The study sample was recruited from three different centers: the Neurology Department in “Hospital Universitario San Carlos,” the “Center for Prevention of Cognitive Impairment,” and the “Seniors Center of Chamartin District” located in Madrid (Spain).

The sample consisted of 131 right-handed and native Spanish speakers. Fifty-one of them were diagnosed as mild cognitive impaired (MCI group), while 80 showed no objective neuropsychological impairment. The latter were further divided in two groups, 39 healthy control elders without any cognitive concern (HC group) and 41 with subjective cognitive decline (SCD group). Table 1 summarizes their demographic data and other relevant characteristics.

To assess the general cognitive and functional status the following set of screening questionnaires were used: The Mini Mental State Examination (MMSE; (Lobo et al., 1979), the Geriatric Depression Scale–Short Form (GDS; Yesavage et al., 1982); the Hachinski Ischemic Score, (HIS; Rosen et al., 1980), and the Functional Assessment Questionnaire (FAQ; Pfeffer et al., 1982).

After the initial screening all subjects underwent an exhaustive neuropsychological assessment including: Direct and Inverse Digit Span Test (Wechsler Memory Scale, WMS-III), Immediate and Delayed Recall (WMS-III), Phonemic and Semantic Fluency (Controlled oral Word Association Test, COWAT), Ideomotor Praxis of Barcelona Test, Boston Naming Test (BNT), Trail Making Test A and B (TMTA and TMTB; Reitan, 1958) and Rule Shift Cards (Behavioral Assessment of the Dysexecutive Syndrome, BADS).

MCI group subjects were diagnosed as MCI according to the criteria established by Petersen (2004) and Grundman (2004). MCI patients did not fulfill criteria for dementia diagnosis.

Cognitive concerns were self-reported by the participants in an interview with clinician experts. The final group assignment was made after neuropsychological evaluation attending to a multidisciplinary consensus (by neuropsychologists, psychiatrists, and neurologists). In order to prevent possible confounders of SCD, problematic medication, psycho-affective disorders or other relevant medical condition lead to the exclusion from the study. Following the recommendations made by the SCD-I-WG, all subjects were older than 60 at onset of SCD, and the onset of SCD occurred within the last 5 years.

The exclusion criteria employed in this study were the followings: (1) history of psychiatric or neurological disorders or drug consumption that could affect MEG activity, such as cholinesterase inhibitors; (2) evidence of infection, infarction or focal lesions in a T2-weighted scan within 2 months before MEG acquisition; (3) a modified Hachinski score equal or higher to 5; (4) a GDS-SF score equal to or higher to 5; (5) history of alcoholism, chronic use of anxiolytics, neuroleptics, narcotics, anticonvulsants or sedative hypnotics. All participants were between 65 and 80 years old at the moment of the MEG acquisition. Besides, additional analyses were conducted to rule out other possible causes of cognitive decline such as B12 vitamin deficit, diabetes mellitus, thyroid problems, syphilis, or Human Immunodeficiency Virus (HIV).

All participants signed an informed consent prior to study enrollment. This study was approved by the “Hospital Universitario San Carlos” ethics committee, and the procedures were performed in accordance with approved guidelines and regulations.

Neurophysiological data was acquired by using a 306 channel (102 magnetometers, 204 planar gradiometers) Vectorview MEG system (Elekta AB, Stockholm, Sweden), placed inside a magnetically shielded room (VacuumSchmelze GmbH, Hanau, Germany) at the “Laboratory of Cognitive and Computational Neuroscience” (Madrid, Spain). All recordings were obtained while subjects were sitting comfortably, resting awake with eyes closed. MEG acquisition consisted of 4 min of signal for each subject.

Head shape was obtained by using a three-dimensional Fastrak digitizer (Polhemus, Colchester, Vermont). Three fiducial points (nasion and left and right preauricular points) and at least 300 points of the surface of the scalp were acquired for each subject. In addition, four head position indication (HPI) coils were placed on the subjects scalp, two in the mastoids and two in the forehead. HPI coils' position was also acquired using the Fastrak device, and continuous head position estimation was used during the recording in order to track head movements. Finally, a vertical electrooculogram of the left eye was used to capture blinks and eye movements.

MEG data was acquired using a sampling rate of 1,000 Hz and an online anti-alias bandpass filter between 0.1 and 330 Hz. Recordings were processed offline using a spatiotemporal signal space separation algorithm (Taulu and Simola, 2006); correlation window 0.9, time window 10 s in order to remove magnetic noise originated outside the head. The algorithm was also used to correct head movements of the subject during the recording.

Ocular, muscular and jump artifacts were first identified using an automatic procedure from the Fieldtrip package (Oostenveld et al., 2011), and then visually confirmed by a MEG expert. The remaining data was segmented in 4 s epochs of artifact-free activity. Subjects with at least 15 clean epochs were selected for further analysis (47.6 ± 7.3 epochs in the HC group, 46.2 ± 9.4 epochs in the SCD group, 42.2 ± 7.0 epochs in the MCI group, mean ± standard deviation). In addition, an ICA-based procedure was employed to remove the electrocardiographic component when it was clearly identified. Due to data redundancy after the spatiotemporal filtering, only magnetometers data were used in the subsequent analysis.

