ORIGINAL RESEARCH article

Front. Neurosci., 10 August 2022

Sec. Brain Imaging Methods

Volume 16 - 2022 | https://doi.org/10.3389/fnins.2022.967760

Aberrant brain gray matter and functional networks topology in end stage renal disease patients undergoing maintenance hemodialysis with cognitive impairment

  • 1. Department of Radiology, The Affiliated Changzhou No. 2 People’s Hospital of Nanjing Medical University, Changzhou, China

  • 2. GE Healthcare, MR Research China, Beijing, China

  • 3. Department of Nephrology, The Affiliated Changzhou No. 2 People’s Hospital of Nanjing Medical University, Changzhou, China

  • 4. School of Computer Science and Artificial Intelligence, Changzhou University, Changzhou, China

Abstract

Purpose:

To characterize the topological properties of gray matter (GM) and functional networks in end-stage renal disease (ESRD) patients undergoing maintenance hemodialysis to provide insights into the underlying mechanisms of cognitive impairment.

Materials and methods:

In total, 45 patients and 37 healthy controls were prospectively enrolled in this study. All subjects completed resting-state functional magnetic resonance imaging (rs-fMRI) and diffusion kurtosis imaging (DKI) examinations and a Montreal cognitive assessment scale (MoCA) test. Differences in the properties of GM and functional networks were analyzed, and the relationship between brain properties and MoCA scores was assessed. Cognitive function was predicted based on functional networks by applying the least squares support vector regression machine (LSSVRM) and the whale optimization algorithm (WOA).

Results:

We observed disrupted topological organizations of both functional and GM networks in ESRD patients, as indicated by significantly decreased global measures. Specifically, ESRD patients had impaired nodal efficiency and degree centrality, predominantly within the default mode network, limbic system, frontal lobe, temporal lobe, and occipital lobe. Interestingly, the involved regions were distributed laterally. Furthermore, the MoCA scores significantly correlated with decreased standardized clustering coefficient (γ), standardized characteristic path length (λ), and nodal efficiency of the right insula and the right superior temporal gyrus. Finally, optimized LSSVRM could predict the cognitive scores of ESRD patients with great accuracy.

Conclusion:

Disruption of brain networks may account for the progression of cognitive dysfunction in ESRD patients. Implementation of prediction models based on neuroimaging metrics may provide more objective information to promote early diagnosis and intervention.

Introduction

End-stage renal disease (ESRD) is the final stage of chronic kidney disease (CKD). Treatment options for ESRD include continuous hemodialysis, peritoneal dialysis, and kidney transplantation. ESRD patients are at high risk of developing cognitive impairment (CI), especially in patients who have received continuous hemodialysis, with a prevalence of 30–60% (; ). Impaired domains include overall cognition, executive function, memory, motor, and attention (). CI in ESRD patients is associated with negative outcomes, including non-adherence to drug treatment and increased rate of suicide (; ). However, the underlying neuropathology of CI in ESRD patients remains largely unknown. Therefore, investigating the neuropathological alterations leading to CI would help to understand the potential mechanisms contributing to CI, which would be beneficial in preparing treatment plans for ESRD patients.

Magnetic resonance imaging (MRI) is a non-invasive technique that can provide structural and functional information of the brain. It has been widely used to study the neurological changes in CI-related diseases, including Parkinson’s disease (), silent cerebral infarction (), Alzheimer’s disease (AD) (), and multiple sclerosis (). Voxel-based morphometry (VBM) is a valuable tool for assessing brain volume changes on a voxel-by-voxel basis (). Recently, reported that a predominant decrease in gray matter (GM) volume is associated with CI in ESRD patients by using VBM based on structural MRI technologies. Diffusion tensor imaging (DTI) is a valuable MRI technique that can estimate the microstructure of tissues by probing the diffusion process of water molecules (). Through the use of DTI, previous studies found that disruption of white matter (WM) integrity correlated with impaired kidney function and CI in ESRD patients (; ; ). Functional imaging techniques, such as resting-state functional magnetic resonance imaging (rs-fMRI) and arterial spin labeling (), have been utilized to detect brain functional alterations in ESRD. Recently, investigated disrupted neurovascular coupling in ESRD patients undergoing hemodialysis and revealed it to be a potential neural mechanism for CI. Moreover, rs-fMRI can detect low-frequency (0.01–0.08 Hz) fluctuations in blood-oxygen-level-dependent (BOLD) signals and can be used to investigate spontaneous neural activity at rest (). In hemodialysis patients, impaired functional integrity of extensive brain regions has been measured by using multiple analytical methods based on rs-fMRI, including regional homogeneity (ReHo), amplitude of low frequency fluctuation (ALFF), and functional connectivity (FC) (; ; ; ). Although the above results were inconsistent due to the diverse population cohorts and methodologies used, a correlation between functional abnormalities and cognitive dysfunction was indicated.

