Static and Dynamic Functional Brain Networks Reveal Distinct Connectivity Alterations Across the Alzheimer's Disease Spectrum

arXiv:2505.03458 · q-bio.NC, q-bio.QM · Submitted 2025-05-06 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Genomics Radio. Generated commentary on the latest computational biology and genomics papers.

Ines: We're looking at the paper: "Static and Dynamic Functional Brain Networks Reveal Distinct Connectivity Alterations Across the Alzheimer's Disease Spectrum".

Marcus: Here's the gist of the paper. Brain network analysis using functional MRI has advanced our understanding of cortical activity and its disruption in neurodegenerative disorders underlying dementia In this study, we analysed age- and sex-matched static and dynamic functional brain networks derived from resting-state fMRI data in 315 individuals with AD, mild cognitive impairment (MCI), and cognitively normal healthy controls (HC) from the ADNI-3 cohort While HC and MCI show similar static and dynamic patterns at the node level, clearer differences emerge in AD We identified stable (stationary) differences in functional connectivity between white matter regions and parietal and somatosensory cortices, whereas temporally varying differences were consistently observed in connections involving the amygdala and hippocampal formation Our results highlight shared and unique functional connectivity patterns in both static and dynamic functional networks, emphasising the importance of incorporating dynamic information into brain network analyses of the Alzheimer’s spectrum Furthermore, we trained a Random Forest model on regional BOLD time series informed by dynamic and static network metrics, achieving robust classification of MCI, AD, and HC groups and demonstrating the diagnostic potential of time-varying connectivity The functional networks were constructed using the Juelich brain atlas (JBA), with static connectivity estimated from full time series and dynamic connectivity derived using a sliding-window approach Cortical regions – defining the network nodes – are based on the Jülich brain atlas (JBA) Amunts et al.

Ines: 2020, resulting in cortical parcellations into 121 region The signal of each node is calculated by averaging the rs-fMRI BOLD time series across all voxels within the corresponding region/node Functional links were defined by the Pearson correlation coefficient, r Sporns 2010; Fornito et al.

Marcus: 2016; Pearson 1895, calculated pairwise between all nodes To construct links of static functional network, we utilise Eq.

Ines: (1) (i.e.,

t0, t0+∆t: =

0, NT: ), and calculate correlation coefficient applied to the full time series (197 points) of each cortical node Sporns 2010; Fornito et al.

Paper summary: Ines: So we've covered how this study used functional MRI to compare static and dynamic brain networks in AD patients versus healthy people. We saw distinct differences between white matter regions and parietal cortices statically, but temporal variations in the amygdala and hippocampus dynamically.

Marcus: The paper shows that these connectivity patterns are not just random noise; they have a specific structure across the spectrum of Alzheimer's disease, linking these changes to underlying pathology like tau and amyloid deposits.

Yuki: It points toward a more complex picture where global brain effects are heterogeneous, meaning different parts of the brain are affected in different ways depending on the stage of the disease.

Ines: The overall conclusion from this paper is that incorporating dynamic information into brain network analyses is really important when studying the Alzheimer’s spectrum because it reveals temporal characteristics of connections between structures like the amygdala and hippocampal formation, which are linked to both neuropathological hallmarks and cognitive decline.

Marcus: And they showed that differences in eigenvector centrality, which they call an early biomarker for how early AD pathology interacts with cognitive decline, were driven by those specific temporal functional networks identified at the link level.

Yuki: So what does this mean for us looking at population genetics? It suggests that the vulnerability isn't just one static feature, but a dynamic interplay of different brain areas over time.

Ines: Exactly. The paper provides evidence that these differences in eigenvector centrality were driven by the amygdala-hippocampus formation connectivity, which they identified as an indicator of abnormal brain function in early AD.

Conclusion: Ines: So we're wrapping up this discussion on the paper titled "Static and Dynamic Functional Brain Networks Reveal Distinct Connectivity Alterations Across the Alzheimer's Disease Spectrum" by Amunts et al.

Marcus: It’s about looking at both static brain connections and dynamic ones in people with Alzheimer's to see how different groups of brain regions are changing differently over time.

Yuki: It’s interesting because it separates the differences into two types: stable patterns that don't change much, and temporal variations that do shift with the disease progression.

Ines: Exactly. The authors found clear differences between white matter regions and parietal or somatosensory cortices in static networks, but then they saw different connections involving the amygdala and hippocampus changing over time.

