Extensive Dynamic Functional Network Connectivity Alterations in Diabetic Retinopathy Among Patients with Type 2 Diabetes

Introduction

Diabetes stands as one of the most prevalent chronic diseases worldwide, presenting a significant public health challenge in the 21st century.¹ It is projected that by 2040, the global diabetic population will soar to 642 million individuals.² Diabetes can trigger various microvascular complications across multiple organs, including the brain,³ eyes,⁴ heart,⁵ and kidneys,⁶ with diabetic retinopathy (DR) emerging as a severe complication. The global prevalence of DR among diabetic patients is 22.27%, ranking as a prominent cause of blindness globally.^7,8 Noteworthy risk factors for DR encompass hyperglycemia, hypertension, dyslipidemia, duration of diabetes mellitus, and pregnancy.⁴ The primary pathological characteristics of DR involve microvascular blockade, microangioma development, capillary nonperfusion, vascular permeability, and retinal neurodegeneration, potentially progressing to neovascularization, proliferative retinal detachment, and eventual blindness. These microangiopathies and neurodegeneration are associated with inflammation, oxidative stress, and a variety of metabolic abnormalities, including abnormalities in the polyol pathway, accumulation of advanced glycosylation end products (AGEs), activation of the protein kinase C (PKC) pathway, and increased flux of the hexosamine pathway.^9–11 Remarkably, the retinal vasculature shares resemblances in anatomy, physiology, and embryology with the cerebral vasculature.¹² Autopsies of diabetic patients often unveil analogous microangiopathy in their brains.¹³ Both diabetes-induced retinal neurodegeneration and cerebral neurodegenerative diseases exhibit common pathogenic pathways,¹⁴ and implying that diabetes could provoke structural and functional neuronal aberrations in the brain.¹⁵ Numerous studies also highlight impaired cognitive functions in individuals with DR,^16,17 suggesting that besides affecting ocular health, DR may also induce central nervous system abnormalities.

Several neuroimaging studies have revealed a spectrum of structural and functional brain changes among patients with DR. Diffusion-weighted imaging techniques have demonstrated notably elevated apparent diffusion coefficient (ADC) values in key visual regions (orbitofrontal cortex, cingulate gyrus, and visual cortex) in DR patients.¹⁸ Reduced gray matter density in the right inferior frontal gyrus and right occipital lobe has been observed in individuals with DR.¹⁹ An examination of the structural gray matter networks in the brains of DR patients showcased diminished local path lengths and local clustering, which are believed to be linked to functional deficits.²⁰ These structural alterations are thought to be associated with some functional impairments in DR patients. Regarding the functional changes of the brain in patients with DR, it has been found that the amplitude of low-frequency fluctuation (ALFF) values in the bilateral calcarine area and the middle occipital gyrus were significantly lower in patients with DR,^21,22 and the regional homogeneity of the right anterior cingulate gyrus and right cuneiform lobe (ReHo) values were also significantly lower than those of the healthy controls group.²³ Altered functional connectivity has been noted in various networks, including the auditory, visual, default mode, salience, and sensorimotor networks in patients with DR, potentially leading to interhemispheric coordination issues and subsequent functional irregularities.²⁴ Additionally, abnormal topological properties in functional brain networks among DR patients may contribute to visual and cognitive impairments.²⁵ Nonetheless, the precise alterations in dynamic large-scale brain networks in DR patients remain ambiguous.

Functional network connectivity (FNC) is a valuable tool for examining the temporal correlations among resting-state networks (RSNs). The resting state means that subjects are instructed to let their minds roam freely in the absence of a task or stimulus, which, unlike studies in the task-based state, avoids the incomparability of experimental results in task-based studies due to differences in the design of the task and variability in its execution by the subjects. In a prior study, we investigated FNC alterations in individuals with DR and discovered significant deficits in both low-level perceptual and high-level cognitive networks in this population. This research provided crucial insights into the neural mechanisms associated with vision impairment and cognitive decline in individuals with DR.²⁶ However, the static nature of traditional FNC analysis assumes that functional interactions remain consistent throughout resting-state scans. In reality, emerging evidence suggests that human brain activity is dynamic, characterized by substantial time-varying fluctuations.^27,28 Notably, functional connectivity can fluctuate in response to various factors such as task demands, learning processes, and significant state transitions including sleep and sedation, showcasing changes in both the strength and directionality of connectivity not only between scanning runs but also on faster timescales (seconds to minutes).²⁹ The exploration of dynamic functional network connectivity (dFNC) unveils a wealth of information embedded in these time-varying features, offering insights into the macroscopic neural activity patterns that underlie critical aspects of cognition and behavior, thus warranting further investigation.