A T1-weighted MRI was available for each subject, acquired in a General Electric 1.5 Tesla magnetic resonance scanner, using 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). MRI images were processed with Freesurfer software (version 5.1.0) and its specialized tool for automated cortical and subcortical segmentation (Fischl et al., 2002) in order to obtain the volume of several brain areas. Hippocampal volumes were selected as anatomical evidences of brain atrophy characteristic for MCI and AD (Dubois et al., 2007), and normalized with respect to the overall intracranial volume (ICV) to account for differences in head volume over subjects.

The source model consisted of 2,459 sources placed in a homogeneous grid of 1 cm in MNI template, then linearly transformed to subject space. Each source was labeled as belonging to one of the 64 areas of the reduced Harvard-Oxford atlas (Desikan et al., 2006). Of the initial 2,459 sources, 970 were discarded due to not being identified as part of any recognized area (i.e., white matter sources). The leadfield was calculated for the remaining 1,489 sources using a three-shell Boundary Element Method (brain-skull, skull-scalp and scalp-air interfaces generated from the subject's T1 MRI) model computed with OpenMEEG software (Gramfort et al., 2010).

Source reconstruction was performed for the alpha band. In order to calculate alpha band in the most representable and comprehensive way for our sample we visually identified the individual alpha frequency (IAF) for every participant as the most prominent alpha peak in the average power spectrum over occipital and parietal channels. The average IAF in the study sample was 9.4 Hz. Then, according to Klimesch considerations about alpha band width (Klimesch, 1999), we set it from 6.9 to 11.4 Hz (i.e., IAF-2.5 Hz − IAF + 2 Hz).

Artifact-free epochs were bandpass filtered in the alpha band using an 1,800 order FIR filter designed using Hanning window. Data was filtered in a two-pass procedure to avoid phase distortion, and using 2,000 samples of data at each side as padding to avoid edge effects. Lastly, a Linearly Constrained Minimum Variance (LCMV) beamformer (Van Veen et al., 1997) was employed to obtain the source time series by using the computed leadfield and the epoch-averaged covariance matrix.

The analysis of FC was performed using the hypothesis of phase synchronization (Rosenblum et al., 2001) and evaluated using the Phase Locking Value (PLV; Lachaux et al., 1999). PLV is calculated using the instantaneous phase difference between a pair of signals. For each temporal point, a vector is constructed with norm unity and phase equal to the phase difference between the instantaneous phases of both signals. Then, PLV-value for each data segment is calculated as the norm of the average vector for the pair of signals k and l: PLVk, l=|1T∑te−j(φk(t)−φl(t))| where φ_k(t) is the instantaneous phase of signal k at instant t, and T the number of temporal points per segment.

In order to reduce the dimensionality of the connectivity matrices, PLV-values are averaged over areas, obtaining a unique PLV-value for each pair of areas defined in the reduced Harvard-Oxford atlas. The final PLV-value between areas A and B is calculated as follows: PLVA, B=1NANB∑NA∑NB|1T∑te−j(φAk(t)−φBl(t))| where N_A is the number of sources in area A, and A_k is the source k inside this area.

In this work the instantaneous phases of the signals were obtained using the Hilbert analytical signal, with 2,000 samples of data at each side as padding in order to avoid edge effects. Mean and standard deviation FC matrices for each group are shown in Figure 1.

We checked the significance of each individual link for each group separately to ensure results interpretability using the analytical method depicted in (Wilkie, 1983; García-Prieto et al., 2017). All the links were significant after FDR correction (Q = 0.05) in the three diagnostic groups (all p < 0.028 for HC, all p < 0.033 for SCD and all p < 0.033 for MCI). Furthermore, in order to evaluate whether between-groups FC differences could be explained by source leakage, the Pearson correlations between beamformer weights were computed for each link and compared across groups (Garcés et al., 2014).

We calculated the FC in two key RSN: DAN and DMN. The coordinates of the nodes forming each RSN were defined in MNI space according to previous work (Watanabe et al., 2013; Jimenez et al., 2016). In turn, ROIs belonging to the DMN were subdivided into an anterior component (aDMN) and a posterior component (pDMN; Andrews-Hanna et al., 2007). The representative FC for each RSN was computed by averaging the FC-values over all pairs of regions comprising the corresponding network.

In order to assess the differences in connectivity between groups we used an ANCOVA analysis with age as covariate. To ensure the accuracy of the results the ANCOVA analysis was non-parameterized using permutations. The original F-value obtained with the ANCOVA analysis and the original grouping were stored, and a series of 10,000 randomizations of subjects were performed, preserving the original group sizes. For each randomization a new F-value was obtained, which allowed creating a null-distribution of F-values adapted to the characteristics of the data. This distribution was used to calculate the permutation-corrected p-value associated to the original F-value.