Marcus: They used a random forest model trained on both static and dynamic metrics to classify people as healthy controls, mild cognitive impairment, or full Alzheimer's disease.

Yuki: The main implication here is that we need to look at the brain not just as a snapshot but how those connections evolve across the entire spectrum of Alzheimer's.

Ines: Because they found that temporal information in dynamic networks helps capture these subtle changes better than just looking at one single point in time.

Marcus: The paper suggests that the way these different parts of the brain are linked over time is a key feature we can use to diagnose or track disease progression more effectively.

Nicolas Rubido, Venia Batziou, Marwan Fuad, Vesna Vuksanovic

University of Aberdeen · Swansea University Medical School

q-bio.NC, q-bio.QM

Submitted: 2025-05-06

Updated: 2026-10-31

License: http://creativecommons.org/licenses/by/4.0/

Importance score: 83/100

The gist: Brain network analysis using functional MRI has advanced our understanding of cortical activity and its disruption in neurodegenerative disorders underlying dementia In this study…

Key concepts

Static Functional Networks (sFNs)
These represent the overall brain connectivity patterns measured from a full time series of brain activity. They show stable, long-term differences in how different brain regions are connected across the entire recording period, providing a snapshot of baseline functional organization.
Dynamic Functional Networks (dFNs)
These networks capture how functional connections change over time using sliding windows. They reveal temporal variations in brain connectivity, showing which connections fluctuate significantly across different segments of the measurement period.
Eigenvector Centrality (EC)
This metric measures a node's importance based on its connections to other nodes. A high EC score means a region is highly influential because it connects to many other important regions within the network, indicating its central role in brain function.

Terminology

Summary

Brain network analysis using functional MRI has advanced our understanding of cortical activity and its disruption in neurodegenerative disorders underlying dementia In this study, we analysed age- and sex-matched static and dynamic functional brain networks derived from resting-state fMRI data in 315 individuals with AD, mild cognitive impairment (MCI), and cognitively normal healthy controls (HC) from the ADNI-3 cohort While HC and MCI show similar static and dynamic patterns at the node level, clearer differences emerge in AD We identified stable (stationary) differences in functional connectivity between white matter regions and parietal and somatosensory cortices, whereas temporally varying differences were consistently observed in connections involving the amygdala and hippocampal formation Our results highlight shared and unique functional connectivity patterns in both static and dynamic functional networks, emphasising the importance of incorporating dynamic information into brain network analyses of the Alzheimer’s spectrum Furthermore, we trained a Random Forest model on regional BOLD time series informed by dynamic and static network metrics, achieving robust classification of MCI, AD, and HC groups and demonstrating the diagnostic potential of time-varying connectivity

Methodology for Network Construction

The functional networks were constructed using the Juelich brain atlas (JBA), with static connectivity estimated from full time series and dynamic connectivity derived using a sliding-window approach Cortical regions – defining the network nodes – are based on the Jülich brain atlas (JBA) Amunts et al. 2020, resulting in cortical parcellations into 121 region The signal of each node is calculated by averaging the rs-fMRI BOLD time series across all voxels within the corresponding region/node Functional links were defined by the Pearson correlation coefficient, r Sporns 2010; Fornito et al. p. 2

Where the paper says it

The signal of each node is calculated by averaging the rs-fMRI BOLD time series across all voxels within the corresponding region/node. Functional links were defined by the Pearson correlation coefficient, r Sporns (2010); Fornito et al.

2016; Pearson 1895, calculated pairwise between all nodes To construct links of static functional network, we utilise Eq. p. 2
Where the paper says it

Functional links were defined by the Pearson correlation coefficient, r Sporns (2010); Fornito et al. (2016); Pearson (1895), calculated pairwise between all nodes: r (s) t∈[t0, t0+∆t] (i, j) = cov xt∈[t0, t0+∆t](i) (s) , xt∈[t0, t0+∆t](j) (s) σi σj , (1) where cov xt∈[t0, t0+∆t](i) (s) , xt∈[t0, t0+∆t](j) (s) is the covariance between the time-windowed signals of nodes i and j over the interval [t0, t0 + ∆t], with t0 + ∆t ≤ NT = 197.