The current study utilized a combination of the sliding window approach and k-means cluster analysis, integrating independent component analysis (ICA) to examine dFNC changes in individuals with DR and to identify inter-group disparities in three dynamic temporal metrics. Our hypothesis posited aberrant functional connectivity within the dynamic large-scale brain networks of DR patients.

Materials and Methods

Participants

Forty-six patients with DR (type 2 diabetes mellitus) and 46 healthy controls matched for sex, age, and education were recruited for this study from the Department of Ophthalmology at Jiangxi Provincial People’s Hospital between January 2022 and January 2024. Ethical approval number (202310114). The inclusion criteria for DR patients were: 1) Fasting blood glucose levels ≥ 7.0 mmol/L, random blood glucose levels ≥ 11.1 mmol/L, or 2-hour glucose levels ≥ 11.1 mmol/L; 2) Patients with nonproliferative DR exhibiting microaneurysms, hard exudates, and retinal hemorrhages.

The exclusion criteria for DR patients were: 1) Presence of proliferative DR with retinal detachment or vitreous hemorrhage; 2) Coexistence of other ocular complications (such as cataracts, glaucoma, high myopia, or optic neuropathy); 3) Diagnosis of diabetic nephropathy (urine albumin/creatinine ratio > 30 mg/g persisting for more than 3 months).

Healthy controls were mainly volunteers without other ophthalmic diseases who met the inclusion criteria. The inclusion and exclusion criteria for healthy controls were: 1) Fasting blood glucose levels < 7.0 mmol/L, random blood glucose levels < 11.1 mmol/L, glycosylated hemoglobin levels < 6.5%; 2) No ocular diseases (such as myopia, cataracts, glaucoma, optic neuritis, or retinal degeneration); 3) Binocular visual acuity ≥ 1.0; 4) No history of ocular surgery; 5) Absence of psychiatric disorders. All subjects had no contraindications to MRI such as claustrophobia, metallic implants or metallic foreign bodies in the body.

This study protocol followed the Declaration of Helsinki guidelines and received approval from the Research Ethics Committee of Jiangxi Provincial People’s Hospital. Before participating, all individuals were fully briefed on the study’s objectives, methodologies, potential risks, and provided written informed consent.

Magnetic Resonance Imaging Acquisition

Magnetic resonance imaging (MRI) scans were performed on a 3-T MRI scanner (Discovery MR 750W system; GE Healthcare, Milwaukee, WI, United States) with an eight-channel head coil. Functional images were obtained utilizing a gradient-echo-planar imaging sequence. Participants were directed to rest quietly with their eyes closed, maintaining a relaxed state without fixating on any particular thoughts or drifting off to sleep. During the scanning process, we monitored the subject’s condition throughout to ensure that they were not experiencing abnormalities. Whole-brain T1-weighted images were obtained through three-dimensional brain volume imaging (3D-BRAVO) MRI with the following parameters: repetition time [TR]/echo time [TE] = 8.5/3.3 ms, slice thickness = 1.0 mm, no interslice gap, acquisition matrix = 256 × 256, field of view = 240 mm × 240 mm, flip angle = 12 degrees, and 240 slices. Functional images were captured using a gradient-echo-planar imaging sequence with these parameters: TR/TE = 2000/25 ms, slice thickness = 3.0 mm, interslice gap = 1.2 mm, acquisition matrix = 64 × 64, flip angle = 90 degrees, field of view = 240 mm × 240 mm, voxel size = 3.6 mm × 3.6 mm × 3.6 mm, and 35 axial slices.

