Dynamic functional connectivity changes in the triple networks and its association with cognitive impairment in hemodialysis patients

Abstract Introduction Cognitive impairment is common in hemodialysis (HD) patients; however, the underlying mechanisms have not been fully understood. The "triple‐network model" that consists of the salience network (SN), central executive network (CEN), and default mode network (DMN) has been suggested to play an important role in various cognitive functions. However, dynamic functional connectivity (FC) alterations within the triple networks have not been investigated in HD patients. Methods Sixty‐six HD patients and 66 healthy controls (HCs) were included in this study. The triple networks were identified using a group spatial independent component analysis, and dynamic FC was analyzed using a sliding window approach and k‐means clustering algorithm. Furthermore, we analyzed the relationships between altered dynamic FC parameters and clinical variables in HD patients. Results The intrinsic brain FC within the triple networks was clustered into four configuration states. Compared with HCs, HD patients spent more time in State 1, which was characterized by weak connections between the DMN and CEN and SN. HD patients showed lower number of transitions across different states than HCs. Moreover, the number of transitions and mean dwell time in State 1 were associated with cognitive performance in HD patients. Conclusion Our findings suggest abnormal dynamic FC properties within the triple networks in HD patients, which may provide new insights into the pathophysiological mechanisms of their cognitive deficits from the perspective of dynamic FC.


INTRODUCTION
Chronic kidney disease (CKD) is a substantial health problem worldwide, affecting 10−12% of the population (Levin et al., 2017). Endstage renal disease (ESRD) is the terminal stage of CKD, and usually requires maintenance hemodialysis (HD) to sustain life over the long term (Webster et al., 2017). Notably, cognitive impairment (CI) is extremely common in HD patients, across numerous domains, particularly in the domains of orientation and attention and executive function, and patients often experience multiple deficits simultaneously (Drew et al., 2017;O'Lone et al., 2016;van Zwieten et al., 2018). Cognitive deficits are associated with poorer quality of life, risk for hospitalization, and increased mortality in HD patients (Drew et al., 2015;O'Lone et al., 2016;van Zwieten et al., 2019). However, up to now, the underlying pathophysiological mechanisms of CI in HD patients have not been completely clarified.
Resting-state functional magnetic resonance imaging (Rs-fMRI)based functional connectivity (FC) analysis, which quantifies intrinsic functional brain organization (Biswal et al., 1995), has become a valuable and non-invasive tool to investigate the network connectivity basis of cognitive deficits in HD patients. For example, widespread weakening of cortical and subcortical network connectivity has been identified in HD patients (G. Zheng et al., 2014). The highly influential "triple-network model," which was proposed by Menon (2011), provides a common framework for understanding the dysfunction in core neurocognitive networks across multiple psychiatric and neurological disorders. This model integrates three key intrinsic brain networksthe default mode network (DMN), central executive network (CEN), and the salience network (SN) into a single cohesive model that serves as the basis for normal behavior and cognition. Evidences from previous studies have demonstrated impaired interactions of the SNcentered "triple-network model" in various brain diseases, such as Alzheimer's disease (C. Li et al., 2019) and major depression disorder (H. Zheng et al., 2015). Our recent study of large-scale network analysis using graph theory-based approaches have demonstrated disrupted topological organizations of brain functional networks in patients with ESRD (Yue et al., 2021); however, we did not focus on cross-network interactions of the triple networks. More importantly, this previous whole-brain FC study on ESRD patients assumed that the FC was static within the entire RS-fMRI scan, and did not consider the important dynamic aspect over time.
In recent years, more and more studies have focused on the dynamic changes of intrinsic FC across large-scale brain functional networks.
Dynamic functional network connectivity has been identified and explored in neuroimaging studies of healthy subjects, which was associated with higher-order cognitions (Kucyi et al., 2017;Shine et al., 2019;Soreq et al., 2019). Moreover, abnormal dynamic brain functional network connectivity has also been demonstrated in clinical patients, such as those with bipolar disorder , Parkinson's disease (Kim et al., 2017), and autism spectrum disorder (Rashid et al., 2018 using a spatial group independent component analysis (ICA), a sliding window approach and k-means clustering algorithm, and explored the underlying neural mechanisms underlying CI in HD patients.

