Multilayer brain connectivity and long-term cognitive changes in individuals at risk for Alzheimer's disease

Participants

The INSIGHT-preAD cohort (INveStiGation of Alzheimer's PredicTors in subjective memory) is a longitudinal, multidisciplinary study conducted at the Pitié-Salpêtrière Hospital in Paris (France) between 2013 and 2020. 318 native French-speaking participants or participants who had learned French during childhood, aged 70–85 years with subjective memory complaints (SMC) for at least six months at the screening visit, ²⁶ were included and followed-up for five years.^27,28 Among the inclusion criteria, participants were required to have unimpaired cognition (Mini-Mental State Examination [MMSE] score ≥27/30; Clinical Dementia Rating [CDR] score = 0; and Free and Cue Selective Reminding Test Total Recall [FCSRT_TR] score ≥44/48). Participants were excluded from the study if they were under guardianship or in a retirement home, had a pre-symptomatic AD or other neurodegenerative disorder diagnosis, had a stroke on the three months prior to the beginning of the study, were illiterate, opposed or unable to undergo the MRI scanning.

Informed consent was obtained after the nature of the study was explained. Amyloid positivity was determined using ¹⁸F-florbetapir PET as a global cortical standardized uptake value ratio (SUVR) > 0.79 calculated as the mean SUVR across predefined cortical regions of interest following the INSIGHT-preAD protocol. ²⁹ AD diagnosis was determined by an independent and specialized committee in cases of cognitive decline across two consecutive neuropsychological assessments and persistent Asymmetrical Hippocampal Segmentation (AHS). ²⁹ In line with previous INSIGHT-preAD investigations using this cohort, ³⁰ a subsample of the cohort that completed all initial neuropsychological tests was selected for this study. This sample included 34 participants without any

Amyloid-β burden (Aβ-/controls); 19 amyloid-β positive non-progressors to prodromal AD (Aβ+/stable) and 13 amyloid-β positive progressors to prodromal AD (Aβ+/progressors). Just like the participants in the Aβ-/controls group, the participants in the Aβ+/stable and Aβ+/progressors groups were cognitively normal at baseline, consistent with the INSIGHT-preAD inclusion criteria.

Amyloid PET acquisition and positivity threshold

Amyloid PET scans were acquired on a 3D PET/CT scanner 50 min after injection of 370 MBq (10 mCi) ¹⁸F-florbetapir. The images were processed using the in-house RACHEL (Région d’Analyse pour la Characterisation En Ligne; Region of Analysis for Online Characterization) software pipeline developed by the Centre d’Acquisition et Traitement des Images (CATI; http://cati-neuroimaging.com) where the SUVRs were calculated by averaging the mean activity of cortical regions of interest most affected by amyloid pathology: left and right precuneus, cingulum posterior, cingulum anterior, and parietal, temporal and orbitofrontal cortices. The SUVR threshold of 0.7918 was extracted by linear correlation of results from CAEN (IMAP study) and AVID's SUVR values, ³¹ and validated through its use in the first study by Dubois et al., 2018. ²⁸

Behavioral measures: neuropsychological assessment

The participants’ cognitive abilities were assessed using the following neuropsychological tests: the MMSE to assess global cognitive function; the FCSRT, ³² and the Rey-Osterrieth Complex Figure to assess episodic memory; the Delayed Matching to Sample 48 (DMS48) to assess visual recognition memory; the Digit Span and Visuospatial Span to assess working memory; the Frontal Assessment Battery (FAB) and the Trail Making Test (TMT) to assess executive functions (for a detailed description of the behavioral measures). ²⁶ Clinical follow-ups were carried out throughout the study: every six months for psycho-behavioral assessments, every twelve months for neuropsychological assessments and EEG recordings, and every two years for fMRI and PET imaging (Figure 1A).

OPEN IN VIEWER

Figure 1.