Complex structural and functional brain networks can provide a physiological basis for information processing among neural elements and mental representation (; ). Diffusion MRI and fMRI are the most extensively used non-invasive imaging methods for reconstructing structural and functional networks, with network nodes representing brain regions and network edges representing structural or functional connectivity (). Currently, graph theory can be used to evaluate the architecture, development, and evolution of brain networks systematically and quantitatively by quantitatively analyzing topological properties (; ). Recent studies have used graph theory to detect the complexity of brain networks in various neurological diseases, including schizophrenia, AD, depressive disorder, acute brainstem ischemic stroke, and epilepsy (; ; ; ; ). In these disorders, the emergence of clinical symptoms or functional impairment was related to the disrupted integration of spatially distributed brain regions in structural and functional networks (). Similarly, the use of graph theory with rs-fMRI and DTI has identified brain aberrations in functional and WM networks correlating with cognitive function in ESRD patients (; ). However, no studies to date have investigated the relationship between topological characteristics of GM networks and cognitive function in ESRD patients.

Diffusion kurtosis imaging (DKI), an extension of the DTI model, was developed to sensitively quantify non-Gaussian water diffusions in a voxel (). Thus, DKI provided a useful way to investigate abnormalities in both GM and WM, where DTI was less capable. Currently, DKI has been used as a practical clinical application in neuroscience research, including normal brain tissues, brain trauma, brain tumors, AD, schizophrenia, and Parkinson’s disease (; ; ; ; ). These studies supported the possibility that DKI may be a sensitive imaging biomarker in many neurological diseases. However, these studies only investigated kurtosis values in regional brain areas, ignoring the coordination of heterogeneous DKI properties between different brain regions. A recent study focused on changes in both GM and WM networks based on DKI in AD found that DKI network damage was related to cognitive performance (). However, few studies have applied DKI to investigate CI in ESRD patients. More specifically, no studies have applied DKI to assess GM networks in ESRD patients.

In this study, we used DKI and rs-fMRI to construct GM functional networks to investigate potential aberrant mechanisms leading to CI in ESRD patients. Further, we tentatively predicted the cognitive function of ESRD patients using topological properties as features based on the individual level. The least squares support vector regression machine (LSSVRM) was used to build a prediction model, and the whale optimization algorithm (WOA) was used to optimize model parameters (). In conclusion, our study attempted to improve the possibility of early diagnosis and neuroprotective treatments for ESRD patients by using predictive models based on neuroimaging techniques.

Materials and methods

Participants

This study was approved by the Ethics Committee of The Affiliated Changzhou No. 2 People’s Hospital of Nanjing Medical University (Number: KY039-01). All participants provided informed consent to participate in the study. All subjects were right-handed and fully capable of completing the Montreal cognitive assessment scale (MoCA) test independently.

Between February 2020 and December 2021, 45 ESRD patients (22 male and 23 female, mean age 49.56 ± 8.02 years) from our hospital were prospectively recruited into our patient group. Inclusion criteria for the patient group included: (1) clinically diagnosed ESRD (estimated glomerular filtration rate less than 15 mL/min/1.73 m2); (2) receiving regular maintenance hemodialysis lasting longer than 3 months (three times a week at the hemodialysis center); (3) age between 30 and 65 years; and (4) MoCA score less than 26. Exclusion criteria included: (1) history of other neuropsychiatric disorders; (2) history of head trauma, intracranial tumors, or cerebral infarction; (3) renal transplant history; (4) drug or alcohol abuse history; (5) contraindications to MRI, such as claustrophobia, dental fixtures, and other exogenous objects in the head; and (6) obvious head motion artifact.

In addition, 37 healthy controls (HCs) (19 male and 18 female, mean age 47.30 ± 7.20 years) without renal disease and other known disorders and with MoCA scores higher than 26 were enrolled at the same time as the control group. Both groups were matched based on age, gender, and education years. The exclusion criteria for HCs were the same as that for the patient group.

