This study was performed in accordance with the local Institutional Review Board. All participants gave written, informed consent in this HIPAA-compliant study.

The study included a total of 15 HC (9 females; mean age [SD]: 74.3 [6.5] years) and 23 individuals with CI (14 females; mean age [SD]: 76.9 [6.8] years). The CI group was comprised of 13 AD patients in the stage of dementia (7 females; mean age [SD]: 77.8 [8.2] years) and 10 MCI patients due to AD (7 females; mean age [SD]: 75.8 [4.8] years). All individuals with CI were referred from local neurology clinics or selected from the Arizona Alzheimer’s Disease Center (AADC) database (15 individuals from neurology clinics and 8 individuals from AADC). The inclusion of subjects in each cohort was based on clinical diagnostic criteria determined by a practicing neurologist (Albert et al., 2011; McKhann et al., 2011). Prior to the MRI scan, all participants underwent cognitive assessments, including the Montreal Cognitive Assessment (MoCA; Nasreddine et al., 2005), and Clock-Draw test (Sunderland et al., 1989). The complete subject characteristics are summarized in Table 1.

MRI data were acquired at 3 T (Ingenia, Philips, Best, Netherlands) with a dedicated 32-channel head coil. Standard T1-weighted (T1-w) anatomical images were acquired using a 3D magnetization-prepared rapid acquisition gradient echo (MP-RAGE) sequence with the following acquisition parameters: repetition time / echo time (TR/TE), 6.7/3.104 msec; acquisition matrix, 256 × 256; voxel size, 1.06 × 1.06 mm; slice thickness, 1.2 mm; 170 sagittal slices; flip angle = 9⁰. rs-fMRI was acquired using an echo-planar imaging (EPI) acquisition with TR/TE, 3.000/30.0 msec acquisition matrix, 215 × 215; voxel size, 2.68 × 2.68 mm; slice thickness, 2.68 mm; flip angle = 80⁰; 140 volumes.

The rs-fMRI datasets were converted to NIFTI format using dcm2niix.¹ To ensure a signal steady state, the first three-time frames were removed from each dataset (3dTcat, AFNI)² and temporal de-spiking was applied using 3dDespike (AFNI).

Image distortions can arise from head movements during data acquisition, potentially compromising data quality and impeding accurate interpretation of results. Even minor head movements can introduce variations in measured blood oxygenation levels, contributing to increased data variability. Therefore, implementing data preprocessing to minimize head movement is essential to improve the final results. In this study, motion parameters were estimated, and realignment of each time series was performed using 3dvolreg (AFNI). Following visual inspection of the estimated motion parameters, participants with motion exceeding thresholds (>3 mm translation, >3 degrees rotation) were excluded from the study. These criteria were established in alignment with voxel size and considerations for the anticipated spatial resolution of BOLD responses, accounting for inherent variability in brain anatomy across subjects.

A brain mask was generated by bet (Smith, 2002; FSL)³ based on the average volume of each motion-corrected time series. Each brain-extracted rs-fMRI was then coregistered to the corresponding relative MPRAGE (also brain extracted by bet) and spatially normalized to the standard MNI template (2 mm) using a 12-parameter affine transformation and mutual-information cost function (FLIRT; Jenkinson and Smith, 2001; FSL). Additionally, the data were resampled to isotropic resolution with a Gaussian kernel of FWHM = 4 mm (FSL). Voxel time-courses were bandpass filtered using AFNI (0.008 Hz < f < 0.1 Hz) to highlight low-frequency correlations during resting state.

The rs-fMRI datasets of the two groups were separately used for group ICA analysis using the Multivariate Exploratory Linear Optimized Decomposition into Independent Components (MELODIC) program (version 3.15; Beckmann and Smith, 2004), which is part of the FSL package. The analysis option of multi-session temporal concatenation was chosen to extract common spatial patterns without assuming a consistent temporal response across subjects. The resulting independent components (IC) maps were thresholded using a mixture model and alternative hypothesis testing approach. The threshold parameter was set to 0.5 to achieve an equal weight in false positives and false negatives (Woolrich et al., 2005).