Schematic overview of the default mode network (DMN), experimental design and multilayer network construction combining EEG and DWI data. A) Design of the experimental protocol detailing the sequence of behavioral sessions (all used for longitudinal analyses), and neuroimaging assessments (M0 EEG data only used, red rectangle); B) Illustration of the DMN visualized using R-Studio scripts according to the Destrieux Atlas; C) Construction of the multilayer network, combining two layers: functional connectivity (FC) derived from EEG data using Phase Lag Index (PLI), and structural connectivity (SC) derived from DWI data using Fractional Anisotropy (FA). EEG: electroencephalography; DWI: diffusion weighted imaging.

MRI and DWI data acquisition

T1-weighted MRI data for each participant were acquired at baseline (M0) on MRI scan 3 T Magnetom VERIO system (Siemens Medical Solutions Erlangen, Germany) over a 1 h period. A 3D MPRAGE sequence was acquired for the anatomical study (sagittal orientation; repetition time 2300 ms; echo time 2.98 ms; inversion time 900 ms; flip angle 9°; 176 slices; slice thickness 1 mm; field of view 256*240 mm; matrix 256*240; bandwidth 240 Hz/Px. ²⁸ DWI data were obtained with the following parameters: repetition time (TR) = 14000 ms; echo time (TE) = 93 ms; field of view (FOV) = 2304 mm×2304 mm; 70 axial slices; voxel size = 2 mm isotropic. The diffusion weighting scheme comprised one non-diffusion-weighted volume (B0 = 0 s/mm²) and 15 diffusion-weighted volumes acquired at B1 = 1500 s/mm².

Neuroimaging DWI data pre-processing

Pre-processing of the diffusion data was performed using ExploreDTI (version 4.8.6), ³³ and included the following steps: (a) images were corrected for eddy current distortions and participant motion; (b) a non-linear least squares method was applied for diffusion tensor estimation, and (c) deterministic DWI tractography was applied using the following parameters: uniform resolution of 2 mm, fractional anisotropy (FA) threshold of 0.2 (limit: 1), angle threshold of 45°, and fiber length range of 50–500 mm. Diffusion tensor imaging (DTI) data were motion corrected during pre-processing. A global motion index was used for quality control and the mean motion across the final sample remained below 0.4 mm. The network analysis tools in ExploreDTI were used to quantify the FA value of the fibers connecting the regions of the Destrieux atlas to obtain matrices comparable to the EEG data, ³⁴ using FreeSurfer's individual cortical parcellation (FreeSurfer/5.3.0). This approach yielded 148*148 SC matrices covering major DMN regions (Figure 1B) such as the precuneus, posterior cingulate gyrus and sulcus, the medial prefrontal cortex, angular gyrus, temporal pole and middle temporal gyrus. A Kruskal-Wallis analysis to control DTI data quality did not reveal any difference among groups (Kruskal-Wallis, p > 0.05).

Neuroimaging EEG data acquisition

High density resting-state EEG (rsEEG) data were acquired at M0 with a 256-channel whole-head cap at 250 Hz and amplified by the EGI NetAmp300. Two separate rsEEG runs, lasting two minutes each with four 30 s sessions alternating between eyes closed (EC) and eyes opened (EO), were acquired at approximately 25 min intervals. An audio cue (low or high frequency tone) indicated the change between EC and EO.