Neuropsychological assessment

MoCA testing was administered to measure the overall cognitive status (out of a possible 30 points), with patients scoring less than 26 diagnosed with CI (). MoCA testing was administered by a well-trained clinical neuropsychologist within 2 h prior to MR scanning.

Laboratory tests

Laboratory tests for all ESRD patients were conducted within 24 h prior to MR scanning and included white and red cell counts, hemoglobin, hematocrit, fasting glucose, urea nitrogen, creatinine, uric acid, cholesterol, triglyceride, and calcium measurements. Laboratory tests were not performed for patients in the HC group.

Research framework

Research framework is shown in Figure 1.

FIGURE 1

Magnetic resonance imaging data acquisition

Magnetic resonance imaging data were acquired using a 3.0T magnetic resonance scanner (Discovery MR750W, General Electric Medical Systems, United States, Milwaukee, WI), equipped with a standard 32-channel head and spine combined coil. All participants were asked to stay still and awake and to keep their eyes closed during the entirety of scanning. Earplugs were used to alleviate the noise from the MR scanner. Foam pads were added on both sides of the head to reduce head motion. High-resolution anatomic T1-weighted images were acquired with the three-dimensional brain volume imaging (3D-BRAVO) sequence [parameters: 152 slices; slice thickness = 1.2 mm (no gap); repetition (TR) = 8.2 ms; echo time (TE) = 3.2 ms; flip angle (FA) = 12°; matrix = 256 × 256; field of view (FOV) = 240 mm × 240 mm; whole scanning time = 3 min 57 s]. rs-fMRI data were acquired with the gradient-recalled echo-planar imaging (GRE-EPI) sequence (parameters: 33 slices; 240 time-points; slice thickness = 4 mm; TR = 2,000 ms; TE = 40 ms; FA = 90°; matrix = 64 × 64; FOV = 240 mm × 240 mm; whole scanning time = 8 min 12 s). DKI data were acquired using a single-shot echo-planar imaging (SS-EPI) sequence, with 3b-values (b = 0, 1,000, 2,000 s/mm2) along 30 diffusion gradient directions (parameters: NEX = 2, slice thickness = 3.6 mm (no gap); TR = 6,500 ms; TE = 95.8 ms; matrix = 128 × 128; FOV = 240 mm × 240 mm; whole scanning time = 14 min 43 s).

Image preprocessing

Diffusion kurtosis imaging data

Diffusion kurtosis imaging data were preprocessed using the FMRIB Software Library (FSL),1 Diffusion Kurtosis Estimator (DKE),2 and Statistical Parametric Mapping 8 (SPM8)3 software. First, DKI data in DICOM format were converted into 4DNIFTI format. Data were corrected for head movement, eddy, and gradient distortion with b = 0 image as a reference using FSL. Then, DKI parameters including mean kurtosis (MK), axial kurtosis (AK), radial kurtosis (RK), and kurtosis anisotropy (KA) were extracted using DKE with the quadratic programming-based (CLLS-QP) algorithm (). Next, after registering each subject’s 3D-T1w anatomic images with the standard Montreal Neurologic Institute (MNI) template, b = 0 images were registered with the standardized 3D-T1w anatomic images using SPM8 with a non-linear co-registration technique. The resulting transformation matrix was used for normalizing all DKI parameter maps. Finally, normalization and smoothing (FWHM = 6 mm) for DKI parameters were conducted using SPM8.

Resting-state functional magnetic resonance imaging data

Resting-state functional magnetic resonance imaging data preprocessing was performed with the data processing assistant for resting-state fMRI (DPARSF-V2.3)4 based on MATLAB 2013. First, after converting data from the DICOM format to the NIFTI format, the first 10 timepoints were discarded to allow for steady-state longitudinal magnetization. Slice timing and head motion correction were conducted for the remaining 230 timepoints collected (subjects were excluded if their head movement was more than 3 mm or 3°). Next, spatial normalization was performed based on unified segmentation to the structural T1 image, and the functional image was warped into the standard space of the (MNI) template (resampling voxel size = 3 mm × 3 mm × 3 mm). Then, spatial smoothing in reference to an isotropic Gaussian kernel (full width at half maximum [FWHW] = 6 mm), detrending, and temporal filtering (bandpass, 0.01–0.08 Hz) were performed. Finally, nuisance covariates including cerebrospinal fluid signals, WM signals, and Friston-24 head motion parameters were regressed out of the data.

