A cross-sectional and observational study was done on 600 participants who were recruited from outpatient and inpatient diabetes clinics at the Guangzhou University of Traditional Chinese Medicine between 2020 and 2023. The inclusion criteria were right-handedness and age 30-70 years. 566 of the 600 participants had MRI scans available. Among the 566 participants with accessible MRI measures, 493 with complete clinical data and images ( Supplementary Figure S1 ). Committee approval (NO. K-2013-146) was obtained from Guangzhou University of Traditional Chinese Medicine’s First Affiliated Hospital for the study. All were asked to sign a written informed consent prior to participating in the study.

Patients with a history of diabetes mellitus were tested for fasting plasma glucose (FPG), a standardized 2-h 75-g oral glucose tolerance test, and HbA_1c levels. Participants without a history of diabetes were tested for FPG and HbA_1c levels. All trials were completed at the First Hospital of Guangzhou University of Traditional Chinese Medicine. The classification of prediabetes and T2DM was determined based on the diagnostic criteria provided by the American Diabetes Association in 2021 (15). Participants were classified as having prediabetes if their FPG was between 5.6~6.9 mmol/L and/or HbA_1c was between 5.7~6.4%. The diagnosis of T2DM was confirmed by a self-reported history of diabetes, use of hypoglycemic drugs, FPG ≥7.0 mmol/L, 2-h postload glucose ≥11.1 mmol/L, or HbA_1c ≥6.5%. Normal glucose metabolism (NGM) is defined as FPG and HbA_1c within the normal range.

The 64-channel 3.0T MRI scanner (Prisma, Siemens, Germany) was used to obtain 3D T1-weighted imaging, diffusion-weighted MRI (dMRI) and resting-state functional MRI (fMRI), and the scanning parameters are shown in Supplementary Table S1 . All neuroimaging data underwent a thorough visual examination to rule out artifacts and inaccurate segmentation and registration. If the head movement was more than 3 mm in three dimensions or 3 degrees in rotation, images were not included. The preprocessing of fMRI data was using SPM 12 software. The processing steps included removing the first 10 time points, correcting for slice timing and head motion, normalizing to Montreal Neurological Institute coordinates using the DARTEL method, applying spatial smoothing with Gaussian kernels of 6 mm full width at half maximum (FWHM), linear detrending, and regressing global signal, cerebrospinal fluid signals, white matter, and Friston-24 head motion parameters. Ultimately, the impacts of high-frequency physiological noise and low-frequency drift were mitigated by using a bandpass filter with a frequency range of 0.01 to 0.08 Hz.

The DTI data underwent preprocessing and analysis using the PANDA v1.3.1 toolbox (http://www.nitrc.org/projects/panda). Rigid-body transforms were performed on the b0 images to correct eddy current distortions and head motion artefacts. The Diffusion Toolkit was used to estimate diffusion tensor models at each voxel using the linear least-squares fitting approach. The Diffusion Toolkit’s Fiber Assignment by Continuous Tracking technique was applied to track whole brain fibers in native diffusion space. Streamlines tracking was stopped when the angle between the current and the previous path segment surpassed 45° or the FA was less than 0.2.

In order to measure the strength of functional connectivity (FC), we calculated the Pearson correlation between the average time series of each pair of brain areas, utilizing the Human Brainnetome Atlas (16). As a result, each participant has a 246×246 FC matrix with a Gaussian distribution after Fisher’s Z transformation. The components of negative correlations were set to zero due to the contentious physiological interpretation of negative correlations (17, 18).

In order to obtain structural connectivity (SC), we defined each region of interest (ROI) as a network node in the Human Brainnetome Atlas. The Human Brainnetome Atlas from the MNI space was registered to the individual’s native space via inverse transformations. In native diffusion space, structural connectivity was considered to be present if streamlines were present in at least 80% of participants in each pair of brain regions. We calculated the weight of each edge by taking the mean FA values of all the fibers.

In order to analyze the alterations in the SC-FC coupling at the network level, we implemented the classic seven-network parcellation according to Yeo atlas (19), which includes the visual network (VIS), somatomotor network (SOM), ventral attention network (VEN), dorsal attention network (DOR), limbic network (LIM), frontoparietal network (FPN), and default mode network (DMN). Subcortical regions were included in the eighth network. The procedure for constructing the network is outlined in Figure 1 .

Correlation coefficients between structural and functional connectivity are used to measure SC-FC coupling. This correlation was limited for each participant by the edges of non-zero structural connection. To be more precise, a vector of structural connectivity values was formed by extracting non-zero structural connectivity edges, and a corresponding vector was created by extracting functional connectivity. The SC-FC coupling values were then obtained by measuring the correlation coefficient between the two priori vectors using Spearman’s correlation. Lastly, tractography algorithms are known to underestimate connectivity across different hemispheres. Consequently, we also compute the SC-FC coupling inside a hemisphere.

