Blind Identification of Altered Functional Subnetworks in Alzheimer’s Disease Using Resting-State fMRI

Abstract

Introduction:

Resting-state functional magnetic resonance imaging (rs-fMRI) is widely used to examine functional connectivity (FC) alterations in neurological disorders such as Alzheimer’s disease (AD). Traditional studies either employ whole-brain analyses or focus on specific regions, yet the vast number of FCs and their interrelations complicate interpretation. This study adopts a data-driven, hypothesis-free approach to detect altered functional subnetworks in AD.

Methods:

Independent component analysis (ICA) was applied to FC matrices from 34 AD patients and 49 healthy controls (HCs) from the Alzheimer’s Disease Neuroimaging Initiative (ADNI). After pruning, significant subnetworks distinguishing AD from HC were identified. Graph theoretical parameters were computed for each subnetwork, and their associations with Mini-Mental State Examination (MMSE) scores were assessed.

Results:

Three subnetworks effectively differentiated AD patients from HCs. One subnetwork showed significant group differences in network strength, clustering coefficient, and local efficiency, despite no whole-brain differences. Abnormal functional lateralization also emerged within subnetworks. Moreover, FC weights in the identified subnetworks positively correlated with MMSE scores, linking cognitive performance to subnetwork connectivity.

Conclusion:

These results demonstrate the utility of a data-driven approach in detecting AD-specific altered subnetworks. By providing a modular perspective, this method facilitates targeted examination of connectivity changes, improves interpretability, and deepens understanding of functional disruptions in AD.

Keywords

blind analysis brain subnetwork functional connectivity Alzheimer rs-fMRI

Introduction

Alzheimer’s disease (AD) is a gradually worsening neurodegenerative disorder that represent the leading cause of dementia, responsible for approximately 60% to 70% of cases worldwide (WHO report [https://www.who.int/news-room/fact-sheets/detail/dementia]). It is marked by cognitive decline, memory impairment and disruptions in daily functioning, stemming from widespread neuronal degeneration.^1
-3

In addition to its clinical manifestations and structural alterations (eg, cortical atrophy), AD is increasingly recognized as a disconnection syndrome, characterized by impaired functional communication among brain regions.^4
-8 Resting-state functional Magnetic Resonance Imaging (rs-fMRI) has become a widely adopted approach for investigating these intrinsic connectivity disruptions.^9
-11

Prior studies investigating alterations in functional connectivity (FC) in neurological conditions such as AD have typically followed 1 of 2 main strategies: hypothesis-driven analysis; targeting specific regions or circuits, and whole brain approaches, which assess connectivity patterns across the entire brain without prior assumptions.^8,12
-17 The success of the latter heavily depends on the effectiveness of the search algorithms, while the former is inherently limited by the accuracy of prior knowledge and assumptions. Moreover, although both approaches have yielded important insights, interpreting the dense web of FC changes—especially from high-dimensional whole-brain matrices—remains a major challenge

The brain, however, is increasingly viewed as a modular system, with functional organization arising from the coordinated activity of interacting subnetworks rather than isolated regions. In line with this view, our previous work introduced an independent component analysis (ICA)–based method to uncover population-level variations in FC at the subnetwork scale, without relying on predefined anatomical or functional templates or prior assumptions.¹⁸ This method was successfully validated in a schizophrenia cohorts, where it identified reproducible and biologically meaningful alterations. Previous ICA applications in Alzheimer’s disease have typically analyzed fMRI images at the individual level,¹⁹ focusing on subject-specific components rather than applying ICA across subjects on functional connectivity matrices (our approach). Moreover, most prior studies aimed to identify common resting-state networks (RSNs).²⁰ In contrast, our study focuses on identifying coherent subnetworks, where each ICA component may capture variations associated with disease, age, gender, or other factors, thus providing a more targeted understanding of connectivity alterations.

In the present study, we build upon this framework to investigate AD. By identifying altered subnetworks, rather than isolated connections, we aim to capture coherent patterns of dysfunction that may reflect underlying disease mechanisms. This data-driven, hypothesis-free approach enables a more interpretable, modular view of connectivity changes in AD, potentially aiding in clinical insight and biomarker development.

Materials and Methods

Dataset

The dataset employed in this study was obtained from the Alzheimer’s Disease Neuroimaging Initiative (ADNI) database (www.adni.loni.usc.edu). ADNI was established in 2003 as a public-private partnership led by Principal Investigator Michael W. Weiner, MD. Specifically, we included subjects from the ADNI-2 phase, in which participants were recruited and assessed between 2011 and 2017. A total of 49 HCs and 34 individuals diagnosed with Alzheimer’s disease (AD) with baseline rs-fMRI data were included in this study (It was downloaded in April 2023). Demographic details, along with Mini-Mental State Examination (MMSE) scores and Clinical Dementia Rating (CDR) scores, are summarized in Table 1.

Table 1.

Information on the Participants Included in This Study.

Demographic Variables	HC (n = 49)	AD (n = 34)	P-value
Male/Female	22/27	16/18	.85
Age, $\frac{K (K - 1)}{2}$ (SD)	75.46 (7.01)	72.7 (7.00)	.081
MMSE, $\bar{X}$ (SD)	28.85 (1.31)	22.76 (2.45)	< .0001
CDR, $\bar{X}$ (SD)	0.01 (0.07)	0.79 (0.24)	< .0001
Years of education $\bar{X}$ (SD)	16.48 (2.15)	15.76 (2.84)	.19

Abbreviations: AD, Alzheimer’s disease; CDR, clinical dementia rating; HC, healthy control; MMSE, mini-mental state examination; SD, standard deviation.

