Data analyzed in this study was previously published in a work on static functional connectivity by Farràs-Permanyer et al. (2019). Ten Mexican patients with MCI diagnosis were recruited in collaboration with Laboratorio Clínico, Centro Integral de Diagnóstico Médico of Guadalajara's Grupo Río Center (Jalisco, Mexico), Instituto de Neurociencias de la Universidad de Guadalajara (Jalisco, Mexico) and Universitat de Barcelona (Barcelona, Spain). All patients met (Petersen et al., 1999) criteria for memory complaints, objective memory impairments, normal general cognitive function, and unaffected daily life activities, and were assessed through Mini-Mental State Examination (MMSE), the Prospective and Retrospective Memory Questionnaire (PRMQ), the Geriatric Depression Scale (GDS), The Pfeffer Activity Questionnaire (PAQ), the Clinical Dementia Rate (CDR), the Boston Naming Test (BNT) and the NEUROPSI (Attention and Memory).

Patients were matched in sex, age group (less than 65, from 66 to 80, more than 80) and education level (up to 6 years, from 6 to 12 years, more than 12 years) in 1:1 ratio to healthy controls screened during patient recruitment.

Patients were excluded from the study in presence of illiteracy, inability to understand the protocol or undergo neuropsychological tests, relevant psychiatric disorders, history of cerebrovascular accidents, alcohol or substance abuse, MRI incompatibilities or advanced cognitive deterioration, dementia or other neurodegenerative diseases other than MCI. Although 27 participants were initially included in the study, three participants in the control group and four participants in the MCI group had to be discarded because the absolute root mean square movement in their rs-fMRI sequences was above half a voxel (Power, Barnes, Snyder, Schlaggar, & Petersen, 2012). Therefore, the remaining 10 participants in the MCI group and 10 participants in the control group were finally analyzed.

Written informed consent was obtained from every individual prior to taking part in the study, according with the Declaration of Helsinki and by the institutional ethics committee. Moreover, this procedure was approved by the Bioethical Committee of the University of Barcelona (03/10/2017).

Neuropsychological assessment was performed in two stages. During a screening stage, participants were administered the Mini-Mental State Examination (MMSE; Folstein, Folstein, & McHugh, 1975) to assess their time and place orientation, attention, calculation, and language and visual construction, while the Prospective and Retrospective Memory Questionnaire (PRMQ; Smith, Della Sala, Logie, & Maylor, 2000) was used to detect prospective and retrospective memory slips in everyday life. More specifically, we used the Spanish adaptation of the MMSE (Blesa et al., 2001) with the adjustments for the Mexican population proposed by Villaseñor-Cabrera, Guàrdia-Olmos, Jiménez-Maldonado, Rizo-Curiel, and Peró-Cebollero (2010), and the Mexican adaptation of the PRMQ (González-Ramírez & Mendoza-González, 2011).

During this stage, the Pfeffer Activity Questionnaire (PAQ; Pfeffer, Kurosaki, Harrah, Chance, & Filos, 1982) was also administered to evaluate their performance of daily activities, if a relative of the participant was also present.

In a second stage, those individuals who were willing to continue their participation in the study went through a complete neuropsychological evaluation that included the administration of the Boston Naming Test (BNT) to assess the language ability (Fernández & Fulbright, 2015), NEUROPSI for memory and attention (Ostrosky-Solís, Ardila, & Rosselli, 1999) and Geriatric Depression Scale (GDS; Brink et al., 1982; Fernández-San Martín et al., 2002).

MR image acquisition and preprocessing protocol has been previously described in Farràs-Permanyer et al. (2019). Functional magnetic images were obtained with a Philips Ingenia 3.0-T system at the Laboratorio Clínico, Centro Integral de Diagnóstico Médico of Guadalajara's Grupo Rio Center (Jalisco, Mexico). A T1 weighted Turbo Field Echo (TFE) structural image was obtained for each subject with a 3-dimensional protocol (repetition time [TR]=2800ms, echo time [TE]=6.3ms, 170 slices, and field of view [FOV]=240×240×170) with sagittal plane acquisition. A functional image T2* weighted (BOLD) was obtained (TR=2000ms, TE=30ms, FOV=230×230×160, voxel size=2.4×2.4×4mm, 29 slices) with transverse plane acquisition.