Participants
This prospective study was approved by the local Ethics Committee and followed the ethical guidelines of the Declaration of Helsinki; written informed consent was acquired from each subject before inclusion. A subset of our prior study sample

Laboratory examinations
Blood and urine tests were performed for all HD patients within 3 days before MRI examinations. The hemoglobin, serum creatinine, and blood urea nitrogen levels were recorded. No biochemistry tests were performed in HCs.

MRI data acquisition
MR imaging data were acquired using a 3T GE Discovery MR750 scanner (General Electric Healthcare, Milwaukee, WI) with a 16channel head coil. All subjects were instructed to be quiet, keep their eyes closed but stay awake, and try not to think about anything.
First, conventional T 1 -weighted, T 2 -weighted, and T 2 -weighted fluidattenuated inversion recovery imaging sequences were acquired to exclude organic brain lesions. Then, the Rs-fMRI data were obtained using an echo-planar imaging (EPI) sequence with the following param-

Image preprocessing
The Data Processing & Analysis for Brain Imaging toolbox (DPABI, version 4.1; http://rfmri.org/dpabi) (Yan et al., 2016) was used for pipeline data analysis of Rs-fMRI. Briefly, the first 10 volumes of the Rs-fMRI dataset of each subject were discarded to allow for MR signal equilibrium; thus, the remaining 170 volumes were used for further analyses. Then, the remaining Rs-fMRI data were corrected for the temporal differences between slices, and were realigned to the first volume for the correction of head motion. Next, the corrected Rs-fMRI data were spatial normalized to the standard Montreal Neurological by using diffeomorphic anatomical registration through exponentiated Lie algebra (DARTEL) (Ashburner, 2007). Finally, the normalized Rs-fMRI images were spatially smoothed with a 6-mm full-width at halfmaximum Gaussian kernel.

Group ICA and identification of intrinsic brain networks
To decompose the preprocessed Rs-fMRI data into different independent components (ICs), spatial group ICA was performed using the Group ICA of fMRI Toolbox (GIFT, version 4.0b; http://icatb. sourceforge.net). The ICA included three steps: (1) data reduction; (2) ICs decomposition; and (3) back reconstruction. Specifically, we first used a two-step principal component analysis (PCA) to reduce the data into 37 ICs. This averaged IC number was determined using the minimum description length (MDL) criteria (Y. O. Li et al., 2007) and was used for each subject for ICA decomposition. Second, ICs estimation was performed using the Infomax algorithm (Bell & Sejnowski, 1995).
In this step, the spatial map and the time course of blood oxygenation level-dependent (BOLD) signal were generated for each IC. This step was repeated 100 times using the ICASSO algorithm to identify the most stable and reliable components (Himberg et al., 2004).
Finally, subject-specific spatial maps and time courses for each IC were obtained using group ICA back reconstruction algorithm (Calhoun et al., 2001), and the subject-specific maps were converted to z-scores.
Subsequently, all ICs were evaluated based on the group IC maps by following the criteria suggested by previous studies (Allen et al., 2014;Cordes et al., 2000;Kim et al., 2017): (1)