Resting-state fMRI images were acquired using 3.0 Tesla Philips Medical Systems Scanners, following an Echo Planner Imaging (EPI) protocol with the following parameters: echo time (TE) = 30 ms, repetition time (TR) = 3000 ms, flip angle (FA) = 80°, pixel size = 3.3 × 3.3 mm², acquisition matrix size = 64 × 64, slice thickness = 3.3 mm, 48 slices per volume, and 140 volumes in total. Further details of the data acquisition description of the ADNI are available at http://adni.loni.usc.edu/.

Data Preprocessing and Matrix Construction

For each subject, the first 5 volumes were discarded to account for scanner stabilization and participant adaptation. The subsequent preprocessing steps were conducted using MATLAB 2021a in combination with SPM12 and DPARSF 5.4.²¹ The retained image data were then processed through a standard pipeline, including slice timing correction, head motion correction, spatial normalization, band-pass filtering (0.01-0.1 Hz) to minimize low-frequency drifts and high-frequency physiological artifacts, nuisance regression to remove confounding covariates, and nonlinear registration to MNI152 standard space. To construct individual functional connectivity (FC) matrices, Dosenboch atlas²² was used to conduct anatomical parcellation (160 ROIs). We then extracted the average time series from each 160 ROIs. Pearson correlation coefficients were then computed between each pair of ROIs time series, resulting in a 160 × 160 FC matrix per participant. The normality of the entire data was subsequently evaluated using the Shapiro–Wilk test.²³ The obtained P-value of .07, which exceeds the significance threshold of .05, suggested that the distribution of FC values is normal.

Subnetwork Extraction

To examine connections within the cerebral cortex, 18 regions associated with the cerebellum were removed, leaving 142 ROIs for subsequent proposed analysis. For each subject, the lower triangular part of the symmetric 142 × 142 FC matrix was vectorized into a 1-dimensional array of length $\frac{K (K - 1)}{2}$ , where K = 142. These FC vectors were stacked across subjects to form the input matrix X, where rows represent subjects (N) and columns correspond to unique pairwise connections. The ICA was then applied to the matrix X to extract latent sources, as in equation (1):

X = A \times S

(1)

where A is the mixing matrix and S is the independent source. We used the Infomax algorithm²⁴ to perform ICA. Each row in S represents one component, and the corresponding values define the “ICA weights” assigned to each edge in that component.

Since only a subset of components consistently contributes to the structure of X, we employed a modified RAICAR (Ranking and Averaging Independent Component Analysis by Reproducibility) algorithm²⁵ to identify reproducible components across 100 ICA runs, mitigating variability in component ordering.

In the next step, edge pruning was applied to each component to identify high-contributing edges—those most critical to ICA reproducibility. The edge pruning algorithm developed by Keyvanfard et al¹⁸ was employed in which the edges with maximum effects on the reversibility of the ICA procedure were kept, and the others were replaced by zero. Each pruned component is henceforth referred to as a “subnetwork.”

Statistical Analysis

The subnetworks were statistically assessed to identify those showing significant group differences between the HC and AD participants. For each subnetwork subNet (i), defined by its selected edges and corresponding ICA weights, we projected each subject’s functional connectivity (FC) vector, FC(j), onto the subnetwork using the inner product, yielding a single scalar projection score per individual for that particular subnetwork. Independent-sample t-test analysis was then applied between these projected values of the 2 groups $(P r j_{H C}, P r j_{S Z})$ .

\begin{array}{l} f o r s u b n e t w o r k (i) a n d s u b j e c t (j) : \\ P r j_{H C} (i, j) = 〈 s u b N e t (i), F C_{H C} (j) 〉 \\ P r j_{A D} (i, j) = 〈 s u b N e t (i), F C_{A D} (j) 〉 \end{array}

(2)

where Prj(i,j) denotes the projection score for subject j on subnetwork i, subNet (i), and ⟨x,y⟩ represents the inner product between the vectorized form of ICA-defined subnetwork and that individual’s functional connectivity. Group differences in these projection scores were tested using independent-sample t-tests. To correct for multiple comparisons across subnetworks, we applied the Bonferroni correction,²⁶ and considered P-value < .005 as statistically significant.

In addition to statistical significance, effect sizes were calculated using Cohen’s d, defined as the mean difference divided by the pooled standard deviation, to quantify the magnitude of group differences.

Graph Parameters

We performed graph-theoretical analysis to evaluate the connectivity characteristics of both the whole-brain and the obtained subnetworks. For whole-brain network analysis, the weighted and fully connected FC matrices of each individual were used without any thresholding or binarization. Five standard topological measures were computed: shortest path length, network strength, global efficiency, local efficiency, and clustering coefficient.