Examining various component options and understanding the reasoning behind their selection is crucial for gaining insights into the robustness of identified resting components. In this study, we analyzed all MELODIC components for each network, averaging them when multiple components were present, to ensure a more resilient identification of the resting network. Furthermore, we scrutinized three key aspects for each component: the spatial map, the time course, and the power spectrum of the time course. Following this procedure, the DMN, VN, ECN, and FPN were the specific networks analyzed in the study.

Dual regression, through the dual_regression tool in FSL, was employed to estimate customized group-level spatial maps and time courses for each participant’s 4D data across the four networks of interest (Nickerson et al., 2017). The process involved two steps. In the first stage, the group spatial maps were regressed onto each participant’s 4D dataset, producing a set of time courses. In the second stage, these time courses were further regressed onto the same 4D dataset, yielding subject-specific spatial maps. Consequently, this analysis generated a component-time course and spatial map for each subject and each component.

Briefly, before performing d-CAP processing, voxel signals of the rs-fMRI data, in MNI standard space, were normalized by demeaning and dividing by the temporal SD. Conventional CAP analysis utilizes a single seed region of interest (ROI) to identify fMRI volumes associated with a specific resting-state network, such as the VN shown in Figure 1A. In this study, we generated automated 3 mm sphere ROIs within each network. Subsequently, the mean signal intensity within each ROI was computed and averaged, yielding a singular time-course per subject for each network. The spatial correlation between the average of selected time frames and the network-seeded correlation map increased with more frames by lowering the threshold. As a result, the spatial correlation led to a threshold of 25% (Figure 1B), and network-associated time points were included when the averaged seed signal intensities exceeded the chosen threshold. Whole-brain fMRI signals at these time points were identified for each subject in both HC and CI groups and were temporally concatenated, as depicted in Figure 1C.

Temporal clustering was performed on the concatenated time frames from both groups to ensure correspondence between CAPs from the HC and CI groups. K-means clustering was used with multiple cluster numbers (k = 1, 2, 3; …). Group-specific CAP sets (Ski1, Ski2, …) were computed by averaging spatial maps corresponding to time frames assigned to each cluster k_i from the HC and CI groups (Figure 1D).

In this study, a separate d-CAP set was calculated for each group, giving equal consideration to every CAP generated from different k-means runs to avoid bias from a single run. The final d-CAP set integrated clustering results from multiple k-means runs, capturing dynamic structures within network-associated time frames. After determining each group’s final d-CAP set, fMRI signals at network-associated time points were assigned to different clusters based on their spatial similarities to the d-CAPs. Spatial correlations were calculated between each network-associated time frame and every d-CAP to evaluate spatial resemblance. This information was utilized to assign the specific cluster (Figure 1E).

From d-CAPs analysis, several metrics can be obtained (Figure 1F):

with j = 1,2,3,…., number of d-CAPs.

Switching probability: the switching probability (SP) can be also computed for each subject to measure the dynamics of d-CAPs associated with each resting-state network. As previously mentioned, network-associated time frames are assigned to respective d-CAPs. If two consecutive time frames are assigned to different d-CAPs, it is considered a d-CAP switch. A reduced switching probability in one group reflects a less dynamic network. The SP for a subject is defined as follows:

with s = 1,2,3,…., number of subjects.

The mean and SD are provided for age and cognitive assessment scores. Age differences were examined using the Student t-test (Shapiro–Wilk: W = 0.97; p = 0.41), while differences in cognitive test scores were analyzed using the Mann–Whitney test (Shapiro–Wilk: p < 0.001 for all tests). Differences in motion during the rs-fMRI acquisition were evaluated by the Mann–Whitney test (Shapiro–Wilk: p < 0.001). Statistical significance was set at p < 0.05, with a power analysis of 0.76 (α=0.05; effect-size = 0.80).