(1) (i.e., [t0, t0+∆t] = [0, NT]), and calculate correlation coefficient applied to the full time series (197 points) of each cortical node Sporns 2010; Fornito et al. p. 2
Where the paper says it

To construct links of static functional network, we utilise Eq. , [t0, t0+∆t] = [0, NT ]), and calculate correlation coefficient applied to the full time series (197 points) of each cortical node Sporns (2010); Fornito et al.

2016; Lei et al. p. 2
Where the paper says it

, [t0, t0+∆t] = [0, NT ]), and calculate correlation coefficient applied to the full time series (197 points) of each cortical node Sporns (2010); Fornito et al. (2016); Lei et al.

2019 Dynamic functional networks were estimated using a half-overlapping sliding window approach applied to the entire time series of each cortical node, as described in Leonardi and Van De Ville 2015 The windows were shifted in steps of ∆t/2, resulting in a number of half-overlapping windows across the entire time series

Node Importance Metrics

We characterised the importance of each node in the dFNs and sFN of each individual in the cohort by calculating the node strength and eigenvector centrality Node Strength, or weighted degree, quantifies the strength of functional connections of a node rather than simply their count Eigenvector Centrality is the centrality measure based on the assumption that connections to more influential nodes are more important than connections to less relevant nodes The elements of this eigenvector centrality provide a global measure of the relative importance of each node in the network, taking into account all connections in the network

Classification Performance

A random forest classifier was employed to distinguish between individuals with AD, MCI and HC based on their fMRI time series To train the classifier we used either the entire time series of 197 time points, or windows of 60, 120 and 180 seconds For classification, three subsets of 121 JBA regions were used as input features The highest accuracy (≈ 79%) was obtained for HC vs AD using the first set of nodes (amygdala and hippocampal formation) with a window length of ∆t = 180 s Overall, classification performance was higher when distinguishing between two groups rather than all three

Link Level Analysis Findings

The analysis of dFNs at the link level quantifies how often each connection shows significant differences across sliding windows The number of significant occurrences ranges from 4 to 12 (out of 18 windows; Fig. 3) Figure 4 shows the frequency of inter-cohort differences across sliding windows: how often each link differs significantly between groups The most consistent differences are driven by connections between WM and somatosensory regions (SPL and SSC) While AD vs HC dFN patterns broadly follow those observed in static networks, MCI vs HC differences are more pronounced in dFNs than in sFNs

Static Network Deviations

To assess directional changes in connectivity, we compared group-level median sFNs of AD and MCI against HC (Fig. 5) For example, connections between WM and somatosensory cortex (SSC) are approximately uncorrelated in HC (r ≃ 0), but become positively correlated in AD (r ≃ 0. p. 8

Where the paper says it

Although the overall structure appears similar, several region-specific deviations are evident. For example, connections between WM and somatosensory cortex (SSC) are approximately uncorrelated in HC (r ≃ 0), but become positively correlated in AD (r ≃ 0.

25) Dynamic FN deviations from healthy controls (Fig. p. 8
Where the paper says it

25) cohorts, but there were no differences between HC and AD. Dynamic Networks between Groups: Deviations from Healthy Controls Dynamic FN deviations from healthy controls (Fig.

6) show less consistent changes compared to static networks, reflecting the non-stationary nature of dFNs Some connections increase while others decrease across time windows, indicating that temporal variability plays an important role in functional connectivity

Discussion and Implications

Our comparative analysis of sFNs and dFNs at the node level, highlighted distinct drivers of local and global functional connectivity patterns between groups A consistent finding across analyses is that inter-cohort differences are largely driven by connectivity patterns involving white matter regions These effects are robust at the link level in both static and dynamic networks, particularly in connections between WM and somatosensory and parietal regions The most consistent results are the group differences that are driven by the connectivity patterns between the WM regions and somatosensory cortices (Figs. 1, 2,3 and 4) dFNs revealed temporal characteristics of the links between the amygdala and the hippocampal formation, associated with neuropathological as well as cognitive hallmarks of AD The variability in eigenvector centrality, which has been identified as an early biomarker of the interplay between early Alzheimer’s disease pathology and cognitive decline in healthy individuals, is driven by temporal FNs identified through the link level analysis We provide evidence that EC differences were driven by the amygdalahippocampus formation connectivity, which has been identified as an indicator of abnormal brain function in early AD

Author Contributions

N.R conducted analysis on the fMRI data, prepared figures, and drafted the initial methods and results sections V. p. 8

Where the paper says it

Discussion In this study, we analysed static and dynamic cortical functional networks derived from rs-fMRI data in a cohort of 315 individuals across three clinical groups: Alzheimer’s disease, mild cognitive impairment, and healthy controls. Our results show that combining static and dynamic analyses provides a more complete description of functional connectivity changes associated with the Alzheimer’s disease spectrum.

