Dynamic multilayer functional connectivity detects preclinical and clinical Alzheimer's disease

Abstract

Increasing evidence suggests that patients with Alzheimer's disease present alterations in functional connectivity but previous results have not always been consistent. One of the reasons that may account for this inconsistency is the lack of consideration of temporal dynamics. To address this limitation, here we studied the dynamic modular organization on resting-state functional magnetic resonance imaging across different stages of Alzheimer's disease using a novel multilayer brain network approach. Participants from preclinical and clinical Alzheimer's disease stages were included. Temporal multilayer networks were used to assess time-varying modular organization. Logistic regression models were employed for disease stage discrimination, and partial least squares analyses examined associations between dynamic measures with cognition and pathology. Temporal multilayer functional measures distinguished all groups, particularly preclinical stages, overcoming the discriminatory power of risk factors such as age, sex, and APOE ϵ4 carriership. Dynamic multilayer functional measures exhibited strong associations with cognition as well as amyloid and tau pathology. Dynamic multilayer functional connectivity shows promise as a functional imaging biomarker for both early- and late-stage Alzheimer's disease diagnosis.

Keywords: AD; cognition; pathology; rs-fMRI; temporal brain networks.

Fig. 1

Visual representation of the methodology. From each group, we extracted the rs-fMRI time-series for 200 regions of the Schaefer atlas and then divided into 19 non-overlapping time windows of 30-s duration in order to obtain an adjacency matrix from each time window (A). Then we calculated the individual multiplex communities (B). Lastly, we evaluated the flexibility and calculated the module allegiance matrix, which represents the probability that two brain regions are part of the same community across the time windows (C). We also computed the dynamic recruitment coefficient and the dynamic integration coefficient of the different RSNs that summarize the results from the module allegiance (C).

Fig. 2

Visualization of our dynamic multilayer FC results for RSNs across different groups. Visualization of summary statistics for RSN-specific (A) integration, (B) recruitment, and (C) flexibility across the 19 time windows. The center black lines represent the median. Statistical analyses were performed while adjusting for sex and age and correcting for multiple comparisons using FDR, and significance levels are denoted as follows: ^*P < 0.05, ^**P < 0.01, ^***P < 0.001. CON—control network; DMN—default mode network; DAN—dorsal attention network; SVAN—salience ventral attention network; LIMB—limbic network; SM—somatomotor; VIS—visual networks.

Fig. 3

Relationship between each pair of dynamic multilayer measures in the AD continuum. Relationship between pairs of dynamic multilayer FC measures in terms of the RSN-specific median values: (A) recruitment-integration, (B) recruitment-flexibility, and (C) integration-flexibility. CON—control network; DMN—default mode network; DAN—dorsal attention network; SVAN—salience ventral attention network; LIMB—limbic network; SM—somatomotor; VIS—visual networks.

Fig. 4

Group classification using dynamic multilayer FC measures and risk factors. ROC curves displaying the AUC scores (left) and confusion matrices (right) for each of the three classification models: CN Aβ + and CN Aβ- (A), MCI Aβ + and CN Aβ− (B), AD Aβ + and CN Aβ− (C) groups. Model F + R is the best model obtained by combining dynamic functional measures and risk factors, followed by model F that contains only the best combination of dynamic functional measures and finally model R that includes only the risk factors.

Fig. 5

PLS regression analysis to assess the relationship of our dynamic multilayer FC measures with cognition and AD pathology. VIP scores determine significant predictors in the PLS model, which are highlighted in darker blue (VIP > 1), for global cognition (A), memory (B), tau-PET (C), and amyloid-PET (D). In light blue we show the non-significant predictors (VIP <= 1). F—flexibility; R—recruitment; I—integration; CogStatus—cognitive status; Edu—education. All PLS models were performed with all the Aβ positive subjects.