Network construction

Gray matter network construction

Gray matter networks were constructed at the group level based on each DKI parameter measurement. The structural connections of the GM network were defined as statistical correlations between pairs of kurtosis values from brain nodes. Nodes were defined as the parcellation of a whole-brain conducted according to the Automated Anatomical Labeling (AAL) atlas, which divided the brain into 90 regions including 78 cortical regions and 12 subcortical regions. These regions were regarded as the nodes of the GM network. For edge definition, regional kurtosis values for every region were extracted to calculate the Pearson’s correlation coefficients across individuals for each pair of brain regions to obtain the correlation matrix (90 × 90) of each group (). Then, a parameter matrix for each DKI metric was assigned to the HC and ESRD patient groups (MK, AK, RK, and KA).

Functional network construction

Functional networks for each participant were constructed using the GRETNA toolbox.5 Nodes were defined (similar to the GM network) according to the AAL atlas (90 regions). Edge definitions were calculated using the mean time series for every region to calculate the Pearson’s correlation coefficients for each pair of regions, which were used to obtain the correlation matrix (90 × 90). Then, Fisher’s Z transformation was performed to increase the normality of the matrix.

Network analysis

Network threshold selection

Using the matrix sparsity as the threshold, the matrix was binarized. The whole range of sparsity thresholds of the GM network was set to 0.06–0.40 with an interval of 0.01 (the sparsity threshold was selected to ensure that all resultant networks have the same number of nodes and edges) (). The whole range of sparsity thresholds of the functional network was set to 0.1–0.4, with an interval of 0.01 (the values of small-worldness of all participants were checked more than once in order to avoid the selection of a threshold range too wide to produce connected nodes and networks with small-worldness features) ().

Network metrics

For each sparsity threshold, global and nodal measures were calculated. Global measures included global efficiency (Eg), local efficiency (Eloc), mean clustering coefficient (Cp), characteristic path length (Lp), standardized clustering coefficient (γ), standardized characteristic path length (λ), and small-world properties (σ). Nodal measures included nodal efficiency (Ne) and degree centrality (Dc). Areas under the curve (AUCs) of the topological parameters of the functional network were calculated within the whole sparsity threshold.

Statistical analysis

Group differences between demographic data, clinical characteristics, and MoCA scores

Demographics, clinical data, and MoCA scores were analyzed using specific software (SPSS version 25.0; SPSS, Chicago, Illinois). Chi-squared test, two-sample independent Student’s t-test, and Mann-Whitney U test were used to compare gender-based differences and quantitative data between the groups. A statistical significance level was set at P < 0.05.

Group differences between all imaging parameters

A two-sample t-test was performed based on GRETNA to detect differences in network measures between the two groups with age, gender, and education years used as covariates. We set the statistical threshold with false discovery rate criterion (FDR)- corrected P < 0.05.

Correlation analysis

Pearson’s correlation analysis was performed based on GRETNA to detect the relationships between significant topological parameters of the functional network and MoCA scores in ESRD patients with age, gender, and education years used as covariates (P < 0.05 and corrected by FDR).

Prediction model construction

Least squares support vector regression machine is an improvement of support vector regression machine. To improve prediction efficiency and accuracy, LSSVRM changed the inequality constraint and the solution of quadratic programming problem into equality constraint and the solution of linear equations, respectively (). Then, to improve operating efficiency, we optimized selection strategy of kernel function parameters by introducing WOA into LSSVRM ().

Optimized LSSVRM was applied to predict cognitive function based on functional networks. Functional network measures (global and nodal) significantly correlated (P < 0.05 and corrected by FDR) with MoCA scores were selected as features. The selected features were fused to build LSSVRM, using two methods: [a] only global measures were fused; [b] both global and nodal measures were fused. Then, selected measures with corresponding MoCA scores of 45 ESRD patients were used as a data set by the leave-one-out method. For our data, leave-one-out method was optimized empirically. The mean square error (MSE), root mean square error (RMSE), mean absolute error (MAE), and mean absolute percentage error (MAPE) were chosen as testing standards to assess the accuracy of models. The lower the MSE, RMSE, MAE, or MAPE, the better the model’s prediction accuracy.

Results

Demographic, clinical, and neuropsychological results

No significant differences were observed in the age, gender, and education years between the participants in the two groups (P > 0.05). When compared to the HC group, patients in the ESRD group had significantly lower MoCA scores (P < 0.001) (Table 1).