Due to the high number of comparisons, of the order of 2,000, the probability of obtaining an elevated number of false positives is high. In order to avoid this problem we complemented the ANCOVA analysis with a False Discovery Rate (FDR; Benjamini and Hochberg, 1995) multiple comparisons correction, using a Q = 0.05 (a 5% rate of false positives).

Finally, for those links who survived the FDR correction we performed a pairwise comparison analysis with Tukey's Honestly Significant Difference correction to determine the groups in which the connectivity values were significantly different. These results were permutation-corrected in a similar way that the ANCOVA results.

Average FC-values for each RSN were compared across groups employing a similar procedure with ANCOVA analysis non-parametrized using permutations.

We performed an ANOVA statistical test with FDR, using a Q-value of 0.05, to identify and discard those links whose connectivity value is below the random threshold. Regarding the beamformer weights correlation, we employed a similar procedure to identify links whose correlation between beamformer filters differed among groups.

Last, we conducted correlation analysis in our sample using Pearson coefficient to characterize the relationship between the observed FC changes and cognitive state, as well as hippocampal volume. In order to address the multiple comparisons problem, all p-values were also corrected using FDR.

The FC analysis in this work was performed between anatomical areas, using a reduced version of the Harvard-Oxford atlas, consisting on the regions and abbreviations shown in Table 2. All the links were included in the analysis, as none of them was discarded because of a non-significant PLV-value. The ANCOVA analysis brought significant between-group differences for 17 links in the overall comparison, FDR (Q = 0.05) corrected (Figure 2). When observed in the pairwise comparison, the results showed a posterior network, with decreased connectivity, and an anterior network, with increased connectivity for both SCD and MCI compared to the HC group.

In this comparison between HC subjects and MCI patients, all 17 links showed significant group effect in the pairwise comparison (p < 0.05; Figure 3). The anterior network presented higher FC in the MCI group compared to the HC group in three links connecting anterior regions, including left Inferior Temporal Gyrus, left Paracingulate and left Anterior Cingulate. The posterior network exhibited lower FC in the MCI group, and comprised 14 links between connecting posterior cortical structures such as: temporal medial structures (both hippocampi and right parahippocampus), parietal (left Postcentral Gyrus, both Supramarginal Gyri), and occipital areas (left Frontal Pole, both Superior Occipital cortices, right Inferior Occipital Cortex, right Lingual Cortex).

In this comparison we identified a similar pattern with two sub-networks, as shown in Figure 4. SCD subjects showed increased FC-values respect to the HC group (p < 0.05) in three links connecting the same regions described in the previous comparison. SCD subjects also showed decreased FC in 11 links (p < 0.05). Those links connected both intra- and inter-hemispherical areas between posterior regions. Interestingly, all the links affected in the SCD group, were also disrupted in the MCI group in a similar manner.

MCI patients showed a network comprising four links where FC-values were significantly lower (p < 0.05) compared to SCD FC-values (Figure 5). This network with reduced FC connected temporal, parietal and occipital regions of the brain, and comprised both intra- and inter-hemispheric links.

Regarding DAN we observed a significant group effect (p < 0.05). Pairwise comparisons revealed a FC decrease in DAN with respect to HC group in both SCD (p < 0.05) and MCI (p < 0.01). Mean FC of the DAN was not significantly different between SCD and MCI (Figure 6).

DMN was divided into two sub-networks and compared separately. We did not observe any significant group effect for the aDMN component. On the contrary, we found a significant group effect in pDMN (p < 0.005). FC was significantly decreased in both SCD (p < 0.005) and MCI (p < 0.005) with respect to control. SCD and MCI did not show significant differences in mean FC in this sub-network (Figure 6).

The correlation analysis between the beamformer weights determined there were no differences in the beamformer filters weights of the three groups. Taking this result into account, it is unlikely that the difference in the connectivity metrics could be caused by source leakage (Garcés et al., 2014).

Due to its extended use in the clinical diagnosis of MCI and AD, we completed this work by analyzing the hippocampal volumes of the participants. The corrected hippocampal volumes were compared between groups by means of an ANCOVA using the age as covariate. The contrast revealed significant differences between groups (p < 0.01; Table 1). Pairwise comparisons showed that MCI patients had significantly lower volumes than both SCD (p < 0.01) and HC group (p < 0.01), while the volumes of SCD and HC participants did not differ (p > 0.05).

Some of the links with a reduction in FC in SCD and MCI exhibited positive associations with hippocampal volume, MMSE, inverse digits test and BNT. This means that those subjects with lower FC-values demonstrated an overall worse cognitive status, performed worse in working memory, executive functioning and language, and had smaller hippocampi. On the other hand, all three links with increased FC in SCD and MCI showed negative associations with BNT performance, meaning that those subject with larger FC hyper-synchronization in these specific links performed worse in language and executive functioning. All significant correlations and links are shown in Table 3.

The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.