B performed the RF analysis, with M. p. 2
Where the paper says it

We also performed analyses on another standard FSL atlas – (Harvard - Oxford Brain Atlas [RRID:SCR_001476]) (H-OBA), which parcellates cortex into 48 regions. Our main results are reported for the JBA analysis, while the H-OBA results are reported in the SI.

F contributing the same RF analysis on the dataset V. p. 0
Where the paper says it

IMPROVED CLASSIFICATION OF ALZHEIMER’S DISEASE AND MILD COGNITIVE IMPAIRMENT THROUGH DYNAMIC FUNCTIONAL NETWORK ANALYSIS A PREPRINT Nicolas Rubido& Aberdeen Biomedical Imaging Centre University of Aberdeen Venia Batziou& Health Data Science Swansea University Medical School Swansea University Marwan Fuad Health Data Science Swansea University Medical School Swansea University Vesna Vuksanovic´ ∗ Health Data Science Swansea University Medical School Swansea University Correspondence:vesna. vuksanovic@swansea.

V supervised the project, conceptualised the study, and wrote the final manuscript

Funding Declaration

N.R and M.F were supported by RSAT (Grant No. DSR1058-100), V. p. 4

Where the paper says it

We then added the labels based on AD, MCI or HC group and we trained a RF classifier using stratified 10-fold cross-validation. The classifier averages the results of 100 decision trees.

B was funded by BRACE (Grant No.

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Enhanced classification accuracy by integrating dynamic functional connectivity metrics into machine learning models using a Random Forest classifier trained on regional BOLD time series informed by both static and dynamic network metrics. This allows the improved AI system to robustly classify individuals into MCI, AD, and HC groups with demonstrated diagnostic potential of time-varying connectivity.

  2. Detection of transient (metastable) functional states in brain networks, which are usually averaged out by conventional analysis. The improved AI system can identify these flexible links and their reconfiguration profiles to capture transient (metastable) states that support cognitive processing, offering a deeper understanding of the underlying pathophysiology in Alzheimer's disease.

  3. Biomarker development for AD progression by utilizing dynamic functional connectivity patterns as biomarkers independent of subjective cognitive scores. The improved system can identify consistent inter-cohort differences driven by the connectivity patterns between the WM regions and somatosensory cortices and appearance of (temporal) links between the Broca area and amygdala with the hippocampal formation.

  4. Characterization of local versus global network organization in neurodegeneration. The system can distinguish that WM functional connections are local in nature by comparing node strength and eigenvector centrality metrics between static and dynamic networks, providing a more nuanced understanding of how AD disrupts connectivity at different scales.

Abstract

Brain network analysis using functional MRI has advanced our understanding of cortical activity and its disruption in neurodegenerative disorders underlying dementia. Dynamic (time-varying) brain networks, which capture spatial and temporal patterns of regional cortical co-activity, remain relatively unexplored across the Alzheimer's disease (AD) spectrum. We analysed age- and sex-matched static and dynamic functional brain networks from resting-state fMRI in 315 individuals with AD, mild cognitive impairment (MCI), and cognitively normal healthy controls (HC) from ADNI-3. Networks were constructed using the Juelich brain atlas, with static connectivity estimated from full time series and dynamic connectivity via a sliding-window approach. Group differences were assessed at link and node levels using non-parametric statistics and bootstrap resampling. Time series from regions identified in these analyses were then used in Random Forest classification to test whether the alterations carried discriminative information at the individual level. HC and MCI showed largely similar node-level static and dynamic connectivity, whereas AD exhibited pronounced alterations in network organisation and temporal variability. Stable (stationary) differences emerged in connectivity between white matter regions and parietal and somatosensory cortices, whereas temporally varying differences consistently involved the amygdala and hippocampal formation. Node centrality analysis suggested that white matter connectivity differences are predominantly local. Our results highlight shared and unique connectivity patterns in static and dynamic networks, emphasising the importance of incorporating dynamic information into brain network analyses of the Alzheimer's spectrum.

Related papers