The topological measures employed in this study are defined as follows:

The shortest path length between 2 nodes, {L_{i, j}} refers to the minimum sum of edge weights needed to connect node i to node j. The characteristic path length of the network (L_net) is obtained by averaging the shortest path lengths across all node pairs (N) in the network²⁷:

L_{n e t} = \frac{1}{N (N - 1)} \sum_{i \neq j} L_{i j}

(3)

The network strength (S) is defined as the mean nodal strength across all nodes, where nodal strength corresponds to the sum of the absolute edge weights (wᵢⱼ) connected to a given node

S = \frac{1}{N} \sum_{i \neq j} w_{i j}

(4)

The global efficiency (E_global) quantifies the overall integration of the brain network.^28
-30 It is calculated as the inverse of the average shortest path length between all node pairs, as follows:

E_{g l o b a l} = \frac{1}{N (N - 1)} \sum_{i \neq j} \frac{1}{m i n {L_{i, j}}}

(5)

The local efficiency (E_{i_local}) reflects the efficiency of information exchange within the subgraph Gᵢ when node i is removed, thereby indicating the fault tolerance of the network.²⁹ It is expressed as:

E_{i_l o c a l} = \frac{1}{N_{G_{i}} (N_{G_{i}} - 1)} \sum_{i \neq j} \frac{1}{m i n {L_{i, j}}}

(6)

The clustering coefficient of a node (C_i) in a weighted graph is given by the ratio of the total triangle intensity (based on wᵢⱼ values) to the number of possible connections within the corresponding subgraph Gᵢ containing k nodes³¹:

C_{i} = \frac{E_{i}}{\frac{k_{i} (k_{i} - 1)}{2}} = \frac{2}{k_{i} (k_{i} - 1)} \sum_{j, k} {(w_{i j}, w_{i k}, w_{j k})}^{\frac{1}{3}}

(7)

In the next step, graph analysis was applied specifically to the significant subnetworks identified in the previous step. For each subject, a weighted subnetwork graph was constructed based on their original FC values at the selected edge locations. The same 5 network metrics were calculated and group comparisons were made using 2-sample t-tests between HCs and AD participants to determine significant differences between the AD and HC groups.

All graph-theoretical measures (for the weighted graphs) were computed using the Brain Connectivity Toolbox,³² and the statistical threshold was set at P < .05.

All data analysis steps were performed using MATLAB version 2021a.

Results

Applying the developed algorithm to the concatenated functional connectivity matrix of both the HC and AD groups resulted in the extraction of distinct brain subnetworks. Based on the robustness and the functional coherence of the connected regions, only the first 10 component were selected for further analysis, as they consistently represent stable and meaningful subnetworks with minimal sensitivity to the number of subjects or the number of components, as shown in prior work.³³ The edge-pruning step was performed using an empirical threshold of 0.18, ensuring that approximately 15% of the strongest edges were retained within each subnetwork. A higher threshold value produced 2 or 3 components with disconnected nodes, and reduced meaningful connectivity information. Conversely, a lower threshold resulted in overly dense subnetworks with numerous spurious connections, making the interpretation difficult.

Statistical Analysis

Among the 10 identified subnetworks, #4 (P = .015), #6 (P = .011) and #8 (P = .00035), exhibited statistically significant differences between AD patients and healthy controls based on the ICA-projected values (as described in equation (2)). However, only subnetwork #8 remained significant after Bonferroni correction (P < .005 threshold) for multiple comparisons. The P-values and corresponding Cohen’s d effect sizes of these 3 subnetworks are presented in Table 2.

Table 2.

The P-Value and Effect Size (Cohen’s d Value) of 3 Significant Subnetworks.

Subnetwork Number	HC (mean ± SD)	AD (mean ± SD)	P-value	Effect size
Subnetwork #4	5.8 ± 2.09	5.26 ± 0.29	.015	0.31
Subnetwork #6	4.94 ± 0.77	5.61 ± 2.44	.011	0.39
Subnetwork #8	4.28 ± 1.71	3.52 ± 0.85	.00035	0.52

Abbreviations: AD, Alzheimer’s disease; HC, healthy control; SD, standard deviation.

The bold row indicates the subnetwork that passed the Bonferroni correction for multiple comparisons.

The Areas/Links Most Affected by AD

For visualization purposes, the significant subnetworks were illustrated based on node strength and important connections. The node strength of all nodes was calculated with respect to the ICA values and then normalized to [0,1]. To better highlight the most affected brain regions and connections, nodes with strength values (ie, absolute ICA values) below the threshold of mean + standard deviation (SD) were excluded from the visualization. Figure 1 presents the 3 identified significant subnetworks. The node colors in Figure 1 represent the resting state networks (RSNs) to which each node belongs. According to the Dosenboch atlas,²² 5 RSN were considered: the default mode network (DMN), fronto-parietal (FP), cingulo-opercular (CO), sensory motor (SM), and occipital (OC) networks.

Figure 1.

Three significant subnetworks: from left to right #4, #6, & #8. The nodes are color-coded according to the networks introduced in the Dosenbach atlas.²² (red: DMN; yellow: fronto-parietal; black: cingulo-opercular; blue: sensory motor; green: occipital).

As illustrated in Figure 1, subnetwork #4 predominantly includes nodes from the cingulo-opercular and frontoparietal RSNs. Some regions from these 2 RSNs also appear in subnetwork #6, along with additional nodes from the DMN (shown in red). In subnetwork #8, the included regions are distributed across 3 RSNs; the DMN, SM, and OC; forming a distinct connectivity pattern. To improve the clarity of the regions involved in each subnetwork, the node strength of each subnetwork is shown in Figure 2 without the connections. Node strength was calculated using the original FC data from the HC group, and the size of each node reflects its relative importance in the obtained subnetworks.

