Participants

The sample consisted of 379 individuals divided into 118 MCI (aged from 58 to 87) and 261 CI participants (aged from 41 to 82). The participants were recruited from the Hospital Universitario San Carlos [17] and from “Centro para Mayores del Distrito de Chamartín”, both located in Madrid (Spain). General inclusion criteria were as follows: a modified Hachinski score ≤ 4, a Geriatric Depression Scale (short form) score ≤ 5, and T1, T2, and diffusion-weighted MRIs within 54 weeks before the MEG recordings (on average, the time period between the MEG and MRI recordings was 3 months) without an indication of infection, infarction, or focal lesions (rated by two independent experienced radiologists [18]). In addition, the criteria for the MCI diagnosis were established according to the NIA-AA clinical criteria [19]. For more information about the diagnostic criteria for MCI, see López et al. [20]. For CI participants, we exclude subjects with evidence of significant hippocampal atrophy in a T1-weighted MRI scan within 2 months before MEG acquisition, as hippocampal atrophy is considered a brain marker associated with neurodegeneration [21]. No one of the participants exhibited a history of psychiatric or neurological disorders other than MCI. Furthermore, we advised subjects to avoid medications that could affect MEG activity, such as benzodiazepines, for 48 h before recordings (A detailed list of the sample characteristics can be found in Table 1).

Table 1 Descriptive measures of the final sampleStandard protocol approvals, registrations, and patient consents

All participants were native Spanish speakers and provided written informed consent. The Institutional Review Board Ethics Committee at Hospital Universitario San Carlos approved the study protocol, and the procedure was performed following the Helsinki Declaration and National and European Union regulations.

Neuropsychological assessment

All participants were screened using standardized diagnostic instruments and received a thorough neuropsychological assessment as formerly detailed in López et al. [20]. The screening consisted of standardized tests that included the Spanish version of the Mini-Mental State Examination (MMSE; [22]), the Geriatric Depression Scale-Short Form (GDS-SF; [23]), and the Logical Memory (I and II) subtest (Wechsler Memory Scale III, WMS-III; [24]).

Due to its effectiveness as a measure of verbal episodic memory, logical memory (LM) is one of the most frequently administered subtests in the Wechsler Memory Scale-III (LM-WMS-III) [24]. In the LM test, the participants presented a text, and the memory ability is divided into immediate recall, delayed recall, and recognition. Our study only included the analysis of the delayed recall score, which consisted of free recall of the passages after a 20 to 30 min delay after the presentation. The narrative nature of the task is sensitive to discriminate between normal aging, MCI [25], and early dementia, due to its tight relationship with other high-level cognitive functions such as episodic memory, conceptual organization, and schema formation [26].

MRI acquisition and volumetric analyses

We used a General Electric 1.5 T system with a high-resolution antenna and a homogenization PURE filter (Fast Spoiled Gradient Echo sequence, TR/TE/TI = 11.2/4.2/450 ms; flip angle 12°; 1-mm slice thickness, 256 × 256 matrix, and FOV 25 cm) to obtain T1-weighted images of our participants. The resulting images were processed using the Freesurfer software (version 5.1.0) and its specialized tool for automated cortical parcellation and subcortical segmentation [27]. The measures that were included in further analyses were total gray matter, total cerebral white matter, and hippocampus (in mm3). The volumes of bilateral structures were collapsed in order to obtain a single measure for each region.

Diffusion tensor imaging

The same scanner was also used to collect diffusion-weighted images (DWI) (single-shot echo planar sequence, TE/TR 96.1/12,000 ms; NEX 3 for increasing the SNR; 2.4-mm slice thickness, 128 × 128 matrix, and 30.7 cm FOV). We acquired 1 image with no diffusion sensitization (i.e., b0 images) and 25 DWI directions (b = 900 s/mm2).

DWI images were processed using probabilistic fiber tractography which was run on the automated tool AutoPtx (https://fsl.fmrib.ox.ac.uk/fsl/fslwiki/AutoPtx) as in Verdejo-Román et al. [28]. Due to its relation to memory performance, we studied the relation between FC, DR, and fractional anisotropy (FA) at the uncinate, forceps major, and forceps minor.

Magnetoencephalography

MEG data was recorded using a 306-channel whole-head MEG system (Vectorview, Elekta AG, Finland), placed in a magnetically shielded room located at the Center for Biomedical Technology in Madrid, following the protocol described in de Frutos-Lucas et al. [29]. First, we applied the Maxfilter software (temporal extension of the signal space separation method, correlation window of 10 s, and correlation limit of 0.9) to remove external noise. Then we used FielTtrip software 28 to automatically scan the data for artifacts, which were visually confirmed by an MEG expert. Artifact-free data were segmented in 4-s epochs, plus 2 s of real data at each side as padding.

Afterwards, we estimated the source level activity for each individual. As source model, we used a 1 cm homogeneous grid of source positions defined in MNI space and labeled according to the automated anatomical labeling (AAL) atlas. This source model consisted of 1202 positions in 78 cortical areas and was transformed to subject space using a linear transformation between the template and the T1-weighted MRI of the participant. This image was also used to generate a single-shell head model defined by the inner skull surface. Then, we combined the head model, the source model, and the sensor definition to create a lead field using a modified spherical solution. As the last step, we used a linearly constrained minimum variance beamformer as inverse method.