Data Preprocessing

All preprocessing steps were conducted using the Data Processing and Analysis of Brain Imaging (DPABI) toolbox,³⁰ which operates on Statistical Parametric Mapping (SPM12) within MATLAB 2013a (MathWorks, Natick, MA, United States). The following procedures were implemented. The first 10 volumes of each scan were discarded to allow for magnetic stability and thus to generate a steady blood oxygenation level-dependent activity signal. Then, the remaining 230 images were slice-time corrected and realigned for head-motion correction. Participants exhibiting head movement beyond 2 mm translation or 1.5° rotation during the scan were excluded.³¹ The generated images were spatially normalized to the Montreal Neurological Institute (MNI) 152 space and resampling at a resolution of 3 mm x 3 mm x 3 mm. Application of spatial smoothing using a 6 mm x 6 mm x 6 mm isotropic Gaussian kernel with full width at half maximum.

An overview of the analytical process is depicted in Figure 1.

Figure 1 Overview of the analytical process for this study: The study first applied the ICA method to extract RSNs from resting-state functional magnetic resonance imaging data, and then combined the sliding time window and k-mean cluster analysis methods to obtain five stable repetitions of dFNC states. Then the dFNC differences between the two groups of patients in each state were compared. Finally, between-group differences in three dynamic temporal metrics were explored. Abbreviations: ICA, independent component analysis; RSN, resting-state network; dFNC, dynamic functional network connectivity.

Group Independent Component Analysis Analysis

Group ICA was conducted to decompose the data into independent components (ICs) utilizing the GIFT toolbox (version 3.0b). Initially, 24 IC maps were estimated in this research, employing the minimum description length criterion to account for spatial correlation. Subsequently, the ICs for each participant were obtained from the group ICA back-reconstruction phase and were standardized into z-scores.³² Components selected from the 24 estimated ICs for further analysis were chosen based on their strongest spatial correlation with specific RSN templates.^33,34 The IC time courses and spatial maps for each individual were then converted into z-scores, leading to the identification of fourteen RSNs in this study.

Dynamic Functional Network Connectivity Analysis

The dFNC matrix was computed using a sliding window approach. In this investigation, the window width was configured to a TR of 30 (60s), with the window sliding along the time axis in steps of 2 TR.

All subjects’ dFNC matrices were subjected to k-means clustering to evaluate the occurrence and configuration of recurring dFNC patterns. The analysis employed sqEuclidean distance to gauge the resemblance between distinct time windows. To enhance the algorithm’s convergence, we established a maximum of 500 iterations and 150 repetitions. The elbow rule guided the selection of the optimal cluster number, which was determined to be k=5. Consequently, the dFNC matrices for all participants were classified into five dFNC states representing recurrent instantaneous functional connectivity patterns across various windows and subjects, with the cluster centroid defined as the central dFNC matrix.

A range of temporal features were computed as: (i) defining the fraction of time as the ratio of the time windows in a state to the total count, (ii) determining the mean dwell time representing the average duration within a specific state, and (iii) counting the transitions reflecting the instances when subjects transitioned between states during the scanning period.

Statistical Analysis

The differences of the clinical and demographic characteristics between DR patients and healthy controls were assessed using the χ² test for categorical variables and an independent-samples t-test for continuous variables by SPSS version 16.0 (SPSS Inc., Chicago, IL, USA).

The independent components (ICs) corresponding to fourteen RSNs were extracted from all participants, and one-sample t-tests were conducted on the spatial maps of each RSN using the SPM12 software, aiming to create masks for two-sample t-tests by limiting the size of the brain voxels to make the results more precise. The statistical significance thresholds were established at P < 0.01, corrected for false discovery rate (FDR).

To assess the variations in dFNC between the DR and healthy control cohorts, direct comparisons were made utilizing two independent samples t-tests within the Stats module of the GIFT software package (p < 0.05, FDR-corrected). Additionally, a two-sample t-test was employed to scrutinize the differences in fraction time, mean dwell time, and number of transitions for each state between the groups (p < 0.05, FDR-corrected).