Structural image was preprocessed using an FSL (FMRIB Software Library v5.0) pipeline adapted under authorization from the authors (Diez et al., 2015), with its parameters adjusted to fit our data, including a motion correction procedure to solve the undesired head's movements in the fMRI sessions according to the same procedure used in Calderón et al. (2020). All non-brain tissue was removed to obtain an anatomic brain mask that would be used to parcel and segment the T1 data images. Moreover, T1 images were reoriented and a resampled AC-PC aligned image with 6 degrees of freedom and all structural data images were corregistered to the normalized space using the Montreal Neurological Institute (MNI) reference brain (Ashburner & Friston, 1999). In this case, the use of a standard atlas such as MNI can be a source of uncontrolled error. To avoid this possible effect, it was studied whether normalization in the MCI group with atlas MNI generated different results if normalization based on a DARTEL model was used. Both processes led to the same results, so the MNI atlas was maintained to guarantee replication options. As usual, the preprocess, apart from maximizing the signal-to-noise ratio, eliminates the presence of extracerebral physiological activities, white matter, and Cerebrospinal Fluid. Regarding the use of low and high pass filters for the elimination of outliers, the elimination of participants with excessive movement allowed confirming that the analysis of outliers was unsuccessful since the final frequencies all ranged between 0.1 and 0.8Hz.

Data treatment and statistical analysis of the participants’ demographic and clinical characteristics, as well as their preprocessed fMRI sequences, were performed with R software (version 3.5.0).

Network Change Points in functional connectivity dynamics were detected through NCPD method (Cribben & Yu, 2017) which performs a spectral clustering to extract the network community structure and principal angles are used as the criterion for change points in the network community structure. Statistical significance of the candidate change points is tested with stationary bootstrap. R function for the NCPD method were provided by the authors and used under their authorization.

For each participant we applied the Automated Anatomical Labelling (AAL) atlas to the preprocessed fMRI sequences to obtain time series of 300 time points (TR=2) for 90 cortical and subcortical regions. Network change points were detected for each participant through spectral clustering with 7 expected communities and a minimum distance of 50 time points between candidate change points. Statistical significance of these change points was tested through stationary bootstrap with 1000 iterations and an outlier threshold of α=.05.

Significant change points were used to segment each participant's rs-fMRI time series. In each segment, we computed the network weighted adjacency matrix though the Pearson correlation between the BOLD signal of each pair of ROIs. These matrices were filtered through TMFG algorithm (Massara, Di Matteo, & Aste, 2016) to obtain a sparse weighted graph between each significant change point.

We characterized the topological properties of each individual's Functional Connectivity networks though their network strength, transitivity, characteristic path length, small-worldness and modularity. Differences in these measures were analyzed with mixed linear models with clinical group and time State as a fixed effects and participant cluster as a random effect. Models were adjusted by neuropsychological variables.

In order to account for the difference in scale between the response variable and the continuous predictors, all continuous numerical variables, including the response, were standardized. Therefore, the model estimates must be interpreted in standard scale.

Network connectivity density was studied with three different approaches. First, we studied the total strength of the network to characterize the global density of the functional connectivity networks in each time segment of each participant. Secondly, we computed the median node strength to identify the central tendency of each network. Median was preferred over mean as it is a robust estimator of central tendency in non-symmetric distributions, as was the case in the node strength distribution of all the participants in this study. Finally, we analyzed the intra-patient variance of the node strength of the networks.

In relation to the total network strength (Figure 4A), the best model in terms of the reduction of the AIC was the baseline model with group and state in the linear predictor (Table 2, Model 1). Although the effect of the state was close to significance (βˆ=0.12; 95% CI: −0.02, 0.26), the marginal R2=.033 indicates an extremely low predictive value of the fixed effects.

Figure 4.

Network strength for control and MCI participants along the fMRI sequence.

(0.22MB).

As an ancillary analysis, we adjusted the models incorporating the results of the neuropsychological assessment to study if these variables modified the effects of group and time. In the case of network total strength, we detected no statistically significant effects of MINIMENTAL (βˆ=0.16; 95% CI: −0.20, 0.53), PRMQ (βˆ=0.29; 95% CI: −0.21, 0.79), PAQ (βˆ=0.29; 95% CI: −0.10, 0.68), GDS (βˆ=0.04; 95% CI: −0.50, 0.58), BNT (βˆ=−0.14; 95% CI: −0.51, 0.22). However, the effect of NEUROPSI was close to statistical significance (βˆ=−0.35; 95% CI: −0.75, 0.06).

Regarding the median node strength, the best model in terms of the reduction of the AIC included the fixed effects of group, state and PRMQ (Table 3, Model 3).