In the dual-regression procedure (ICA), the group-level statistical inference was performed through a linear model at the voxel level, with motion, age and sex as covariates, using an in-house R script (version 4.3.1) and RStudio (version 2023.06.0). p-values were corrected using the False Discovery Rate (FDR) method to address multiple comparisons. Probabilistic threshold-free cluster enhancement (pTFCE; Spisák et al., 2019) was used to control for multiple comparison correction on each component separately with corrected probability of 0.05 determined as the significance threshold. Furthermore, given our analysis of four distinct networks, we applied the Bonferroni correction to adjust the final p-values. For ICA, effect-size and associated confidence interval were calculated by Cohen-d with R and RStudio (large effect for d > ±0.80).

d-CAPs were processed using an in-house R and RStudio script, employing the algorithm described in Zhuang et al. (2018). In this analysis, we explored a range of cluster numbers in k-means, spanning from 2 to 20. To mitigate the instability of single-trial k-means outcomes, we repeated the clustering process 100 times. The clustering result with the lowest within-cluster sum of point-to-centroid distances was selected as the optimal outcome.

To investigate differences in the switching probability of each network, a Student t-test was conducted with age and gender as covariates. FDR correction was used to address multiple comparisons. Differences in the temporal fraction of the 1^st d-CAPs for the networks with the same number of d-CAPs was also evaluated by Student t-test with age and gender as covariates, followed by FDR correction.

Correlations between switching probabilities and cognitive scores were evaluated by a linear model, with age and sex as covariates, by R and RStudio. The p-values were corrected by FDR. Effect size for correlations was evaluated by Spearman’s rank correlation coefficient (ρ).

The resulting clusters were labeled according to the automated anatomical labeling atlas (AAL; Tzourio-Mazoyer et al., 2002).

The Z-values, from the dual-regression procedure, for the DMN are shown in Figure 2. Panel A shows the statistical differences between HC and CI for ICA analysis. Significantly lower connectivity was found in the CI group than HC in the right precentral (<t > =2.979) and right postcentral gyrus (<t > =3.096). The complete results, with effect sizes and confidence intervals, are reported in Table 3A.

Panel B displays the spatial patterns of the blood oxygen level-dependent (BOLD) signal for the final d-CAP sets of the DMN. For this network, for both groups, we found 3 d-CAPs. Each d-CAP’s spatial map was also converted to effect size (Cohen’s d) maps, which were thresholded at d = ±0.80, representing the large effect size. For each d-CAP, between-group differences in Cohen’s d maps at d ≥ ±0.80 are reported in panel C and Table 3B. Increased coactivation (indicated by large effect size) for all d-CAPs was observed in the HC group. On the other hand, for d-CAP-2, small clusters of larger activity were observed in CI groups inside the left cuneus and precuneus and inside the left superior occipital gyrus. Spatial similarities between d-CAPs are represented by correlation matrices and are shown in panel D. The between-group d-CAP spatial similarities are shown inside the ‘red’ box. For this network, high spatial similarity was found for d-CAP-1 between HC and CI (r > 0.90) and for d-CAP-1 (HC) and d-CAP-3 (CI; r > 0.80).

Figure 3 displays the Z-values obtained from the dual-regression procedure for the VN. In panel A, the statistical differences between the HC and CI groups are reported for ICA analysis. Significantly lower connectivity was found in the CI group than HC, including in the right hippocampus (<t > =3.367), right parahippocampal gyrus (<t > =3.868), right amygdala (<t > =3.347), and right temporal pole (superior temporal gyrus; <t > =3.199). The complete results, with effect sizes and confidence intervals, are reported in Table 4A.

Panel B displays the spatial patterns of the BOLD signal and the thresholded effect-sizes for the final d-CAP sets for this network. In this case, 5 d-CAPs for HC and only 3 d-CAPs for CI were found. For each d-CAP, between-group differences in Cohen’s d maps at d ≥ ±0.80 are reported in panel C and Table 4B. Differences in coactivation (mainly with HC > CI) were found in several brain regions (Table 4B). Spatial similarities between d-CAPs are reported in panel D. For this network, high spatial similarities were found across the d-CAPs.