TABLE 1

VariableESRD (n = 45)HC (n = 37)Statistic valueP-value
Demographic
Age, years49.56 ± 8.0247.84 ± 6.71T = 1.038P = 0.303
Gender, male (%)22 (49)19 (51)x2 0.032P = 0.858
Education, years (25%, 75%)12 (9, 15)12 (8, 15)Z = -0.932P = 0.351
Neuropsychological scores
MoCA scores (25%, 75%)21 (20, 23)27 (26, 28)Z = -7.807P < 0.001*
Laboratory data
Systolic pressure, mmHg150.38 ± 27.00---
Diastolic pressure, mmHg91.44 ± 16.17---
White cell count, 10^9/L5.90 ± 2.17---
Red cell count, 10^12/L4.37 ± 4.21---
Hemoglobin, g/L107.60 ± 18.20---
Hematocrit, %37.17 ± 26.60---
Fasting glucose, mmol/L5.42 ± 1.89---
Urea nitrogen, mmol/L19.83 ± 8.39---
Creatinine, μmol/L867.49 ± 411.83---
Uric acid, μmol/L343.73 ± 134.11---
Cholesterol, mmol/L4.17 ± 1.57---
Triglyceride, mmol/L1.66 ± 0.89---
Calcium, mmol/L2.25 ± 0.21---

Demographic and clinical characteristics and neuropsychological test scores.

*Significantly difference.

x2, Chi-square test; T, two-sample t-test; Z, Mann-Whitney U test;MoCA, Montreal cognitive assessment scale.

Group differences of network measures

Gray matter network

The images of the group-level interregional correlation matrices using DKI metrics of AK, MK, RK, and KA are shown in Figure 2, and the between-group differences of global topological properties are presented in Table 2 and shown in Figure 2. The small-world organization of GM networks with AK, MK, RK, and KA metrics in the ESRD and HC groups was identified over a wide range of sparsity (0.06–0.4). The small-worldness values (σ = γ/λ) were larger than 1. However, when compared with the HC group, decreased topological properties of the GM network were found in ESRD patients. Decreased Cp was found in the GM network with the AK metric; decreased γ and σ were found in the GM network with the MK metric; decreased γ was found in the GM network with the RK metric; and decreased γ, σ, and Eloc were found in the GM network with the KA metric (P < 0.05; corrected by FDR).

FIGURE 2

TABLE 2

Structural networkParametersESRDHCt valueP-value
AKCp0.687 ± 0.0810.722 ± 0.046-2.1940.028
MKγ1.960 ± 0.4182.866 ± 0.945-5.190<0.001
σ1.596 ± 0.1312.240 ± 0.310-6.311<0.001
RKγ2.183 ± 0.8842.624 ± 1.046-3.749<0.001
KAγ2.519 ± 0.9633.423 ± 1.788-2.6320.010
σ1.790 ± 0.2772.459 ± 0.660-5.532<0.001
Eloc0.811 ± 0.0780.848 ± 0.039-2.5690.011

Significant between-group differences of topological parameters of gray matter (GM) network.

Data are expressed as mean ± SD. Corrected by false discovery rate criterion (FDR) and set at p < 0.05.

AK, axial kurtosis; MK, mean kurtosis; RK, radial kurtosis; KA, kurtosis anisotropy; Cp, mean clustering coefficient; γ, standardized clustering coefficient; λ, standardized characteristic path length; σ, small-world properties; Eloc, local efficiency.

Functional network

The between-group differences of topological properties are shown in Table 3 and Figure 3. Global measures: both the ESRD and HC groups demonstrated economical small-world properties (γ greater than 1, λ approximately 1, and σ greater than 1). However, in ESRD patients, significantly decreased γ and σ were observed in the functional network (P < 0.05; corrected by FDR). There were no significant differences in Eg, Eloc, Cp, and Lp between the two groups (P > 0.05; corrected by FDR). Nodal measures: regions with significantly changed Ne and Dc (9 and 15 regions, respectively) were identified in the ESRD group and distributed laterally (P < 0.05; corrected by FDR), which were mainly involved in the default mode network, limbic system, frontal lobe, temporal lobe, and occipital lobe. Specifically, all right regions showed significantly decreased Ne and Dc (right-lateralized), and most left regions showed increased Ne and Dc (left-lateralized), with the left superior temporal gyrus having decreased Ne and Dc.