Although the effects of group (βˆ=−0.57; 95% CI: −1.48, 0.34) and state (βˆ=0.03; 95% CI: −0.13, 0.18) were not statistically significant, we detected a significant effect of PRMQ where an increase of one unit implied a 0.53 increase (95% CI: 0.08, 0.97) in the network median strength. However, according to the marginal R2, the fixed effects of this model can only explain a 13.30% of the variability in the response variable. The changes in median node strength in both groups can be seen in Figure 4(B).

We detected no statistically significant effects of MINIMENTAL (βˆ=0.13; 95% CI: −0.23, 0.149), PAQ (βˆ=0.25; 95% CI: −0.13, 0.64), GDS (βˆ=0.18; 95% CI: −0.34, 0.70), BNT (βˆ=−0.11; 95% CI: −0.47, 0.24), or NEUROPSI (βˆ=−0.23; 95% CI: −0.63, 0.17).

As to the intra-patient variance of the node strength, the best model in terms of the reduction of the AIC included the effects of group, state and the interaction between group and state (Table 4, Model 2).

According to this model, there was a statistically significant positive effect (βˆ=0.34; 95% CI: 0.05, 0.63) of the interaction between group and state on the variability of the network strength. However, the marginal R2=.104 indicates a low predictive value of the fixed effects.

The changes in intra-individual variability of the node strength in both groups can be seen in Figure 4(C).

As a complementary analysis, we adjusted the models incorporating the results of the neuropsychological assessment to study if these variables modified the effects of group and time. In the case of intra-patient variance of node strength, we detected no statistically significant effects of MINIMENTAL (βˆ=0.17; 95% CI: −0.17, 0.50), PRMQ (βˆ=−0.07; 95% CI: −0.53, 0.39), PAQ (βˆ=0.22; 95% CI: −0.14, 0.58), GDS (βˆ=−0.19; 95% CI: −0.67, 0.30), BNT (βˆ=−0.23; 95% CI: −0.55, 0.09), or NEUROPSI (βˆ=−0.35; 95% CI: −0.71, 0.01).

The best model for the characteristic path length of the networks in terms of the reduction of the AIC included only the effects of group, state and NEUROPSI score (Table 5, Model 3). Although the effects of group and state were not statistically significant, an increase of one standard unit in the NEUROPSI score implied a 0.39 (95% CI: 0.06, 0.71) increase in the characteristic path length of the network. The marginal R2=.123 indicates a low predictive capacity of the fixed effects.

The incorporation of PAQ scores in the model increased the marginal R2 to 0.184 but failed to reduce the deviance of the model (Table 5, Model 4). However, given the limitations of our sample size, this effect should be considered in further research.

The fluctuations in the characteristic path length in both groups can be seen in Figure 5(A).

Figure 5.

Characteristic path length, transitivity and Small-worldness. (A) Characteristic path length, (B) transitivity and (C) small-world coefficient. Each line represents a single individual and brakes in the lines correspond to changes in the network measure after a significant change point.

(0.2MB).

Ancillary analysis concluded that there was no effect of MINIMENTAL (βˆ=−0.15; 95% CI: −0.45, 0.15), PRMQ (βˆ=0.19; 95% CI: −0.33, 0.43), GDS (βˆ=0.26; 95% CI: −0.16, 0.67) and BNT (βˆ=0.07; 95% CI: −0.22, 0.37).

The best model for the transitivity of the networks in terms of the reduction of the AIC included only the effects of group and state (Table 6, Model 1). However, none of these effects were statistically significant and the marginal R2=.022 indicates a low predictive capacity of the fixed effects. The fluctuations in the transitivity coefficient in both groups can be seen in Figure 5(B).

Complementary analysis led to conclude the absence of significant effects of NEUROPSI (βˆ=−0.28; 95% CI: −0.66, 0.10), MINIMENTAL (βˆ=0.21; 95% CI: −0.12, 0.54), PRMQ (βˆ=0.36; 95% CI: −0.11, 0.83), PAQ (βˆ=0.29; 95% CI: −0.08, 0.65), GDS (βˆ=0.13; 95% CI: −0.38, 0.64) and BNT (βˆ=0.08; 95% CI: −0.43, 0.27).

Given that some of the variance components for the random structure are zero, the linear mixed models for the network small-worldness dynamics are compromised and the R2 cannot be computed. Therefore, the models shown in Table 7 should be considered with caution.

The fluctuations in the small-worldness coefficients in control and MCI participants can be seen in Figure 5(C) and they do not appear to indicate different patterns in the small-world network topology across groups.