Figure 4 shows the Z-values resulting from the dual-regression procedure for the ECN. Panel A focuses on the statistical differences observed between the HC and CI groups in ICA analysis. Notably, a significant decrease in connectivity was discovered within the right superior temporal gyrus (<t > = − 3.177) and both sides of the middle temporal gyrus (<t > = − 3.140) in the HC group compared to the CI group. For comprehensive findings, including effect sizes and confidence intervals, please refer to Table 5A.

Panel B shows the spatial patterns of the BOLD signal and the thresholded effect sizes for the final d-CAP sets for this network. In this case, for both groups, we found 3 d-CAPs. Additionally, for each d-CAP, between-group differences in Cohen’s d maps at d ≥ ±0.80 are reported in panel C and Table 5B. Differences in coactivation were found in several brain regions. For d-CAP-3, we found only clusters with higher coactivation in CI than HC (Table 5B). Lastly, spatial similarities between d-CAPs are reported in panel D. High spatial similarity was found for d-CAP-1 between HC and CI (r > 0.80).

Figure 5 shows the Z-values resulting from the dual-regression procedure for the FPN. Panel A shows the statistical differences observed between the HC and CI groups in ICA analysis. Compared with CI, significant clusters with higher connectivity were observed in HC, mainly in the left posterior cingulate gyrus (<t > =3.055), left superior parietal gyrus (<t > =3.068), precuneus (<t > =3.141), and thalamus (<t > =3.439 and < t > =3.198, for left and right sides, respectively). A small cluster with higher connectivity in CI was found inside the left inferior frontal gyrus pars orbitalis (<t > = − 3.103). For comprehensive findings, including effect sizes and confidence intervals, please refer to Table 6A.

Panel B shows the spatial patterns of the BOLD signal and the thresholded effect sizes for the final d-CAP sets for this network. For this network, for both groups, we found 2 d-CAPs. Moreover, for each d-CAP, between-group differences in Cohen’s d maps at d ≥ ±0.80 are reported in panel C and Table 6A. Compared with the CI group, higher coactivation was found in HC inside several brain regions (Table 6B). Lastly, spatial similarities between d-CAPs are reported in panel D. High spatial similarity was found for d-CAP-1 between HC and CI (r > 0.80).

Switching probabilities were computed for each participant within the HC and CI groups. Box plots illustrating the switching probabilities associated with all networks are depicted in Figure 6, while the complete results can be found in Table 7. Notably, when comparing the CI group to the HC group, significantly lower switching probabilities were observed for the DMN (t = 2.239; FDR = 0.023), VN (t = 2.864; FDR = 0.007), and FPN (t = 2.645; FDR = 0.017). However, for the ECN, although a decrease in switching probabilities was observed in the HC group, the corresponding FDR-value was greater than 0.05 (t = 0.461; FDR = 0.764). The decreased switching probabilities in the CI group, particularly for the DMN, VN, and FPN networks, indicate an overall reduction in dynamic activity within these networks.

Table 8A presents the spatial consistency of each d-CAP associated with all networks. Across the analyzed networks, the mean spatial consistency for the first d-CAP was higher in the HC group compared to the CI group, which can indicate resting-state networks with reduced dynamics in the latter.

In Table 8B, the temporal fraction of the d-CAPs associated with each network is shown. Additionally, Table 8C provides the mean and SD of the temporal fraction for the 1st d-CAPs in each group for only networks with the same number of d-CAPs. Statistical differences were found in the DMN and FPN networks, where the HC group exhibited lower temporal fractions compared to the CI group, indicative of decreased network dynamics for the CI group.

We examined the correlations between switching probability and cognitive scores, as shown in Figure 7. Our analysis revealed statistical correlations, without correction for multiple comparisons, between MoCA scores and SP within the DMN (t = 2.372, p = 0.028, FDR = 0.084, ρ = 0.43). Similarly, a significant correlation was observed between the clock draw and SP within the ECN (t = 2.467, p = 0.036, FDR = 0.107, ρ = 0.656). It is important to note that both correlations did not survive FDR correction.