Dynamic FC computation
A sliding window approach was applied for the computation of dynamic FC using the Temporal dFNC toolbox in GIFT. In each window, the Pearson's correlation coefficients of time courses of each pair of the five resting-state networks were calculated, resulting in a 5 × 5 correlation matrix. We used a tapered window, created by convolving a rectangle (width = 30 TRs, i.e., 60 s) with a Gaussian (σ = 3 TRs), and the window was shifted with step size of 1 TR each time, resulting in 140 consecutive windows across the entire Rs-fMRI scan. The window length of 30 TRs was chosen based on a previous study (Yu et al., 2015), which suggests that this window size may provide a good trade-off between the quality of the FC estimate and the temporal resolution.
To characterize the full covariance matrix, we estimated covariance from the regularized precision matrix or the inverse covariance matrix (Smith et al., 2011). To promote sparsity in estimation, a penalty on the L1 norm (i.e., the sum of the absolute values of the elements of the precision matrix) was imposed in the graphic LASSO framework with 10 repetitions (Friedman et al., 2008). Values in the resulting FC matrices were converted to z-scores using Fisher's r-to-z transformation to improve the normality of the distribution of Pearson's r and were then residualized with known confounding variables, such as age and gender. The resulting individuals' 140 FC matrices represent the dynamic changes of FC during the whole Rs-fMRI scan period and were used for further FC state analysis.

Clustering analysis
We applied a k-means clustering algorithm on all the windowed 5 × 5 FC matrices for all subjects to estimate reoccurring FC patterns (states) with a random initialization of the centroid positions. The kmeans clustering algorithm was repeated 500 times to increase the chance of escaping the local minima. The optimal number of clusters was estimated to be four (k = 4) using the elbow criterion (Allen et al., 2014). The correlation distance method was chosen because it is more sensitive to the dynamic FC pattern irrespective of magnitude (Rashid et al., 2018). Subsequently, these resulting centroids were used as starting points to cluster the dynamic FC windows for all the subjects.

FC strength
The subject-specific centroid of each state was computed by calculating the median value of each FC matrix for that state. To determine between-group differences in connectivity strength, we calculated the group-specific centroids of the four states by averaging subjectspecific centroids of both HD patients and HCs, respectively.

Statistical analysis
2.9.1 Group differences in demographic and clinical data A chi-square test was used to determine gender-ratio difference between the HD group and HC group, and an independent two-sample t-test was used to compare the clinical variables and neuropsychological test results between the two groups. Statistical analysis was performed using software (SPSS, version 21.0; IBM Corp, Armonk, NY).
Statistical significance was defined as p < .05.

Group differences in dynamic FC
Between-group differences in temporal properties of dynamic FC states strength were tested using a Mann-Whitney U-test in SPSS software, and between-group differences in FC strength of each state were tested using an independent two-sample t-test in GRETNA software (http://www.nitrc.org/projects/gretna). The significance level was set to false-discovery rate-corrected p < .05.

Relationship between altered dynamic FC properties and clinical variables
A partial correlation analysis was used to estimate the relationships between those dynamic FC parameters showing significant betweengroup differences and the clinical variables in the HD group (p < .05, uncorrected), with age, sex and educational level as covariables.
The clinical variables comprised the neuropsychological test results, duration of HD, hemoglobin level, serum creatinine level, and serum urea level.

Demographic and clinical characteristics
The demographic and clinical characteristics are summarized in Table 1.
There were no significant differences in age (p = .923), sex (p = .486), and education level (p = .295) between the two groups. The MoCA (p < .001) and SDMT (p < .001) scores of the HD group were significantly lower than those of the HC group, and HD patients spent longer time in the completion of TMT-A than HCs (p < .001). Moreover, the HAMD-17 score of the HD group was significantly higher than that of the HC group (p < .001).

Intrinsic FC networks
As shown in Figure 1, the five ICs, aDMN (IC12), pDMN (IC8), RCEN (IC27), LCEN (IC20), and SN (IC8), were selected from the 37 ICs. The spatial maps of those ICs were similar to those found in previous studies Wang et al., 2020) and covered most of the grey matter.