TABLE 3

Functional networkParametersAALESRDHCt valueP-value
Global measuresγ0.633 ± 0.0640.676 ± 0.048-3.2050.002
σ0.579 ± 0.0620.617 ± 0.048-3.0150.003
Nodal measuresNodal efficiency
Frontal_Mid_L70.180 ± 0.0130.173 ± 0.0122.2720.009
Calcarine_L430.168 ± 0.0150.162 ± 0.0102.1550.012
Lingual_L470.169 ± 0.0140.161 ± 0.0162.3220.018
Temporal_Sup_L810.175 ± 0.0160.182 ± 0.015-2.1450.012
Rolandic_Oper_R180.172 ± 0.0150.179 ± 0.012-2.3920.006
Insula_R300.175 ± 0.0160.182 ± 0.013-2.4490.004
Fusiform_R560.158 ± 0.0140.168 ± 0.013-3.283<0.001
Parietal_Sup_R600.163 ± 0.0170.171 ± 0.013-2.3320.004
Temporal_Sup_R820.171 ± 0.0160.179 ± 0.013-2.4900.004
Degree centrality
Frontal_Mid_L77.886 ± 2.1436.936 ± 1.9292.0690.014
Frontal_Mid_Orb_L97.483 ± 1.8416.169 ± 1.6903.036<0.001
Frontal_Sup_Medial_L237.113 ± 1.9246.243 ± 1.2312.2200.003
Calcarine_L436.579 ± 1.9205.767 ± 1.6702.0650.019
Lingual_L476.627 ± 1.9015.794 ± 1.8841.9970.022
Temporal_Sup_L817.127 ± 1.9077.917 ± 1.781-2.1790.020
Frontal_Inf_Tri_R145.991 ± 1.7526.737 ± 1.830-2.0370.021
Rolandic_Oper_R186.705 ± 1.6167.607 ± 1.516-2.6010.003
Insula_R307.110 ± 1.7788.031 ± 1.535-2.5180.004
Amygdala_R426.427 ± 1.5517.266 ± 1.894-2.2370.011
Fusiform_R565.100 ± 1.4806.239 ± 1.570-3.254<0.001
Parietal_Sup_R605.659 ± 1.9416.572 ± 1.476-2.2460.005
Putamen_R746.755 ± 1.7657.589 ± 1.425-2.3610.004
Heschl_R805.811 ± 1.5626.745 ± 1.719-2.3960.003
Temporal_Sup_R826.728 ± 1.7347.787 ± 1.373-2.8490.001

Significant between-group differences of topological parameters of functional networks.

Data are expressed as mean ± SD. Corrected by false discovery rate criterion (FDR) and set at p < 0.05.

γ, standardized clustering coefficient; σ, small-world properties; AAL, automated anatomical labeling.

FIGURE 3

Correlation analysis

In the ESRD group, decreased γ and σ of the functional network positively correlated with MoCA scores (r = 0.42, r = 0.44, respectively; P < 0.05, corrected by FDR). Further, a decreased Ne of the functional network in the right insula and the right superior temporal gyrus was positively correlated with MoCA scores (r = 0.34, r = 0.33, respectively; P < 0.05, corrected by FDR) (Figure 4).

FIGURE 4

Prediction results

Optimized LSSVRM based on functional networks predicted the cognitive function of ESRD patients with great accuracy. When using selected global measures as features, MSE, RMSE, MAE, and MAPE between the actual scores and predicted scores were 0.85, 0.92, 0.84, and 4.05%, respectively, with an R-squared (R2) of 0.69. When both selected global and nodal measures were used as features, the MSE, RMSE, MAE, and MAPE between the actual scores and predicted scores were 0.77, 0.88, 0.78, and 3.80%, respectively, with an R2 value of 0.65 (Figure 4).

Discussion

The current study investigated the changes of GM and functional network topology in ESRD patients based on DKI and rs-fMRI technologies. The changed network topology was further used to predict ESRD patient cognitive function. Our results showed the following: (1) both the GM and functional networks presented small-world characteristics in ESRD patients and HC controls; (2) both the GM and functional networks tended to be less optimized in ESRD patients, indicated by significantly decreased global properties and aberrant nodal properties; (3) functional network construction demonstrated a lateralized distribution of changed Ne and Dc in ESRD patients compared to HCs; and (4) in the functional network, decreased γ, σ, and Ne in the right insula and superior temporal gyrus were significantly correlated with cognitive dysfunction in ESRD patients. Interestingly, related topological parameters could be selected as features to predict cognitive function in ESRD patients in our study.