Figure 2.

Node strength in the 3 subnetworks. The size of each node is indicative of its strength, which was computed as the summation of the absolute weights of all edges connected to it. Nodes with a larger size have a greater node strength.

Correlation of Subnetworks With Clinical Measures

The relationships between the MMSE scores of the AD and HC groups and the original edge weights in the obtained subnetworks were examined using Pearson correlation analysis. The connections that exhibited significant correlations (P-value < .05) are visualized in Figure 3A and B. Figure 3A displays the correlation coefficient (r) of the FCs across the 3 subnetworks, while Figure 3B presents the corresponding P-values. Furthermore, Figure 4 illustrates the mean value of this relationship between MMSE scores and the original weights of the selected edges within the 3 subnetworks. A significant positive correlation was observed between MMSE scores and the mean value of these connections (subnetwork 4: r = .28, P = .01; subnetwork 6: r = .65, P < 10⁻⁵; subnetwork 8: r = .45, P < 10⁻⁴).

Figure 3.

Illustration of connections that are significantly correlated with MMSE scores in 3 subnetworks. The correlation coefficient (r-value) (A) and the P-value (B) are color coded as depicted in the color bar.

Figure 4.

Correlation analysis of the mean FC values and MMSE. A significant positive correlation was observed for all 3 subnetworks (A-C). The red line represents the fitting line for the data, and the red dashed line represents the 95% confidence interval of the fitting line.

Graph Parameters

There were no significant differences in characteristic path length, network strength, global efficiency, local efficiency, or the clustering coefficient of the whole-brain network between the HC and AD groups (Table 3). Nevertheless, subnetwork #8 showed a significant group differences in 3 graph metrics: network strength, clustering coefficient, and local efficiency (all P < .05), highlighting localized topological disruptions in AD (Figure 5 and Table 3). All measured graph metrics of the whole brain and subnetwork #8 are presented in Table 3. It should be noted that the p-values for these graph parameters in other subnetworks were greater than 0.1. Together, these findings indicate that while global network topology remains relatively preserved in AD, specific functional subnetworks—especially subnetwork #8—exhibit both altered connectivity and reduced efficiency, with meaningful associations to cognitive decline.

Table 3.

Network Topology Metrics of subnetwork #8 as well as the Whole Brain.

	Subnetwork #8			Whole Brain
Graph Metrics	HC (mean ± SD)	AD (mean ± SD)	P-value	HC (mean ± SD)	AD (mean ± SD)	P-value
Network strength	4.73 ± 0.76	4.41 ± 0.52	.03	27.86 ± 3.5	27.08 ± 3.45	.32
Shortest path length	7.05 ± 0.84	7.35 ± 0.71	.08	4.16 ± 0.36	4.22 ± 0.30	.39
Clustering coefficient	0.035 ± 0.007	0.032 ± 0.0038	.02	0.16 ± 0.02	0.15 ± 0.02	.32
Global efficiency	0.16 ± 0.02	0.15 ± 0.017	.06	0.27 ± 0.026	0.26 ± 0.024	.34
Local efficiency	0.086 ± 0.016	0.079 ± 0.0096	.03	0.16 ± 0.02	0.16 ± 0.02	.32

Abbreviations: AD, Alzheimer’s disease; HC, healthy control; SD, standard deviation.

Subnetwork #8 showed significant group differences (P < .05) in network strength, clustering coefficient, and local efficiency. No significant differences were observed at the whole-brain level for these metrics.

Figure 5.

The significant graph metrics, including the clustering coefficient (CC), local efficiency (E-local) and network strength (str), were obtained from subnetwork #8. The right y-axis values show the network strength measure, and the left y-axis values are related to the CC and E-local values.

Discussion

In the present study, we applied the previously proposed ICA-based algorithm¹⁸ to 83 individuals (AD and HC) to identify the AD-associated subnetworks. The first 10 components were selected for further analysis. Statistical comparison of the functional connectivity in the subnetworks revealed 3 subnetworks with significant group differences. After multiple comparisons correction, only subnetwork #8 remained statistically significant.

The algorithm was designed to decompose the variations in functional connectivity data into distinct components, which serve as the fundamental building blocks of group variations. Consequently, a particular brain disorder, identified as one of the intergroup variations, is expected to manifest within a limited subset of these derived components. The remaining subnetworks, which did not show significant group differences between AD and HC groups, may be associated with unrelated interindividual variability such as age. As the main goal of this study was to identify subnetworks that are particularly useful for distinguishing between AD patients and healthy individuals, no detailed discussion of the non-discriminative components is included.

Anatomical Distribution of Subnetworks

The regions linked to subnetwork #4 are primarily associated with the cingulo-opercular (CO) and frontoparietal (FP) RSNs according to the Dosenbach atlas. These networks are recognized as key cognitive control systems that play important roles in the executive function.³⁴ Prior research has demonstrated that their activation increases during complex cognitive tasks,³⁵ and stronger connectivity within these network predicts cognitive performance.^36
-39 Some studies propose that CO and FP form a dual top-down control system supporting higher-order cognitive abilities.^40,41 While these RSNs (FP and CO) are not primary targets of AD pathology, they have been linked to cognitive changes in AD, including associations with apathy,⁴² fatigue,⁴³ and deficits in Schizophrenia.⁴¹ Their interactions have also been studied in older adults and in relation to working memory. Altogether, CO and FP may play a secondary or compensatory role in cognitive processes affected by AD.