Dynamic FC state analysis
Using the k-means clustering algorithm, we identified four highly structured FC states that recurred throughout individual scans and across subjects ( Figure 2 are much weaker than in other states.
The group-specific medians for each state are shown in Figure 3.
Notably, not all subjects had the windows assigned to each state, which contributed to changes of the number of subject-specific matrices in different states (see Figure 3 for subject counts of each state). No significant between-group differences in FC strength were found in each state after FDR correction.
Between-group differences in temporal properties of FC states are shown in Table 2

DISCUSSION
By using Rs-fMRI in combination with clustering algorithm and FC state analysis, the present work is the first study to analyze the dynamic FC alterations in HD patients based on the "triple-network model," focusing on the temporal properties and FC strength of dynamic FC states.
Three main findings were found in our study. First, the dynamic FC TA B L E 2 Between-group differences in temporal properties of dynamic functional connectivity states HCs dynamically switch between different FC states and are therefore probably faster in recruiting necessary resources in the face of changing task demands (Yu et al., 2015). Thus, a lower total number of transitions across states in HD patients may suggest a slower speed to recruit necessary resources when faced with different cognitive tasks, which may lead to the decline in cognitive performance. Indeed, recent studies have demonstrated the association between network flexibility and cognitive task performance (Garrett et al., 2013;Madhyastha et al., 2015;Spreng & Schacter, 2012;Thompson et al., 2013).  (Kim et al., 2017). In the present study, we used a typical TR (TR = 2s) to acquire Rs-fMRI images. Previous studies have demonstrated the reliability of this typical TR in sampling the dynamics of low-frequency fluctuation (Allen et al., 2018;Liao et al., 2014). However, to increase the estimation power of FC matrices calculated within small windows in the sliding window approach, it would be beneficial to use a rapid Rs-fMRI acquisition, such as the simultaneous multiband multi-slice EPI acquisition (Moeller et al., 2010), that could sample more dense time series. Third, although we found an association between the number of transitions and the SDMT score, one might have expected the number of transitions to be associated also with the completion time of TMT-A given its relevance to processing speed.
Actually, our partial correlation analysis revealed a negative correlation between these two measures; however, this result was not significant, which might be attributed to the relatively small sample size.

F I G U R E 4
Between-group comparison of temporal properties and the partial correlation analysis results. The temporal properties (a-c) of functional connectivity state analysis are shown using violin plots. Transverse solid and dotted lines represent medians and quartiles, respectively. Asterisks indicate a significant between-group difference (p < .05, FDR corrected). For hemodialysis patients, (d) the number of transitions positively correlated with the SDMT score , (e) and the mean dwell time in State 1 positively correlated with the completion time of TMT-A. Abbreviations: SDMT, Symbol Digit Modalities Test; TMT-A, Trail Making Test A Finally, the sample size of our study was relatively small, which may affect the statistical power. Future studies with large sample sizes are needed to verify the reproductivity of our findings.

CONCLUSIONS
In summary, the present study investigated the dynamic FC properties in the triple networks in HD patients using Rs-fMRI data in combination of ICA, sliding windows approach, and k-means clustering analysis.
Our study demonstrated altered dynamic FC properties in the triple networks in HD patients compared with HCs, including increased time in the weakly connected State 1 and decreased number of transitions.
In addition, the altered dynamic FC properties were associated with cognitive performance in HD patients. Thus, this study provides new insights into the pathophysiological mechanisms of CI in HD patients from the perspective of dynamic FC.

This study was supported by the Henan Province Medical Science and
Technology Research Project (2018020349).

CONFLICT OF INTEREST
The authors declare no conflict of interest.

AUTHOR CONTRIBUTIONS
Baolin Wu conceived and designed the study, supervised the conduct of the study, and took responsibility for the paper. Xuekun Li, Zheng Yue, and Jipeng Ren were responsible for data acquisition. Jianghui Cao and Guangzhi Liu analyzed the data. Xuekun Li assisted with the literature review. Jianghui Cao and Guangzhi Liu drafted the initial manuscript.
Baolin Wu and Wei Zhu reviewed and revised the manuscript. All authors read and approved the final manuscript.

DATA AVAILABILITY STATEMENT
The data that support the findings of this study are available from the corresponding author upon reasonable request.

PEER REVIEW
The peer review history for this article is available at https://publons. com/publon/10.1002/brb3.2314