Neuroimaging EEG data pre-processing

Pre-processing was performed using Brainstorm (version 3.221117 Nov 2022). ³⁵ First, the electrodes were re-referenced to consider the 257^th electrode (i.e., Cz) as the reference. The data from the other 256 electrodes were then filtered (0.5–40 Hz bandpass; harmonic filters at 50 Hz, 100 Hz, 150 Hz, 200 Hz, 250 Hz, 300 Hz). Only the electrodes that covered the scalp were used in the analyses (electrodes outside the scalp were removed), thus only 169 electrodes were considered. Bad channels were detected and rejected using power spectral density (PSD).^11,35,36 Because all subsequent analyses were performed in source space, bad channels were excluded and not reconstructed. ³⁷ Physiological artefacts (blinks, saccades and cardiac artefacts) were identified and removed by independent component analysis (ICA) as implemented in Brainstorm and muscle artefacts were identified and discarded after visual inspection of ICA components. ³⁸ Four epochs of 30 s each, alternating between eyes closed and eyes opened (EC, EO), were extracted. EO epochs were used to normalize the activity measured during the EC (e.g., (EC-EO)/EO for each subject and in each band, at each grid point, using the EO condition as a subject specific physiological baseline). ³⁹ This ensures that the reported connectivity and power values reflect state-specific neural dynamics. ⁴⁰ This normalization was applied at the level of PSD, prior to connectivity computation. The EEG head model was obtained using a symmetric boundary element method. ⁴¹ The Phase Lag Index (PLI) was computed in source space, after source reconstruction with a Boundary Element Model (BEM) and cortically constrained sLORETA inverse solution procedure to estimate the cortical origin of scalp EEG signals, ⁴² using individual T1-weighted anatomical MRI priorly processed using FreeSurfer. ⁴³ For each region of interest (ROI), a representative time series was extracted using principal component analysis of all source vertices within that ROI, as implemented in Brainstorm. This approach captures the dominant temporal dynamics of each cortical region and reduces the influence of local noise and spatial leakage.

The individual alpha-peak frequency (IAF) measured as the average of peaks detected between 7–13 Hz over occipital electrodes was used to define the delta band for each participant (between IAF minus 8 Hz and IAF minus 6 Hz).^3,13,44 The PLI was computed as implemented in Brainstorm, ⁴⁵ and was used to assess the functional synchrony between ROI defined using the Destrieux atlas brain parcellation.

Multiplex network segmentation and construction

Following Jauny et al., ²⁰ we used BRAPH software 1.0.0 (http://braph.org/)46 to construct a two-layer multiplex network for each subject: a “structural” layer with DWI tract FA data and a “functional” layer with PLI EEG data. This framework allows direct assessment of cross-modal integration while preserving the topological correspondence between anatomical and functional networks.^46,47 A multiplex network is a multilayer framework in which several types of connectivity (i.e., structural and functional) are represented as distinct but interconnected layers. Each layer encodes a specific type of interaction between the same set of nodes. In this case, the cortical regions defined by the Destrieux atlas share identical nodes across layers (i.e., the same ROIs across layers) (Figure 1C). Therefore, each brain region (e.g., the precuneus, the medial prefrontal cortex and the posterior cingulate cortex) appears once in each layer, and the edges connecting these nodes represent modality-specific interactions. Considering the prominent changes in healthy and pathological aging,^3,8 the association of connectivity in the delta frequency band with the structural layer was specifically investigated. For each participant, a binary multiplex matrix was calculated from the individual matrices of DWI and EEG data. To ensure comparable network density across modalities and participants, both structural and functional connectivity matrices were binarized using a subject independent threshold. Matrices were binarized according to the minimum density of 1.9% observed for DTI across all participants. Specifically, the minimum non zero (1.9%) tract density value observed across all participants in the DWI matrices was used as common threshold. The density or number of connections for all matrices was the same across layers. Connections with values above this threshold were assigned a weight of 1, and all others were set to zero. This procedure ensures that all multiplex networks share the same edge density, facilitating across layer comparisons. We chose the minimum non-zero tract density as the binarization threshold as it was the number observed for DTI data across all participant, ⁴⁸ in order to preserve all structural connections detected by deterministic tractography in any participant, on the basis that a connection is considered present whenever at least one streamline is reconstructed between two regions. Fixed percentage thresholds such as retaining the top 10% of connections can introduce participant dependent variability by including weak connections in some individuals while excluding potentially genuine ones in others, thereby affecting topology in a non-uniform way. This approach is also consistent with previous work from our group. ²⁰ Since the nodes are identical across layers, it is possible to quantify how similar the connectivity pattern of each region is between modalities, for example by computing the multiplex participation coefficient (PC), ⁴⁷ which reflects the degree of structure-function similarity for a given node. The PC was investigated, allowing the quantification of the connectivity similarity of a node across the different layers. The multiplex PC of a node i is defined as:

p i = M M − 1 [ 1 − ∑ ∝ = 1 M ( k i [ ∝ ] o i ) 2 ]