Subnetwork #6 includes connections between the CO, FP and DMN, although the resulting regions are mostly different from those in subnetwork#4. The fusiform gyrus is one of the DMN regions in this subnetwork, which has been reported to be affected in AD.^44
-46

Subnetwork #8, showing the most significant differences between HCs and AD patients, spans DMN, SM, and CO nodes. The ventromedial prefrontal cortex (vmPFC),one of the nodes in this subnetwork, is a major hub within the DMN. It is recognized as a key brain area involved in decision-making, memory representation, emotional processing and regulation,^47
-50 as well as relational memory and inferential reasoning.⁵¹ In addition, the vmPFC has been shown to correlate with insight loss in AD⁵² and age-related DMN changes are more pronounced in AD compared to healthy older adults.^53
-55 It have also reported that connections between the vmPFC and other regions—including the precentral gyrus (PRG) and occipital nodes—are disrupted in aging-related disorders such as AD.^56,57 Moreover, previous studies have confirmed functional alterations in the CO^58,59 and SM network^11,12,16 in AD patients.

Interestingly, the spatial distribution of subnetwork #8, and to a lesser extend subnetwork #6, shows a right-lateralized pattern of disruption. Altered connections are mainly in the right hemisphere (Figure 1), reflected by larger node sizes (Figure 2) and more significant MMSE-related connections (Figure 3). Previous studies report similar right-lateralized abnormalities in MCI and AD during task and resting-state assessments.^34,60
-62 This atypical lateralization suggests compensatory mechanisms in the right hemisphere to counteract left-hemispheric network degradation.^60
-62 The emergence of this pattern from a fully data-driven, hypothesis-free approach underscores the power of blind methods to detect disease-specific functional reorganizations.

Graph Analysis

Another interesting finding concerns the topological properties of the identified subnetworks. Network strength, reflecting overall connectivity, was reduced in the AD-specific subnetwork. Similarly, local efficiency and clustering coefficients—indices of fault tolerance and local information exchange—were also significantly lower in subnetwork #8, indicating reduced resilience in AD patients. Such reduced across the entire brain have already been reported in prior fMRI studies on AD.^63
-67 Notably, many previous studies have focused on significant alterations in these graph-based metrics at the whole-brain level in individuals with AD. However, depending on the dataset, such global changes may not always reach statistical significance. Nonetheless, analyzing targeted subnetworks can reveal meaningful alterations. Such focused analysis may improve diagnostic sensitivity and help identify potential targets for intervention.

Limitations and Future Work

In ICA, the number of extracted components generally matches the dimensionality of the data (eg, number of subjects), but only a subset of these components is expected to be reproducible and functionally relevant. Employing a larger dataset leads to the extraction of more reliable and consistent components. Therefore, by analyzing a larger cohort and providing their connectivity matrices as input to the algorithm, it becomes possible to identify a greater number of reproducible subnetworks that exhibit higher intergroup variability. Accordingly, the relatively small sample size in the present study constitutes a key limitation. Nevertheless, our analysis successfully identified an AD-specific subnetwork whose graph-theoretical metrics were significantly different from those of healthy controls and were also significantly correlated with MMSE scores, thereby demonstrating the method’s utility even with a modest sample size.

Reproducibility and the ordering of components in the ICA algorithm represent another important challenge in this study. To address this limitation, we employed the RAICAR algorithm,^18,25 which enhances the stability and consistency of component identification by ranking them across multiple ICA runs. Although this procedure substantially improves reliability, it requires repeated executions of the algorithm, which makes the process time-consuming. Nevertheless, this trade-off was considered acceptable, as it allowed us to prioritize robust and reproducible components for subsequent analyses, thereby strengthening the overall validity of our results.

In addition to the small sample size and variability in ICA decomposition, preprocessing choices may also influence the reproducibility of our findings. Steps such as motion correction, spatial normalization, and temporal filtering can affect the resulting functional connectivity patterns and, consequently, the extracted components. While we followed standard preprocessing pipelines recommended for fMRI data, future studies could explore the sensitivity of the results to alternative preprocessing strategies to further strengthen the generalizability of the findings.

The outcomes of this study suggest that adopting a modular, network-based perspective of the brain can facilitate a more in-depth understanding of the connectivity alterations induced by neurological abnormalities. Rather than focusing solely on isolated altered connections, this approach enables the identification of specific functional subnetworks associated with particular conditions, such as the AD-specific subnetwork highlighted in this study. This finding opens promising avenues for future investigations aimed at utilizing these identified subnetworks as potential biomarkers for the classification and early diagnosis of AD. In addition, tracking the temporal dynamics of such subnetworks could serve as a foundation for assessing disease progression and prognosis. Moreover, future work should aim to investigate the connections among the nodes identified within the subnetwork, highlighting the importance of analyzing these relationships from a neuroscience perspective. Furthermore, the relationships between these identified regions can be explored, particularly since they are classified as a cohesive subnetwork (eg, Subnetwork #8) following a data-driven, blind approach. Future work should also aim to replicate these findings across larger and independent cohorts to confirm the generalizability of the identified subnetworks.