Results

Demographic and Clinical Characteristics

The clinical and demographic characteristics are detailed in Table 1. No significant differences in sex, education, or age were observed between the DR patients and healthy controls. However, notable disparities were detected in bilateral best-corrected visual acuity, showing statistical significance (P < 0.001).

Table 1 Demographics and Visual Measurements Between Two Groups

Spatial Pattern of Resting State Networks in Each Group

The typical spatial patterns in each RSN of both DR and healthy control groups, as illustrated in Figure 2. Fourteen of these components coincided with RSNs included: DMN (IC3, IC15, IC16, IC17): posterior cingulate cortex, precuneus, medial prefrontal cortex, inferior parietal lobule, and lateral temporal cortex.; DAN (IC7): the inferior parietal cortex, frontal eye motor area, auxiliary motor area, insular lobe, and dorsolateral prefrontal cortex; VN (IC8, IC9, IC13, IC20, IC24): middle occipital gyrus, superior occipital gyrus, the temporal-occipital regions, and fusiform gyrus; sensorimotor network(SMN)(IC10): precentral gyrus, postcentral gyrus, and supplementary motor area; ECN (IC11, IC21): dorsolateral frontal cortex and posterior cingulate cortex; CN (IC23): cortex and medulla (parietal nucleus, intermediate nucleus, and dentate nucleus).

Figure 2 Spatial maps of the identified RSNs, including DMN (IC3, IC15, IC16, IC17), DAN (IC7), VN (IC8, IC9, IC13, IC20, IC24), SMN (IC10), ECN (IC11, IC21), CN (IC23). Different colors pass spatial information. Abbreviations: RSNs, resting-state networks; IC, independent component; DMN, default mode network; DAN, dorsal attention network; VN, visual network; SMN, sensorimotor network; ECN, executive control network; CN, cerebellum network.

Dynamic Functional Network Connectivity Analysis

Cluster Analysis

Five recurrent states of dFNC matrices were identified through the scans using the k-means clustering algorithm. The distribution of these states varied among subjects: State 1 (24%) (Figure 3A), State 2 (7%) (Figure 3B), State 3 (26%) (Figure 3C), State 4 (11%) (Figure 3D), and State 5 (32%) (Figure 3E). The visualization of functional network connectivity in each state is illustrated in State 1 (Figure 3F), State 2 (Figure 3G), State 3 (Figure 3H), State 4 (Figure 3I), and State 5 (Figure 3J).

Figure 3 Five recurrent states of dFNC matrices were identified through the scans using the k-means clustering algorithm. The distribution of these states varied among subjects: State 1 (24%) (A), State 2 (7%) (B), State 3 (26%) (C), State 4 (11%) (D), and State 5 (32%) (E). The visualization of functional network connectivity in each state is illustrated in State 1 (F), State 2 (G), State 3 (H), State 4 (I), and State 5 (J).(K and L) Significant differences in dFNC between DR and HCs were observed in state1 and state2 (p<0.05, FDR correction). Abbreviations: dFNC, dynamic functional network connectivity; DR, diabetic retinopathy; HCs, health controls; DMN, default mode network; DAN, dorsal attention network; VN, visual network; SMN, sensorimotor network; ECN, executive control network; CN, cerebellum network.

Comparison of Dynamic Functional Network Connectivity Between Groups

Two independent sample t-tests were utilized for further comparison of the dFNC matrix in each state between groups, revealing significant differences in state 1 and state 2 between DR patients and healthy controls (p < 0.05, FDR corrected). When compared to healthy controls, the DR group exhibited increased connectivity between default mode network-ventral network (DMN-VN), sensorimotor network-ventral network (SMN-VN), DMN-executive control network (DMN-ECN), and cingulo-opercular network-ventral network (CN-VN) in state 1, whereas decreased connectivity was observed between dorsal attention network-ventral network (DAN-VN) (Figure 3K and Table 2). State 2 demonstrated enhanced connectivity between CN-VN and reduced connectivity between SMN-VN and VN-VN in the DR group (Figure 3L and Table 2).