The characteristics of the participants were expressed as mean ± SD (continuous variables) and percentage (categorical variables). Comparisons between glucose metabolism status were conducted using statistical analyses appropriate for the type of data. ANOVA was used for continuous variables that followed a normal distribution; the Kruskal-Wallis test was used for continuous variables that exhibited a skewed distribution; and the categorical variables were analyzed using the Pearson χ² test. A post hoc test was then performed, using two-sample t tests for means and χ² test for proportions. If the SC-FC coupling values conform to the normal distribution, a general linear model (GLM) is used for statistical analysis; otherwise, a non-parametric permutation test is used, followed by a post hoc test (P <.05). The analyses were controlled for potential confounding variables, specifically age, sex, and education level (model 1). Additional adjustments were made for cardiovascular disease risk factors, including BMI, office systolic blood pressure, the ratio of total cholesterol to HDL, lipid-modifying medication, antihypertensive medication, and history of cardiovascular disease (model 2). The variance inflation factor was used to test for multicollinearity of the independent variables. Interaction terms for sex * glucose metabolism status were added to explore the interaction. The analyses were conducted using SPSS (IBM, SPSS, version 27).

To investigate the association between diabetes-related measures, regional SC-FC coupling, and the Montreal Cognitive Assessment (MoCA) score (20), and to explore whether the brain SC-FC coupling acts as a potential mediator in the association between diabetes-related measures and MoCA score, we adopted a two-stage statistical analysis. In the first stage, multiple linear regression analysis was used to examine the relationship between glucose metabolism status, diabetes-related measures, and regional SC-FC coupling. Dummy variables were then used to evaluate regression coefficients for different glucose metabolism groups. The correction for confounding variables is consistent with the above. Secondly, based on the significant association between exposure (diabetes-related measures) and mediator (regional SC-FC coupling), as well as the association between the mediator and the outcome (MoCA score), we performed the mediation analysis to examine whether brain SC-FC coupling metrics mediated the association between exposure and outcome. To calculate bias-corrected 95% CIs, we used bootstrapping (5000 samples) and the PROCESS statistical package for SPSS (21).

Table 1 displays the baseline demographic information. Out of 493 people who were part of the study, 119 were NGM, 125 were prediabetes, and 249 were T2DM. The average age was 50.6 years, and 50.1% were female. Participants with prediabetes and T2DM were significantly older and had higher cardiovascular risks than NGM individuals ( Table 1 ).

Figure 2 illustrates the average SC-FC coupling among 493 unrelated individuals. Regional SC-FC coupling varied substantially between subcortical and cortical regions but was generally positive at the group level. In comparison to the other networks, the SC-FC coupling in the ventral attention and somatomotor areas was generally higher (0.29 ± 0.40 and 0.23 ± 0.20, respectively). In contrast, the coupling in the subcortical and dorsal attention areas was significantly weaker (-0.06 ± 0.16 and -0.12 ± 0.23, respectively) ( Figure 2 ).

T2DM was significantly associated with weaker regional SC-FC coupling in the limbic areas (β = -0.110; 95% CI: -0.196 to -0.024; P = 0.012) and higher coupling in the default mode areas (β = 0.093; 95% CI: 0.028 to 0.158; P = 0.005) after full adjustment. The regression coefficients for the relationship between prediabetes and coupling in the limbic and default mode areas were roughly one-third to two-thirds of the coefficient for T2DM. There were no correlations between T2DM or prediabetes and SC-FC coupling in the remaining domains ( Table 2 ).

After full adjustment, fasting glucose and HbA_1c were associated with weaker regional SC-FC coupling in the limbic areas (β = -0.095; 95% CI: -0.189 to -0.001; P = 0.049; and β = -0.106; 95% CI: -0.201 to -0.011; P = 0.029). Fasting insulin was associated with higher coupling in the limbic areas (β = 0.115; 95% CI: 0.021 to 0.209; P = 0.017), and fasting insulin contributed the most to the change of coupling. In addition, HbA_1c was associated with higher coupling in the dorsal attention areas (β = 0.113; 95% CI: 0.018 to 0.209; P = 0.020) ( Table 3 ).

Considering the association of “diabetes-related measures—regional SC-FC coupling” and “regional SC-FC coupling—MoCA score” ( Supplementary Table S2 ), we estimated the single mediating effect of SC-FC coupling levels in limbic areas on the association of “diabetes-related measures—MOCA score”. The study revealed that the suppressing effect of SC-FC coupling in limbic areas on the relationship between fasting insulin and MoCA score was strongest (β_{indirect effect} = -0.019, 95% CI: -0.034, -0.001), with a suppressing effect of 25.3%. In addition, significant indirect effects could also be found in fasting glucose (β_{indirect effect} = 0.018, 95% CI: 0.001, 0.040) and HbA_1c (β_{indirect effect} = 0.018, 95% CI: 0.002, 0.039), corresponding to a suppressing effect of 9.8% and 8.9%, respectively ( Figure 3 , Supplementary Table S3 ).

Tractography algorithms are recognized to underestimate connections between the two hemispheres. When we consider the unilateral brain separately, we found differences in SC-FC coupling regions only in the dominant brain (left brain); T2DM was significantly associated with higher coupling in the default mode areas (β = 0.150; 95% CI: 0.073 to 0.227; P <0.001) compared with NGM ( Supplementary Table S4 ). No associations between prediabetes and T2DM were found with regional SC-FC coupling in the nondominant brain (right brain) ( Supplementary Table S5 ). Finally, we found similar results when we used diabetic status instead of diabetes-related measures to assess whether SC-FC coupling mediates the relationship between diabetic status and cognition ( Supplementary Table S6 ).