Conclusion

In conclusion, this study employed a systematic, hypothesis-free approach to identify functional subnetworks affected by Alzheimer’s disease. The resulting subnetworks primarily involved the DMN, SM, CO, and FP networks, and showed significant positive correlations with MMSE scores. Notably, one subnetwork demonstrated a particularly strong association with AD, reflected in both altered graph-theoretical metrics and patterns of hemispheric lateralization. These findings suggest that the data-driven method effectively isolates functionally meaningful modules of brain connectivity disruption. Overall, this modular framework offers valuable insights into network-level alterations in Alzheimer’s disease and may hold promise for broader applications in other neurological disorders.

Footnotes

Acknowledgements

Data used in the preparation of this article were obtained from the Alzheimer’s Disease Neuroimaging Initiative (ADNI) database (www.loni.ucla.edu/ADNI). As such, the investigators within the ADNI contributed to the design and implementation of the ADNI and/or provided data but did not participate in the analysis or writing of this report. A complete listing of ADNI investigators can be found at . We especially thank Dr. Ataollah Ebrahimzadeh (Department of Electrical and Computer Engineering, Babol Noshirvani University of Technology, Babol, Iran) for providing preprocessed data.

ORCID iD

Farzaneh Keyvanfard

Ethical Considerations

Data used in the preparation of this article were obtained from the Alzheimer’s Disease Neuroimaging Initiative (ADNI) database (adni.loni.usc.edu), (National Institutes of Health Grant U01 AG024904) and DOD ADNI (Department of Defense award number W81XWH-12-2-0012). ADNI is funded by the National Institute on Aging, the Foundation for the National Institutes of Health, the Alzheimer’s Association, etc. A complete listing of ADNI investigators can be found at . The ADNI study was approved by the Institutional Review Boards (IRBs) of all participating centers, and written informed consent was obtained from all participants or their authorized representatives, in accordance with the Declaration of Helsinki. As the present analysis used only de-identified ADNI-2 data, no additional ethical approval or study-specific approval number was required. For further information on ADNI ethics approvals, see the ADNI website (adni.loni.usc.edu).

Consent to Participate

All participants provided written informed consent to participate in this study.

Consent for Publication

We consent to the publication of this research.

Author Contributions

FK and AN: methodology and investigation. FK: formal analysis, writing—original draft, and writing—review and editing. AN: conceptualization, investigation, writing—review and editing, supervision. All authors contributed to the article and approved the submitted version.

Funding

The authors received no financial support for the research, authorship, and/or publication of this article.

Declaration of Conflicting Interests

The authors declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

Data Availability Statement

All ADNI data are deposited in a publicly accessible repository and can be accessed at ADNI ().

Code Availability

The code and materials generated or used during this study is publicly available at .

References

Tarawneh

Holtzman

DM.

The clinical problem of symptomatic Alzheimer disease and mild cognitive impairment. Cold Spring Harb Perspect Med. 2012;2:a006148.

Barnes

Dickerson

Frost

Jiskoot

Wolk

van der Flier

WM.

Alzheimer’s disease first symptoms are age dependent: evidence from the NACC dataset. Alzheimer’s Dement. 2015;11:1349-1357.

Better

MA.

Alzheimer’s disease facts and figures. Alzheimer Dement. 2023;19:1598-1695.

Delbeuck

Collette

Van der Linden

Is Alzheimer’s disease a disconnection syndrome?: Evidence from a crossmodal audio-visual illusory experiment. Neuropsychologia. 2007;45:3315-3323.

Brier

Thomas

Ances

BM.

Network dysfunction in Alzheimer’s disease: refining the disconnection hypothesis. Brain Connect. 2014;4:299-311.

Mousa

Zayed

Yassine

IA.

Alzheimer disease stages identification based on correlation transfer function system using resting-state functional magnetic resonance imaging. PLoS One. 2022;17:1-17.

Jing

Chen

Wei

, et al. Altered large-scale dynamic connectivity patterns in Alzheimer’s disease and mild cognitive impairment patients: a machine learning study. Hum Brain Mapp. 2023;44:3467-3480.

Lama

Kwon

GR.

Diagnosis of Alzheimer’s disease using brain network. Front Neurosci. 2021;15:1-13.

Ibrahim

Suppiah

Ibrahim

, et al. Diagnostic power of resting-state fMRI for detection of network connectivity in Alzheimer’s disease and mild cognitive impairment: a systematic review. Hum Brain Mapp. 2021;42:2941-2968.

10.

Alorf

Khan

MUG

. Multi-label classification of Alzheimer’s disease stages from resting-state fMRI-based correlation connectivity data and deep learning. Comput Biol Med. 2022;151:106240.

11.

Fathian

Jamali

Raoufy

MR.

The trend of disruption in the functional brain network topology of Alzheimer’s disease. Sci Rep. 2022;12:14998.

12.

Lin

Huang

, et al. Abnormal dynamic functional connectivity in Alzheimer’s disease. CNS Neurosci Ther. 2020;26:962-971.

13.

