05 were kept and included in the following steps of the analysis. Then, we aimed to extract a robust significant network, also called a motif in graph theory 12 . These motifs consisted of several consecutive, significant links, which systematically showed a diminished or enhanced FC in the CN+ group compared with the CN-group. We considered a motif to be significant only when 1) at least 25% of the nodes that composed the ROI were involved, 2) at least 10% of the links among them had significant FC differences between groups, and 3) the motifs were connected, that is, a path existed between each pair of nodes in the motif 12 . The first two conditions set the minimum dimensions of the motif, and the third condition fixed a constraint in the morphology, dismissing the insulated links. If more than two motifs survived within a single ROI or single inter-ROI region, we selected the largest motif as being representative. Then, to control for the multiple comparisons problem, we estimated a proper null distribution of F-values by randomizing the original data. First, we randomized the group's configuration (maintaining the same number of subjects in each group). Next, we shuffled the FC matrices. In each randomized dataset, we extracted the new motifs and then calculated the corresponding F-values. The F-values, computed by ANCOVA adjusted for age, over each motif in the original data set, were compared with the equivalent measure from the randomized data. Therefore, for each motif, the proportion of randomizations with F-values higher than the ones in the original data corresponded to the permutation p-values. Only motifs with p < 0.05 were kept as significant motifs.  Hypo -0.564*** †The p values were Bonferroni corrected by multiplying with 5 (the number of frequency bands). ‡FC changes in the CN+ group compared with the CN-group. Hypo, decreased connectivity in the CN+; Hyper, increased connectivity in the CN+. §Correlation coefficient (r) for each FC value with mean SUVR value within the DMN ROIs (DMNSUVR). The asterisks indicate statistically significant correlations (* p <0.05, ** p <0.01, *** p < 0.001). PCu: precuneus, rIPL: right inferior parietal lobule, lIPL: left inferior parietal lobule.