Statistics

To assess EEG data quality, the number of rejected ICA components and bad channels were recorded for each participant. Group comparisons on these quality metrics were performed using the Kruskal-Wallis to verify that data quality did not differ systematically across groups. Descriptive statistics were calculated for each neuropsychological measure and are reported as median values with interquartile range (IQR). These descriptive values are presented in Table 1. The normality and homogeneity of variance of each dependent variable were verified using respectively Q-Q plot visualization and Levene's test. The individual mean of subjective memory complaints was first computed for each participant across all available visits. Group level descriptive statistics (mean ± standard deviation) were then computed from these individual means separately for the three amyloid groups. Between-group comparisons were performed using ANOVA despite unequal group sizes (n = 34, n = 19 and n = 13 respectively). Consistent with previous reports from the INSIGHT pre-AD cohort all participants were cognitively normal at inclusion and showed comparable neuropsychological performance at baseline (M0) independently of amyloid burden.^28,30,50 The analyses therefore focused on longitudinal trajectories rather than baseline group difference. To assess the five-year longitudinal cognitive trajectory for each participant's neuropsychological test, variation rates were calculated as a percentage change, defined as the difference between the score at the participant's last follow-up session (e.g., M48 or M60) and their baseline (i.e., M0). This difference was then divided by the score at M0 and multiplied by 100. (e.g., (MMSE M60–MMSE M0)/MMSE M0) * 100). The variation rates normalize change relative to the initial score, enabling a direct comparison across participants with different baseline levels. However, because this approach captures the total magnitude of change between two time points rather than the rate of change per unit time, they do not account for differences in the duration of follow-up across participants.

OPEN IN VIEWER

Table 1.

Neuropsychological performance at the initial session and after 5 years of follow-up (last session) for the three groups’ cognitive scores showing significant results.

		Initial session			Last session
		Aβ-/Controls	Aβ+/Stable	Aβ+/Progressors	Aβ-/Controls	Aβ+/Stable	Aβ+/Progressors
Score names		Median [Q1-Q3]	Median [Q1-Q3]	Median [Q1-Q3]	Median [Q1-Q3]	Median [Q1-Q3]	Median [Q1-Q3]	Group differencesUncorrected p	n2
MMSE		29 [28–29]	28 [28–29]	28 [28–29]	30 [29–30]	29 [28–29]	27 [25–28]	0.01	0.139
FCSRT	TFR	32 [29–34]	31 [27–33]	23 [20–28]	33 [29–36]	33 [29–36]	13 [11–15]	<0.001a,b	0.461
	TR	47 [45–48]	47 [46–48]	45 [44–46]	47 [46–48]	47 [45–48]	36 [29–39]	<0.001a,b	0.599
	DFR	12 [10–13]	12 [10–13]	10 [8–12]	13 [11–14]	12 [11–13]	1 [0.3–5]	<0.001a,b	0.395
	DTR	16 [16–16]	16 [15–16]	16 [15–16]	16 [16–16]	16 [16–16]	12 [10–13]	<0.001a,b	0.617
DMS48	IR	47 [45–47]	48 [46–48]	47 [46–48]	48 [47–48]	48 [46–48]	46 [44–46]	<0.001a,b	0.229
	DR	46 [44–47]	46 [45–47]	46 [45–48]	46 [45–47]	46 [45–48]	46 [45–46]	ns	0.042
Rey-Osterrieth	3min	15 [12–22]	19 [16–22]	11 [11–18]	21 [17–27]	18 [15–23]	5 [5–6]	<0.001a,b	0.205
	30min	16 [13–23]	19 [15–23]	16 [10–18]	21 [16–25]	19 [16–24]	7 [5–12]	0.006a,b	0.158
Visuospatial span	Forwards	5 [4–6]	6 [5–6]	5 [4–6]	5 [5–6]	5 [5–6]	5 [4–5]	0.02a,b	0.120
	Backwards	5 [4–6]	5 [4–6]	4 [3–5]	5 [4–6]	5 [4–6]	4 [4–6]	ns	0.040