Palesi

Castellazzi

Casiraghi

, et al. Exploring patterns of alteration in Alzheimer’s disease brain networks: a combined structural and functional connectomics analysis. Front Neurosci. 2016;10:1-16.

14.

Filippi

Spinelli

Cividini

Agosta

Resting state dynamic functional connectivity in neurodegenerative conditions: a review of magnetic resonance imaging findings. Front Neurosci. 2019;13:464333.

15.

Weiler

de Campos

de Ligo Teixeira

, et al. Intranetwork and internetwork connectivity in patients with Alzheimer disease and the association with cerebrospinal fluid biomarker levels. J Psychiatry Neurosci. 2017;42:366-377.

16.

Wang

Zhou

Yao

, et al. Aberrant intra- and inter-network connectivity architectures in Alzheimer’s disease and mild cognitive impairment. Sci Rep. 2015;5:14824.

17.

Keyvanfard

Schmid

A-K

Nasiraei-Moghaddam

Functional connectivity alterations of within and between networks in schizophrenia: a retrospective study. Basic Clin Neurosci J. 2023;14:397-410.

18.

Keyvanfard

Nasab

Nasiraei-Moghaddam

Brain subnetworks most sensitive to alterations of functional connectivity in schizophrenia: a data-driven approach. Front Neuroinformatics. 2023;17:1175886. doi:10.3389/fninf.2023.1175886

19.

Chou

Chang

Remedios

Butman

Chan

Pham

DL.

Automated classification of resting-state fMRI ICA components using a deep Siamese network. Front Neurosci. 2022;16:768634.

20.

Soman

Raghavan

Rajesh

, et al. Does resting state functional connectivity differ between mild cognitive impairment and early Alzheimer’s dementia? J Neurol Sci. 2020;418:117093.

21.

Chao-Gan

Yu-Feng

DPARSF: a MATLAB toolbox for ‘pipeline’ data analysis of resting-state fMRI. Front Syst Neurosci. 2010;4:1-7.

22.

Dosenbach

NUF

Nardos

Cohen

, et al. Prediction of individual brain maturity using fMRI. Science. 2010;329:1358-1361.

23.

Shapiro

Wilk

Chen

HJ.

A comparative study of various tests for normality. J Am Stat Assoc. 1968;63:1343-1372.

24.

Bell

Sejnowski

TJ.

An information-maximization approach to blind separation and blind deconvolution. Neural Comput. 1995;7:1129-1159.

25.

Yang

LaConte

Weng

Ranking and averaging independent component analysis by reproducibility (RAICAR). Hum Brain Mapp. 2008;29:711-725.

26.

Benjamini

Hochberg

Controlling the false discovery rate: a practical and powerful approach to multiple testing. J R Stat Soc Ser B. 1995;57:289-300.

27.

Watts

Strogatz

SH.

Collective dynamics of ‘small-world’ networks. Nature. 1998;393:440-442.

28.

Achard

Bullmore

Efficiency and cost of economical brain functional networks. PLoS Comput Biol. 2007;3:e17.

29.

Latora

Marchiori

Efficient behavior of small-world networks. Phys Rev Lett. 2001;87:198701-198704.

30.

Latora

Marchiori

Economic small-world behavior in weighted networks. Eur Phys J B Condens Matter Complex Syst. 2003;32:249-263.

31.

Onnela

J-P

Saramäki

Kertész

Kaski

Intensity and coherence of motifs in weighted complex networks. Phys Rev E. 2005;71:065103.

32.

Rubinov

Sporns

Complex network measures of brain connectivity: uses and interpretations. Neuroimage. 2010;52:1059-1069.

33.

Keyvanfard

Nasiraei-Moghaddam

Hagmann

Interindividual covariations of brain functional and structural connectivities are decomposed blindly to subnetworks: a fusion-based approach. J Magn Reson Imaging. 2020;51:1779-1788.

34.

Hausman

Hardcastle

Albizu

, et al. Cingulo-opercular and frontoparietal control network connectivity and executive functioning in older adults. GeroScience. 2022;44:847-866.

35.

Dosenbach

Visscher

Palmer

, et al. A core system for the implementation of task sets. Neuron. 2006;50:799-812.

36.

Seeley

Menon

Schatzberg

, et al. Dissociable intrinsic connectivity networks for salience processing and executive control. J Neurosci. 2007;27:2349-2356.

37.

Kelly

Uddin

Biswal

Castellanos

Milham

MP.

Competition between functional brain networks mediates behavioral variability. Neuroimage. 2008;39:527-537.

38.

Rypma

Berger

Prabhakaran

, et al. Neural correlates of cognitive efficiency. Neuroimage. 2006;33:969-979.

39.

Song

Zhou

, et al. Brain spontaneous functional connectivity and intelligence. Neuroimage. 2008;41:1168-1176.

40.

Dosenbach

Fair

Cohen

Schlaggar

Petersen

SE.

A dual-networks architecture of top-down control. Trends Cogn Sci. 2008;12:99-105.

41.

Sheffield

Repovs

Harms

, et al. Fronto-parietal and cingulo-opercular network integrity and cognition in health and schizophrenia. Neuropsychologia. 2015;73:82-93.

42.

Tumati

Marsman

De Deyn

Martens

Aleman

Functional network topology associated with apathy in Alzheimer’s disease. J Affect Disord. 2020;266:473-481.