Table 2 Significant Dynamic Functional Network Connectivity Differences Across States

Comparison of Dynamic Functional Network Connectivity Temporal Metrics Between Groups

Figure 4 illustrates that compared to healthy controls, patients with DR exhibited a significant decrease in the fraction of time (Figure 4A) and mean dwell time in state 1(Figure 4B), along with a significant increase in the fraction of time (Figure 4A) and mean dwell time in state 3 (Figure 4B) (p < 0.05). Notably, there was no significant difference in the number of transitions between DR patients and healthy controls (Figure 4C and D) (Table 3).

Table 3 Differences in Dynamic Temporal Metrics Across States Between DR and HC Groups

Figure 4 Dynamic temporal metrics (fractional time (A), mean dwell time (B), number of transitions (C), probability of the transitions (D)) differences between DR and HC groups. *P < 0.05, **P < 0.01, ***P < 0.001. Abbreviations: DR, diabetic retinopathy; HC, health control.

Discussion

This study explored dynamic changes in large-scale brain networks in patients with DR using ICA with a sliding time window approach and k-means cluster analysis. The analysis identified five recurring states over time, with notable differences in dFNC between DR and healthy control groups in states 1 and 2. In these states, compared to healthy controls, DR group exhibited reduced functional connectivity within the VN and among the DAN and VN, paired with compensatory increases in connectivity between the DMN and VN, Central Network (CN) and VN, and DMN and Executive Control Network (ECN). Assessment of three dynamic time metrics revealed that DR patients, in comparison to healthy controls, displayed significantly reduced fraction of time and mean dwell time in state 1, and increased fraction of time and mean dwell time in state 3. These results provide valuable insights into the neural mechanisms underlying dynamic alterations in large-scale brain networks in individuals with DR.

In the cluster analysis we identified five states recurring over time, each with their own distinct dFNC patterns, which reflects the flexibility of functional coordination between brain networks. Significant differences in dFNC between the two groups were found in state 1 and state 2, which were mainly characterized by a decrease in functional connectivity within VN and between DAN-VN, along with a compensatory elevation of functional connectivity between DMN-VN, CN-VN and DMN-ECN. Patients with DR often have varying degrees of visual impairment or even blindness.^35,36 This is primarily thought to be caused by microangiopathy and neurodegeneration of the retina. However, studies have shown similar neuronal abnormalities in the proximal and distal optic nerves and brain.³⁷ Combined with the anatomical, physiologic, and embryologic similarities of the retinal vasculature to the cerebral vasculature, we suggest that the cause of visual impairment in patients with DR is not limited to these retinal lesions. The VN is located in the occipital lobe and includes brain regions such as the lingual gyrus, the fusiform gyrus, and the middle occipital gyrus, which are mainly responsible for the processing of visual information.³⁸ DR patients were found to have significantly reduced occipital grey matter density (GMD).¹⁹ Yu et al showed that the functional connectivity between the left primary visual cortex (V1) and the cuneus/calcarine/precuneus was weakened in patients with DR, which may cause visual information processing abnormalities.³⁹ Huang et al found large-scale neural network dysfunction in DR patients, including reduced FC within the VN, which was thought to be related to vision loss in DR patients.²⁶ Weakened functional connectivity within the VN was also found in our previous studies of thyroid-associated ophthalmopathy and primary angle-closure glaucoma,^40,41 and in conjunction with these findings, we hypothesized that reduced functional connectivity within the VN is associated with impaired visual acuity in patients with DR. This provides useful insights into the mechanisms of visual impairment in DR patients and is favorable to the preservation of vision in patients with DR. The DAN, which includes areas such as the intraparietal sulcus, the middle frontal gyrus, and the frontal eye fields, is responsible for attentional orienting and is also capable of allocating cognitive resources in response to stimuli.^42,43 The DAN belongs to the high-level cognitive network, while the VN belongs to the low-level perceptual network. According to the previous inference, DR patients have abnormal visual function, and changes in visual information processing may cause cognitive impairment,⁴⁴ which may be caused by reduced input of visual information and stimulation of higher-level cognitive networks. On the other hand, the processing of perceptual information requires the support of top-down attentional mechanisms,^45,46 so the normal operation of visual function also requires the support of corresponding cognitive resources, otherwise it may lead to visual dysfunction. Combining these analyses, we suggest that functional connectivity between the DAN-VN is decreased in DR patients, which may lead to cognitive deficits and impaired visual function.