MMSE: Mini-Mental State Examination (range: 0–30 points); FCSRT: Free and Cued Selective Reminding Test; DMS48: Delayed Matching to Sample Task 48; Rey-Osterrieth Complex Figure (score based on the accuracy of reproducing the different 18 components of the figure after 3 min and 30 min, range: 0–36). Only neuropsychological tests with significant results are listed. TFR, total free recall (number of correct items, range: 0–48); TR, total recall (number of correct items, range: 0–48); DFR, delayed free recall (number of correct items, range: 0–16); DTR, delayed total recall (number of correct items, range: 0–16); IR, immediate recall (total correct image selections, range: 0–48); DR, delayed recall (total correct image selections, range: 0–48); Forwards, Backwards (maximum sequence length correctly recalled, range: 2–7). The number represents the median and interquartile (Q1 for the 25% interquartile and Q3 for the 75% interquartile) of each group at the first (initial session) and at the last follow-up and the longitudinal evaluation of the cognition among the groups. p-value significance represents a difference between the three groups using ANOVA; the differences were further revealed with emmeans post-hoc corrected for multiple comparison using Tukey adjustment within each cognitive measure:

^a

[Aβ+/progressors < Aβ-/controls]: MMSE (p = 0.008), TFR (p ≤ 0.001), TR (p ≤ 0.001), DFR (p ≤ 0.001), DTR (p ≤ 0.001), DMS48_IR (p ≤ 0.001), ROCF_3 min (p ≤ 0.001), ROCF_30 min (p = 0.005), Visuospatial Span_Fowards (p = 0.013).

^b

[Aβ+/progressors < Aβ+/stable]: TFR (p ≤ 0.001), TR (p ≤ 0.001), DFR (p ≤ 0.001), DTR (p ≤ 0.001), DMS48_IR (p = 0.001), ROCF_3 min (p = 0.048), ROCF_30 min (p = 0.023). There is no significant difference between the stable groups (Aβ-/CTRL vs. Aβ+/stable). ns for non-significant results. No significant difference for Aβ-/controls < Aβ+/stable. n² for size effect.

Group-based analyses were conducted at two levels: (i) a whole-brain level using the average PC across all cortical regions from the Destrieux atlas, and (ii) a network-specific level restricted to regions of the DMN, defined a priori based on anatomical labels and prior literature. ²⁰ To investigate the specificity of reported results, additional control analyses were conducted in the alpha band (IAF minus 2/IAF plus 2) and in the Salience network (SN) (composed of the insula, the anterior cingulate cortex, the ventral striatum, the thalamus, the amygdala).

To assess group differences in the PC, one-way analyses of variance (ANOVA) were conducted, followed by pairwise post hoc comparisons using Tukey's test which is more appropriate than Bonferroni because of the unequal group size to explore specific group effects. Age, level of education and sex were included as covariates in these analyses to control for potential confounding effects. Association between the PC (global and local) and alpha peak frequency were conducted through Pearson's correlation. The relationship between the PC and cognitive variation rates was assessed using partial Pearson correlations, controlling for age, sex and level of education. Partial correlations were computed to isolate the unique association between the PC and each cognitive outcome while removing the shared variance attributable to these demographic covariates. These analyses were conducted both across all participants and separately within each group. An FDR correction was applied to the p-values derived from the partial correlation analyses to control for false positives. Educational level was characterized using the estimated number of years of formal education, with 12 years corresponding to a high school diploma. All statistical tests were conducted using JASP software (https://jasp-stats.org/ ; version 0.17.2.1). The networks specific analyses were FDR corrected for multiple comparisons. ⁵¹ To assess the internal consistency of the results across EEG recordings, a 2 (Group)×2 (Run) mixed ANOVA was performed. In addition to correlation analyses, additional exploratory hierarchical clustering analyses were performed to determine whether group differences in longitudinal cognitive trajectories could be identified based on the structure/function coupling profile, independently of their a priori amyloid-based group assignment. Hierarchical cluster analyses were performed using the between-groups linkage method of clustering (with squared Euclidean distance). The squared Euclidean distances were computed between each subject and subjects with the smallest distance were classified together in a cluster. The range of number of clusters started at 2 and stopped at 4, but only clusters that would divide the group in subgroups with sufficient number of subjects per subgroup (N > 15) were selected for statistical analyses.

