This was an extension of our previous cross-sectional study (11). The recruited patients with type 2 diabetes (age: < 40 years) and non-diabetic controls were followed up. Patients had no peripheral microvascular complications such as diabetic retinopathy, diabetic nephropathy, and diabetic peripheral neuropathy at recruitment. Participants with a history of diabetic ketosis or ketoacidosis, hypoglycemia within 48 hours, clinical evidence of cardiovascular or cerebrovascular diseases, a history of alcohol consumption, thyroid dysfunction, anemia, and any MRI contraindications were excluded. After recruitment, patients were provided diabetes education and individualized treatment. All licensed antidiabetic medications were permitted. No target hemoglobin A1c was predetermined. The follow-up visits targeted at 18 months after recruitment. Of the 67 participants (35 patients with type 2 diabetes and 32 controls) recruited at baseline, 50 (74.6%) participants (26 patients with type 2 diabetes and 24 controls) consented to participate in the extensional study. Participants were given separate informed consents for this longitudinal study. The study protocol was reviewed and approved by Shanghai General Hospital Ethics Committee, approval number 2019SQ082.

Cognitive function was assessed at both time-points using the same methodology (11) and by the same physician. Briefly, the cognitive battery included tests for global cognitive function (Montreal Cognitive Assessment [MoCA]), executive function (the accuracy and reaction time of Stroop Color Word Test - part C), memory function (Rey Auditory Verbal Learning Test [RAVLT]), and language function (Verbal Fluency Test [VFT] and Boston Naming Test [BNT]).

Age, sex, education level, and follow-up interval (i.e., the time from baseline to follow-up MRI scan) were recorded for statistical correction. The levels of hemoglobin A1c (HbA1c), fasting plasma glucose, and fasting serum C-peptide were assessed to determine the glycemic control at the two time points. Homeostasis model assessment (HOMA) (14) was used to assess β-cell function (HOMA-%β) and insulin resistance (HOMA-IR). Antidiabetic drugs used by patients with type 2 diabetes at recruitment and follow-up visits were recorded. The laboratory data were collected a day before the MRI scan.

Microvascular complications were assessed in patients with diabetes at the two time points. A dilated fundus examination using a 90-diopter pan-fundus lens was performed by an ophthalmologist. Diabetic retinopathy was defined as the presence of any of the following lesions: retinal microaneurysms, hemorrhages, hard exudates, soft exudates, neovascularization, or evidence (or history) of laser photocoagulation. The urinary albumin-to-creatinine ratio (UACR) was measured in first-void clean-catch urine samples collected on two consecutive days. The average UACR for two consecutive days was used for the analysis.

The baseline and follow-up MRI were performed using a 3-Tesla MRI scanner (Ingenia, Philips Healthcare, Best, NL). fMRI data were acquired using a gradient-echo-planar imaging sequence with the same parameters and instructions (e.g., eyes closed): repetition time (TR), 2000 ms; echo time (TE), 30 ms; field of view, 224 × 224 mm²; flip angle, 90°; slices, 33; slice thickness, 3.5 mm; slice spacing, 0.7 mm; matrix, 64 × 62; volumes, 240; and acquisition time, 8 min. High-resolution, T1-weighted images were obtained using a magnetization-prepared rapid gradient-echo sequence (TR = 7.0 ms, TE = 3.2 ms, flip angle = 7°, inversion time = 1100 ms, and voxel size = 1 × 1 × 1 mm³).

The fMRI data were preprocessed using Data Processing Assistant for Resting-State fMRI software (15) and the following parameters, as in our previous study (11): 1) slice timing; 2) spatial realignment to correct for head motion; 3) spatial normalization of images to the Montreal Neurological Institute (MNI) space using a unified segmentation algorithm (16) and resampling to 3-mm isotropic voxels; 4) spatial smoothing with a 6-mm full-width-half-maximum (FWHM) Gaussian kernel; and 5) linear detrending, bandpass filtering (0.01–0.08 Hz), and regressing out several covariates such as six head-motion parameters and white matter (WM) and cerebrospinal fluid (CSF) signals.

Subjects with both baseline and follow-up data were enrolled for the statistical analysis. The between-group differences at each time point were analyzed using a two-tailed two-sample t-test, Mann–Whitney U test, or chi-square test. Neuropsychological scores were normalized by logarithmic transformation. Thereafter, linear mixed-effects models were performed to compare the longitudinal changes of neuropsychological scores in patients with diabetes and non-diabetic controls between the two time-points, with adjustment for sex, age at baseline, education level and follow-up interval.