43.

Ramage

Ray

Franz

Tate

Lewis

Robin

DA.

Cingulo-opercular and frontoparietal network control of effort and fatigue in mild traumatic brain injury. Front Hum Neurosci. 2021;15:1-17.

44.

Cai

Chong

Zhang

, et al. Altered functional connectivity of fusiform gyrus in subjects with amnestic mild cognitive impairment: a resting-state fMRI study. Front Hum Neurosci. 2015;9:150842.

45.

Bokde

Lopez-Bayo

Meindl

, et al. Functional connectivity of the fusiform gyrus during a face-matching task in subjects with mild cognitive impairment. Brain. 2006;129:1113-1124.

46.

Mondragón

Maurits

De Deyn

PP.

Functional connectivity differences in Alzheimer’s disease and amnestic mild cognitive impairment associated with AT(N) classification and anosognosia. Neurobiol Aging. 2021;101:22-39.

47.

Lopez-Persem

Verhagen

Amiez

Petrides

Sallet

The human ventromedial prefrontal cortex: sulcal morphology and its influence on functional organization. J Neurosci. 2019;39:3627-3639.

48.

Mayeli

Misaki

Zotev

, et al. Self-regulation of ventromedial prefrontal cortex activation using real-time fMRI neurofeedback-influence of default mode network. Hum Brain Mapp. 2020;41:342-352.

49.

Mai

Liu

The default mode network and social understanding of others: What do brain connectivity studies tell us. Front Hum Neurosci. 2014;8:52017.

50.

Jobson

Hase

Clarkson

Kalaria

RN.

The role of the medial prefrontal cortex in cognition, ageing and dementia. Brain Commun. 2021;3:fcab125. doi:10.1093/braincomms/fcab125

51.

Wing

D’Angelo

Gilboa

Ryan

JD.

The role of the ventromedial prefrontal cortex and basal forebrain in relational memory and inference. J Cogn Neurosci. 2021;33:1976-1989.

52.

Hornberger

Yew

Gilardoni

, et al. Ventromedial-frontopolar prefrontal cortex atrophy correlates with insight loss in frontotemporal dementia and Alzheimer’s disease. Hum Brain Mapp. 2014;35:616-626.

53.

Jones

Machulda

Vemuri

, et al. Age-related changes in the default mode network are more advanced in Alzheimer disease. Neurol. 2011;77:1524-1531.

54.

Wang

Risacher

West

, et al. Altered default mode network connectivity in older adults with cognitive complaints and amnestic mild cognitive impairment. J Alzheimers Dis. 2013;35:751-760.

55.

Yan

, et al. Aging-related changes in the default mode network and its anti-correlated networks: a resting-state fMRI study. Neurosci Lett. 2011;504:62-67.

56.

Xue

Yuan

Yue

, et al. Distinct disruptive patterns of default mode subnetwork connectivity across the spectrum of preclinical Alzheimer’s disease. Front Aging Neurosci. 2019;11:483881.

57.

Dennis

Thompson

PM.

Functional brain connectivity using fMRI in aging and Alzheimer’s disease. Neuropsychol Rev. 2014;24:49-62.

58.

Huang

Beach

Bozoki

Zhu

DC.

Alzheimer’s disease progressively reduces visual functional network connectivity. J Alzheimer’s Dis Reports. 2021;5:549-562.

59.

Plaza-Rosales

Brunetti

Montefusco-Siegmund

, et al. Visual-spatial processing impairment in the occipital-frontal connectivity network at early stages of Alzheimer’s disease. Front Aging Neurosci. 2023;15:1097577.

60.

Liu

Zhang

, et al. Changes in brain lateralization in patients with mild cognitive impairment and Alzheimer’s disease: a resting-state functional magnetic resonance study from Alzheimer’s disease neuroimaging initiative. Front Neurol. 2018;9:1.

61.

Wang

Zhao

Shi

Zeng

Research on the lateralization of brain functional complexity in mild cognitive impairment-Alzheimer’s disease progression based on multiscale lateralized brain entropy. Biomed Signal Process Control. 2023;86:105216.

62.

Tyrer

Gilbert

Adams

, et al. Lateralized memory circuit dropout in Alzheimer’s disease patients. Brain Commun. 2020;2:fcaa212.

63.

Sanz-Arigita

Schoonheim

Damoiseaux

, et al. Loss of ‘small-world’ networks in Alzheimer’s disease: graph analysis of fMRI resting-state functional connectivity. PLoS One. 2010;5:e13788.

64.

Ahmadi

Fatemizadeh

Motie-Nasrabadi

Identifying brain functional connectivity alterations during different stages of Alzheimer’s disease. Int J Neurosci. 2022;132:1005-1013.

65.

Brier

Thomas

Fagan

, et al. Functional connectivity and graph theory in preclinical Alzheimer’s disease. Neurobiol Aging. 2014;35:757-768.

66.

Zhao

Tang

Feng

Sui

The effects of neuron heterogeneity and connection mechanism in cortical networks. Phys A Stat Mech its Appl. 2008;387:5952-5957.

67.

Farahani

Karwowski

Lighthall

NR.

Application of graph theory for identifying connectivity patterns in human brain networks: a systematic review. Front Neurosci. 2019;13:1-27.