It has been found that spontaneous activity in the default mode network is negatively correlated with spontaneous activity in the attention network.⁴⁷ Previously, we found a weakening of the functional connectivity between the DAN and the VN, while at the same time, we found an enhancement of the functional connectivity between the DMN and the VN. Therefore, we believe that the decrease in the input of visual information and the stimulation of the DAN led to the weakening of the spontaneous activity of the DAN, which caused the corresponding dysfunction. In contrast, the spontaneous activity of the DMN was enhanced, and the functional connectivity with the VN was also enhanced, compensating for the impairment of cognitive and visual functions through this compensatory mechanism. In the past, cerebellar function was mainly understood in terms of motor control, but current research has found that the CN plays an important role in multimodal integration, for example, it contains afferent fibers of the visual sensory system that convey rich visual information and direct visual attention.⁴⁸ CN-VN functional connectivity is enhanced in DR patients, which may be a compensatory mechanism in order to maintain normal visual-motor functional integration in the presence of abnormal visual function. The ECN and DMN are active during task and off-task periods,^49,50 respectively, and switch under the control of the salience network.^51,52 Enhanced functional connectivity between the DMN and ECN may imply that the state of functional isolation between them has been disrupted and functional crosstalk is occurring, which is detrimental to the correct allocation of cognitive resources and may exacerbate cognitive impairment.⁵³ It is worth noting that the above between-group differences in dFNC may be a potential neuroimaging indicator in patients with DR and may help in the diagnosis of DR.

For the three dynamic time metrics, the results showed a significant decrease in fraction of time and mean dwell time in state 1 and a significant increase in fraction of time and mean dwell time in state 3 in DR patients compared to healthy controls. By looking at the dynamic functional connectivity matrix diagram, we find that each of these five stable and repeatable states has its own characteristics. Among them, state 1 mainly exhibits stronger positive connectivity, especially stronger functional connectivity within VN and between DMN and VN. DR patients had significantly less fraction of time and mean dwell time in state 1, and this short dwell time may not be conducive to the functional interactions between the above networks,⁵⁴ which may be one of the mechanisms causing visual and cognitive impairment in DR patients. State 3, on the other hand, mainly showed sparser functional connectivity. Inter-network functional connectivity reflects the ability of functional integration between different regions of the brain.⁵⁵ Sparse functional connectivity often implies inefficient functional integration, which is not conducive to the synergistic performance of various higher cognitive functions by different brain regions,^53,54 and may be one of the mechanisms of cognitive impairment in DR patients.

Several limitations are inherent in this study. First, the sample size of the study remains small and there is no comparative study of patients with different severities and stages of the disease, which limits the generalization of the findings to some extent. In the future, we will include larger sample sizes and study patients with varying degrees of disease severity. Secondly, the RSN values derived from blood oxygen level-dependent signals are susceptible to physiological noise, such as heartbeat and respiratory activity. Thirdly, our exploration was confined to the resting-state network extracted through ICA without considering the impacts of other brain networks. Lastly, while all participants were instructed to maintain a state of rest and minimal movement during the scanning procedure, complete assurance of compliance with these instructions could not be guaranteed.

Conclusions

In conclusion, our study illustrates the existence of a extensive altered dynamic functional network connectivity in patients with DR, which are mainly focused on the visual and higher cognitive networks and may be related to visual impairment and cognitive deficits in patients with DR. At the same time, dynamic temporal metrics were altered in DR patients in different brain states. These insights shed light on the neural mechanisms underlying the large-scale brain network alterations in DR patients and have positive implications for the diagnosis and treatment of the disease.