Demographic and clinical characterization of the groups of interest

Participants were assigned to three groups according to their amyloid status and clinical progression over time: Aβ-/controls, Aβ+/stable and Aβ+/progressors.^30,50 Demographic characteristics (i.e., age, sex, level of education) and the assessment of the subjective memory complaint of the sample are presented in Tables 2 and 3, respectively. Among the 66 participants initially included in the study, 21.2% (n = 14) did not complete the full five-years follow-up (M60). The mean follow-up duration was 4.70 years (± 0.72). Consistent with previous INSIGHT pre-AD studies,^28,30,50 the groups did not differ in global cognitive performance at baseline (M0) as measured by the MMSE (p = 0.46), confirming that cognitive functioning was comparable among groups at study entry, as per protocol design.

OPEN IN VIEWER

Table 2.

Demographics characteristics of the groups of interest: Aβ-/controls, Aβ+/stable, Aβ+/progressors.

		Initial session
		Controls	Stable	Progressors	p
		Aβ- (n = 34)	Aβ+ (n = 19)	Aβ+ (n = 13)
		mean ± SD	mean ± SD	mean ± SD
Age		76.4 ± 2.9	76.7 ± 2.6	78.6 ± 3.0	0.06
Sex (%)	Women	21 (31.8%)	10 (15.2%)	8 (12.1%)	0.8
	Men	13 (19.7%)	9 (13.6%)	5 (7.6%)
Education level	Years of education	12.5 ± 5.4	13.47 ± 4.6	13.69 ± 4.9	0.7

Values represent the mean and standard deviation (SD) or count (%) of each group at the initial session. Educational level was characterized using an estimated number of years of formal education, derived from the categorical socio-cultural level, with 12 years corresponding to a high school diploma and 17 years corresponding to a Master's degree.

OPEN IN VIEWER

Table 3.

Subjective feelings about memory and cognition.

Assessments of the subjective memory complaints
				F(2,63)	p
INSIGHT-preAD questionnaire of cognitive decline	13.5 ± 5.8	13.0 ± 4.9	13.0 ± 4.6	0.069	0.934
Healthy Aging Brain Care Monitor	12.7 ± 9.1	10.8 ± 6.4	9.1 ± 6.7	1.078	0.347
Self-evaluation of complaints	24.6 ± 11.4	23.6 ± 9.8	19.5 ± 12.7	0.997	0.375

Values represent the mean and ± standard deviation of individual scores averaged across all available study visits (M0 to M60) for each participant. Between-group differences were assessed using one-way ANOVA; F-statistics are reported with degrees of freedom df(2,63). A Bonferroni correction was applied to account for multiple comparisons (3 tests; corrected α = 0.017). All p-Bonferroni = 1.00.

A clear divergence in cognitive trajectories emerged over time. After the five years follow-up, Aβ+/progressors showed significantly reduced performance in global cognition (MMSE) and episodic memory (FCSRT, DMS48 and Rey-Osterrieth) compared to both Aβ-/controls and Aβ+/stable as shown in Table 1. In contrast, Aβ+/stable and the Aβ-/controls groups remained globally stable across the follow-up period. This distinct longitudinal pattern highlights the emergence of cognitive decline specifically in the Aβ+/progressors group, despite comparable baseline performance.