We estimated FC by means of the phase locking value (PLV), a phase synchronization metric that evaluates the distribution of the phase difference between two-time series. Briefly, after source reconstruction, the dataset consisted of matrices of 1202 nodes by 4000 samples by epochs for each of the 4 frequency bands studied here. Then, for each frequency band and epoch, we calculated the PLV [30] via the following procedure: firstly, we used the Hilbert transform to extract the instantaneous phase φj(t) for each node j = 1…1202 and time t = 1…4000 ms:

$$_\left(t\right)=_\left(t\right)+i\bullet Hilbert\left(_\left(t\right)\right)=_\left(t\right)\bullet ^_\left(t\right)}$$

Secondly, we estimated the synchronization between each pair of signals \(j\) and \(k\) by means of their difference of phases \(_\left(t\right)\) and \(_\left(t\right)\) using the following expression:

$$PLV= \frac\left|\sum\limits_^^_\left(_\right)-_(_)\right)}\right|$$

where T = 4000 is the number of samples in the time series (4 s per epoch at 1000 Hz sampling rate). Lastly, we averaged the PLV matrices across epochs to obtain a more robust estimator of resting-state FC.

This algorithm provided symmetrical whole-brain matrices of 1202 × 1202 nodes per participant and frequency band (theta, between 4 and 8 Hz; alpha, between 8 and 12 Hz; beta, between 12 and 30 Hz; and gamma, between 30 and 45 Hz) [31]. Then, we calculated the nodal strength (also known as weighted global connectivity), which is defined for each node as the sum of its FC with the rest of the nodes. To account for the number of links, the strength of each node was then normalized by dividing the number of links connected to it. This procedure resulted in one brain map of normalized node strengths per each participant and frequency band.

APOE genotype

Genomic DNA was extracted from 10 ml blood samples in ethylenediaminetetraacetic acid. Detection of APOE genotype was performed with TaqMan technology using an Applied Biosystems 7900 HT Fast Real-Time PCR machine (Applied Biosystems, Foster City, CA). See the genotyping method previously described in Cuesta et al. [32] for more information. All the sample was included independently of APOE genotype in the initial analysis. Then, to evaluate the potential moderation role of genotype, the participants were classified as APOE ɛ4 carriers and noncarriers (i.e., ɛ3ɛ3). Participants who presented less frequent allele combinations (i.e., ɛ2ɛ2, ɛ2ɛ3, ɛ2ɛ4, and ɛ4ɛ3ɛ4) were excluded from the sample.

Statistical analysesFunctional connectivity strength (strength FC)

Cluster-based permutation test (CBPT) was carried out separately for each frequency band [33]. We defined a cluster as a set of spatially adjacent nodes that presented a significant partial correlation (Spearman correlation using age as a covariate, p < 0.001) in the same direction between the strength FC values and each DR variable. In this framework, a cluster can be considered as a functional unit. Only clusters including at least 1% of the grid (i.e., a minimum of 12 nodes) were considered. The Spearman rho values were transformed into Fisher Z values, and the cluster-mass statistics were computed as the sum of the Z values of all nodes within the cluster. The p value for each cluster was calculated in a nonparametric fashion, using a null distribution generated by the mass of the main cluster obtained over 5000 random permutations (shuffled versions) of the data [30]. Only those clusters that resulted significant (p < 0.05) after this step were considered in further analyses. Then, we used the average of the strength FC values of the members of the cluster to obtain a representative FC marker. Of note, this FC marker would be indicating that the global FC of the possible significant clusters appeared to be associated with memory performance.

Seed-based analyses (seed link FC)

To examine whether the strength FC results were caused by global or region-specific effects, we performed complementary seed analyses, using the previous clusters as seeds. For it, we calculated the average FC of each source position with the sources in the cluster. Then, we repeated the statistical CBPT analysis using these seed-based FC values instead of the strength FC values.

Correlations between FC and measures of white matter integrity and brain volume

We used the aforementioned cluster markers in subsequent correlation analyses with measures of AD-specific signatures. As to this, we used both the whole sample and a stratification of the cohort by diagnosis (MCI and CI). To account for multiple comparisons, the resulting p values were corrected using a false discovery rate (FDR). All statistical analyses were carried out using MATLAB R2020b (MathWorks Inc.).

Moderation analysis

Additionally, we analyzed the impact of FC on DR focusing on the possible influence exerted by APOE genotype. Each group was divided into ɛ4 carriers and ɛ4 noncarriers to evaluate multiple regression analysis (using age and years of education as covariates), and we calculated the increase in variance explained after including the interaction into the model. Then, we examined the effect of FC on delayed recall scores in the APOE genotype subgroups and used the Johnson-Neyman technique to identify the threshold where the synchronization shows a dysfunctional pattern in CI (CI33/CI34) and MCI (MCI33/MCI34). Statistical analyses were carried out using Process Macro, an extension for SPSS that calculates X’s direct, indirect, and total effects on Y and unstandardized and standardized regression coefficients, standard errors, t, p values, and R2 for the models [34].