Graph theory analysis can quantitatively analyze the segregation and integration of brain networks (). Functional segregation represents the ability of functional specialization within interconnected brain areas and is measured by Cp, γ, and Eloc. Functional integration represents the ability of global communication within distributed brain regions and is measured by Lp, λ, and Eg. Consistent with previous work, we demonstrated the common small-world organization of both GM and functional networks in ESRD patients and HCs (). However, in our study, changed topological parameters of both GM and functional networks were found in ESRD patients compared with HCs. The data revealed a downward trend of the highly optimized balance of small-world network organization. Specifically, our results indicate that this organization may have shifted toward a hemodialysis-associated network randomization, to some extent ().

Based on the group level, disruption of GM networks within four DKI metrics was found in the ESRD group, as indicated by decreased global measures. In a previous study, similar disruption of the GM network had been observed in AD (). In addition, the changed topological properties were thought to be caused by the pathology of AD, including astrogliosis, microglial activation, vascular hyalinization, and axonal loss (). Thus, we assumed that alterations of GM networks in ESRD patients could be linked to changes in brain microstructure. Decreased γ and σ were found in the GM network with MK, which may reflect reduced complexity of the brain microstructure (). Decreased Cp and γ were also found in the GM network with AK and RK, which may indicate the loss of nerve axons and a breakdown of myelin membranes (). Furthermore, the GM network with KA involved more topological parameters in our study, which may suggest that changes in fiber bundle crossover regions were more severe. To some extent, this result was consistent with a previous study focused on the WM network demonstrating the disrupted fiber integrity (). Interestingly, GM networks with four DKI metrics showed various aspects of disrupted microstructure, indicating significantly impaired functional segregation and relatively preserved functional integration of GM networks in ESRD patients. These results suggest that DKI may be sensitive enough to detect microstructural differences from the perspective of the GM network.

Based on the individual level, disruption of functional networks was found in ESRD patients, indicated by decreased global measures and abnormal nodal measures. Consistent with previous studies, we observed decreased γ and σ of the functional network in ESRD patients (; ). A similar decline in separation capabilities in both GM and functional networks was found in our study. This finding may suggest a link between microstructural connectivity and functional connectivity in the brain of ESRD patients (). Moreover, decreased γ and σ of functional networks were positively correlated to MoCA scores in ESRD patients, which indicates that these parameters and potential imaging markers may reflect CI.

In contrast to our global measure findings, ESRD patients had more subtle changes in nodal measures of the functional network. Altered Ne and Dc were mainly observed in the default mode network, limbic system, frontal lobe, temporal lobe, and occipital lobe. This result was in accordance with previous studies and expanded our understanding of network organization degradation in ESRD patients (; ; ). First, we observed a lateralized distribution of these brain regions. All brain regions with increased nodal parameters were included in the left cerebral hemisphere, and most brain regions with decreased nodal parameters were included in the right cerebral hemisphere. This may be related to the left-lateralization in righted-handed people (). Our data strongly indicate that left regions of the brain may show more strong associations with brain compensatory mechanisms, and the right regions may be more susceptible to brain network impairment in right-handed ESRD patients. While our analyses could not directly address the relationship between alterations of nodal measures and pathophysiological changes, our results may be partly attributed to the ESRD-related compensatory mechanism. Second, decreased Ne of the functional network in the right insula and the right superior temporal gyrus showed a significantly positive correlation with MoCA scores. The insula is a part of the highly interconnected cognitive control network involved in various functional tasks, including auditory and vestibular functions, motor plasticity, cognitive control, and individual and social emotions (). The superior temporal sulcus is a crucial hub for speech perception and emotional and memorial information processing (). Taken together, these data suggest that nodal properties of brain regions may be used as sensitive imaging markers to detect CI, and non-invasive neural intervention may be feasible in the clinic.