Longitudinally, for global cognition, Aβ+/progressors have reduced MMSE scores compared to the Aβ-/controls (p_tukey = 0.008, ±1.688) but non statistically different compare to Aβ+/stable (p_tukey = 0.062, ±1.854) groups while Aβ-/controls did not differ from Aβ+/stable groups (p_tukey = 1, ±1.440) (see Table 3 for main group effects). A similar pattern was observed for memory performance, with lower FCSRT_TR total recall in Aβ+/progressors compared with Aβ-/controls and Aβ+/stable groups respectively (p_tukey<0.001, ±2.080) and (p_tukey<0.001, ±2.284), but no difference on performance between Aβ-/controls and Aβ+/stable (p_tukey = 0.755, ±1.774).

Reduced performance was also observed in Aβ+/progressors for DMS48 immediate recognition compared with Aβ-/controls (p_tukey<0.001, ±1.544) and Aβ+/stable (p_tukey = 0.001, ±1.713). For the Rey-Osterrieth 3-min recall, Aβ+/progressors showed a slower performance compared to Aβ-/controls (p_tukey<0.001, ±13.861) but a slightly slower performance compared Aβ+/stable (p_tukey = 0.048, ±15.169) with no difference between Aβ-/controls and Aβ+/stable (p_tukey = 0.303, ±12.206). Finally, for the Rey-Osterrieth 30minutes recall, the Aβ+/progressors showed a stronger decline compared to the Aβ-/controls (p_tukey = 0.005, ±15.129) and Aβ+/stable (p_tukey = 0.023, ±16.481), while the Aβ-/controls groups and the Aβ+/stable did not differ (p_tukey = 0.9336 ± 13.341), see Supplemental Table 1 for more information on size effect and bootstrapping 95%CI.

EEG functional connectivity PLI analyses

EEG data quality was comparable across the three groups. The number of rejected ICA components was minimal across participants ranging from 0 to 7 (median = 1) and did not differ significantly (Kruskal-Wallis: H(2) = 5.52, p = 0.063) between Aβ-/controls, Aβ+/stable and Aβ+/progressors groups. Exploratory pairwise comparisons with Bonferonni correction (Mann-Whitney tests, α = 0.017) did not identify any significant differences between specific groups pairs (all corrected p > 0.017). Similarly, the number of bad channels did not differ across groups (Kruskal-Wallis, p > 0.05), confirming that group differences in EEG connectivity measures were not attributable to systematic differences in data quality. The one-way ANOVA performed on delta band PLI values within the DMN revealed no significant group effect (F(2,63) = 1.65, p = 0.20, η² = 0.049). While mean PLI values were lower in Aβ+/progressors compared to the others two groups, these differences were not statistically significant. However, when examining the association between delta band within the DMN and FCSRT (Total Free Recall), a significant (after FDR correction) and positive partial correlation was observed (r = 0.262, p = 0.033). No significant results were found for the SN or in the alpha frequency (p > 0.1). Overall, participants with higher delta synchrony within the DMN showed better recall performance, suggesting that delta within the DMN supports episodic memory efficiency.

Multiplex participation coefficient analyses

A global multiplex ANOVA analysis to assess the PC in the entire brain was conducted for the delta band. No group differences for the delta frequency band (F(2,63) = 0.218, p = 0.805, η² = 0.007) or the alpha band frequency (F(2,63) = 0.953, p = 0.391, η² = 0.029). Then, a local network multiplex analysis was conducted to assess whether a difference in the PC was notable in the DMN for the delta band among our participants (F(2,63) = 4.363, p = 0.017, η² = 0.122). Post Hoc results of Tukey showed an increased PC for Aβ+/progressors compared to Aβ-/controls (p_tukey = 0.018, ±0.025) and Aβ+/progressors compared to Aβ+/stable (p_tukey = 0.035, ±0.028; see Figure 2). No significant differences were found between Aβ+/controls and Aβ+/stable groups (p_tukey = 1, ±0.022).

OPEN IN VIEWER

Figure 2.

Level of the multiplex participation coefficient in the DMN as a function of group. Higher and statistically significant PC in Aβ+/progressors participants compared to Aβ+/stable and Aβ-/controls. F(2,63) = 4.363, p = 0.017, η² = 0.122); significant p after post-hoc of Tukey, ns: non-significant.

No correlation was observed between global or local multiplex PC and the individual alpha peak frequency (ps > 0.40). To assess frequency and network specificity, the same multilayer connectivity analyses (e.g., additional ANOVA analyses) were conducted in alpha frequency band and the SN reported in AD.^8,52 No significant findings were observed. Furthermore, a 2(Group)×2(Run) mixed-design ANOVA including both rsEEG recordings (first and second rsEEG sessions; see Methods) was conducted to examine the robustness and internal replicability of the reported results. ⁵³ The results revealed no main effect of Run and no Group×Run interaction (ps > 0.16) indicating that the group effects remained stable across both recordings with no influence of the timing of the EEG acquisition, meaning the PC values remains stable across sessions.

Negative association between cognitive performance and multiplex participation coefficient

To further determine the impact of network changes on cognitive performance, partial correlation analyses were conducted between the PC and the scores of the neuropsychological measures listed in Table 1. p-values were adjusted for multiple comparisons using FDR across the eleven cognitive outcomes tested.

The correlation analyses revealed a significant negative association between global brain multiplex PC and FCSRT total recall variation rate within the delta band frequency with higher network similarity associated with larger episodic memory decline (r = -0.371, p_{_uncorrected} = 0.0082, p_{_FDR} = 0.016; Figure 3A). A similar negative association was observed when restricting the PC correlation to the DMN regions (r = -0.358, p_{_uncorrected} = 0.003, p_{_FDR} = 0.033; Figure 3B). Notably, among individual DMN areas, only the left dorsal PCC showed a significant correlation with the FCSRT_TR. The correlation also highlighted that this relationship was mainly driven by the stronger decline in Aβ+/progressors participants relative to the other groups. No other significant results were observed for other neuropsychological measures after FDR correction.

OPEN IN VIEWER

Figure 3.

Negative association of the multiplex participation coefficient with the FCSRT variation rate. A) A higher PC for the DMN within the delta band frequency is associated with lower cognitive function mostly for the Aβ+/progressors as shown by the FCSRT test, B) a higher PC in the whole-brain within the delta band frequency associated with lower cognitive function as revealed by the FCSRT.

Individual within-group correlations (performed separately for each group) did not yield significant associations, indicating that the global brain effect was driven by between group variability rather than within group variance. Additionally, to examine whether the PC tracks amyloid burden along AD continuum, partial correlations were computed between global cortical SUVR and the delta-band DMN PC, controlling for age, sex and education. No significant positive association was observed across the full sample nor within the amyloid positive participant suggesting that the relationship is primarily categorical rather than linear across the continuum of amyloid burden (see Supplemental Figures 1 and 2).

Hierarchical clustering analyses

Finally, as an additional data driven analysis, hierarchical clustering analyses based on PC within the DMN for the delta identified two distinct subgroups: 47 individuals with high PC and 17 individuals with low PC (p < 0.001). Importantly, the high PC subgroup included the majority of Aβ+/progressors (12 out of 13) whereas the low PC subgroup consisted predominantly of Aβ-/controls individuals (12 out of 17). Notably, individuals in high PC group exhibited a significantly greater cognitive decline in FCSRT scores compared to those in low PC (p = 0.003). This analysis does not constitute an independent validation of the group classification but it provides converging evidence for interindividual variability in delta band structure-function coupling, suggesting that PC meaningfully distributes individuals along a continuum of network rigidity that aligns with both amyloid status and cognitive trajectory.

Overall, the results showed that a high similarity between the structure-function coupling in the delta band frequency whether assessed in the DMN or at the whole-brain level is associated with greater cognitive decline, especially those who subsequently progress to AD. These findings support the relevance of multiplex coupling as a potential marker of progression along AD continuum within amyloid positive group based solely on connectivity features (Figure 3).