Although we observed significant correlations between network measures and MoCA scores in ESRD patients, there were limitations that limited the ability of these results to guide early clinical diagnoses based only on correlation analyses. Thus, we further applied optimized LSSVRM to tentatively predict cognitive function in ESRD patients. LSSVRM is an extension of the support vector regression machine (SVRM) that improves prediction efficiency by altering the inequality constraint and quadratic programming problems in the SVRM model into the equality constraint and solution of linear equations. Furthermore, the LSSVRM improves prediction accuracy by taking the error square and loss function to effectively fit the cognitive scores with non-linear characteristics (). WOA can optimize the selection strategy of kernel function parameters and further improve the operating efficiency of the model (). The combination of LSSVRM and WOA takes work efficiency and prediction accuracy into account (). In our study, the prediction models based on functional networks showed a relatively great prediction accuracy, with the prediction accuracy based on combined global and nodal measures being slightly higher than that based only on a single type of measures. This may suggest that critical nodal changes have a certain influence on the overall CI observed in ESRD patients. Given that brain network changes were associated with cognitive performance and may be able to be selected as features to construct prediction models, it would be of further interest to investigate the feasibility of clinical application in future studies.

Despite the novelty of our study findings, we acknowledge some limitations of our study. First, due to the cross-sectional design and relatively small sample size, the power of the statistical analysis and prediction results may be affected. And the parameters of prediction models should be used with caution. Longitudinal studies with larger sample sizes are needed to provide dynamic perspectives in the future. Second, further neuropsychological tests are essential to provide more information. Third, exploring the occurrence and progression of CI will further contribute to the efficient development of clinical intervention. Thus, in our future researches, we will include ESRD patients with varying degrees of CI, including normal cognitive function, subjective cognitive function, mild CI, and moderate CI. Finally, the feature extraction, selection, and fusion methods utilized can be improved to better mine the information of structural and functional brain networks and enhance the prediction accuracy achieved.

Conclusion

In summary, our study demonstrated that disrupted GM and functional networks in ESRD patients on maintenance hemodialysis contribute to and can be used to predict CI. In contrast to most diffusion MRI studies that focus merely on WM networks, our study paid attention to GM networks. Moreover, related functional network metrics were found to be significant imaging markers capable of predicting cognitive function in ESRD patients. Ultimately, this study highlighted the feasibility and necessity of early diagnosis of CI in ESRD patients from the perspective of quantitative imaging parameters.

Statements

Data availability statement

The raw data supporting the conclusions of this article will be made available by the authors, without undue reservation.

Ethics statement

The studies involving human participants were reviewed and approved by The Affiliated Changzhou No. 2 People’s Hospital of Nanjing Medical University. The patients/participants provided their written informed consent to participate in this study.

Author contributions

JZ and HS made substantial contributions to the conception and design of this study. JZ and XW made substantial contributions to the acquisition of data. JZ made substantial contributions to the writing of manuscript. JD, TL, and ZJ made substantial contributions to the revision of the manuscript. All authors revised the draft for intellectual content, gave their final approval of the final version for publication, and agreed to be accountable for all aspects of the work in ensuring that questions related to the accuracy or integrity of any part of this study were appropriately investigated and resolved.

Funding

This study was funded by Changzhou Municipal Health Commission Major Project (ZD201913) and Jiangsu Province’s 2020 High-level Health Talents Six One Project Top-notch Talent Research Project (LGY2020036).

Conflict of interest

JD was employed by GE Healthcare, MR Research China. The remaining authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

Publisher’s note

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.

References

Summary

Keywords

end-stage renal disease, resting-state functional magnetic resonance imaging, diffusion kurtosis imaging, graph theoretical analysis, predict

Citation

Zheng J, Wu X, Dai J, Pan C, Shi H, Liu T and Jiao Z (2022) Aberrant brain gray matter and functional networks topology in end stage renal disease patients undergoing maintenance hemodialysis with cognitive impairment. Front. Neurosci. 16:967760. doi: 10.3389/fnins.2022.967760

Received

13 June 2022

Accepted

18 July 2022

Published

10 August 2022

Volume

16 - 2022

Edited by

Jiaojian Wang, Kunming University of Science and Technology, China

Reviewed by

Xin Zhang, Nanjing Drum Tower Hospital, China; Feifei Zhang, Sichuan University, China

Updates

Copyright

*Correspondence: Haifeng Shi, Tongqiang Liu, Zhuqing Jiao,

This article was submitted to Brain Imaging Methods, a section of the journal Frontiers in Neuroscience

Disclaimer

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article or claim that may be made by its manufacturer is not guaranteed or endorsed by the publisher.

Outline

Figures

Cite article

Copy to clipboard


Export citation file


Share article

Article metrics