Functional connectivity analysis using the Resting-State fMRI Data Analysis Toolkit (REST, http://restfmri.net) was performed using bilateral hippocampus and PCC as the seed region of interest (ROI) in our previous study (11). The three enhanced functional connectivities which were founded in previous between-group analysis were further investigated. The value of each functional connectivity for each subject at both time points was extracted. We first performed between-group analysis at each time point using a two-tailed two-sample t-test. We then performed linear mixed-effects models to compare the longitudinal changes of brain functional connectivity in patients with diabetes and controls between the two time-points, adjusted for sex, age at baseline, education level and follow-up interval.

Finally, to determine the associations between the changes in the three functional connectivities and the changes in neuropsychological performance, we performed linear regression analyses adjusting sex, age at baseline, education level, and follow-up interval in each group. The associations between changes in the aforementioned connectivity and changes in hyperglycemia-related variables (i.e., HbA1c level, fasting plasma glucose level, fasting serum C-peptide level, HOMA-%β, and HOMA-IR) were also determined in patients with diabetes.

We performed VBM analysis on T1-weighted images using the VBM8 toolbox in the SPM8 software package (http://www.fil.ion.ucl.ac.uk/spm/). Cerebral tissues from each participant were segmented into gray matter (GM), WM, and CSF. Images of each participant’s GM, WM, and CSF were nonlinearly registered using the DARTEL method and transformed into the MNI standard space. The warped GM images were modulated by Jacobian determinants to include the information of volume in these modulated images. Finally, the resultant maps were smoothed with a 6-mm FWHM Gaussian kernel. In addition, the GM, WM, CSF, and total brain volumes for each participant were obtained simultaneously.

At the voxel level, between-group differences in the GM volume were determined using random-effects two-sample t-tests with age, sex, educational level, and BMI as nuisance covariates. The threshold was set at a corrected P value of < 0.05, with multiple comparisons corrected using the AlphaSim program (http://afni.nimh.nih.gov/pub/dist/doc/manual/AlphaSim.pdf) determined by 1000 Monte Carlo simulations (i.e., single voxel P < 0.001, combining a minimum cluster size). In addition, we performed between-group analysis for the mean volumes of GM, WM, and CSF, as well as total brain volume.

The mean follow-up intervals for patients and controls were 18.0 ± 5.5 and 18.0 ± 4.9 months respectively (P = 0.891). The characteristics of subjects who withdrew from the study are presented in Supplemental Table 1 .

Compared with the controls, the patients with type 2 diabetes had higher body mass index (P = 0.017), HbA1c level (P < 0.001), fasting blood glucose (P = 0.011), HOMA-%β (P = 0.003), and lower HDL cholesterol level (P = 0.002) at the follow-up visit as expected. However, there were no differences in the level of HOMA-%β (P = 0.965), triglyceride (P = 0.083), and uric acid (P = 0.129) at the follow-up visit, and these variables were statistically different between groups at baseline. Notably, the number of patients who were taking antidiabetic medicines increased from 42.3% to 92.3% from baseline to follow-up. A large proportion of patients were using metformin (73.1%) and/or glucagon-like peptide 1 receptor agonists (42.3%) at the follow-up visit. The demographic and clinical data of the participants are shown in Table 1 .

The between-group differences in the neuropsychological tests at baseline are consistent with our former study, which revealed longer Stroop Reaction Time in patients with type 2 diabetes than in controls (P = 0.033 in the current sample), indicating poorer executive function in the patients group. At the follow-up visit, interestingly, patients with type 2 diabetes had a higher MoCA than the controls (P = 0.039), which indicated better global cognitive function in the patients group. No significant differences were noted in any other neuropsychological performances between the two groups at the follow-up visit ( Figure 1 ).

The linear mixed-effects model demonstrated that the longitudinal trajectories of log-transformed MoCA and RAVLT differed significantly between the two groups. Compared with controls, patients with diabetes had a 1.5% greater incremental rate of MoCA score (β = 0.015, 95% CI: 0.00013 – 0.030) and a 5% greater incremental rate of RAVLT (β = 0.050, 95% CI: 0.0002 – 0.100) on a logarithmic scale from baseline to follow-up visit. Furthermore, significant increase of log-transformed Stroop Accuracy ((β = -0.009, 95% CI: -0.016 – -0.002) and decrease of Stroop Reaction Time ((β = 0.030, 95% CI: 0.010 – 0.049)) were observed among patients with diabetes, which indicated that patients had better executive function at follow-up compared to that at baseline.

Baseline between-group analyses revealed that the three functional connectivities were significantly higher in patients with type 2 diabetes than in the controls (P < 0.001 for the connectivity of the left hippocampus with the left IFG and the left IPL; P = 0.001 for the connectivity of the PCC with the left IPL), which were also consistent with our former study. At the follow-up visit, however, the between-group differences disappeared ( Figure 2 ).

Moreover, the linear mixed-effects model demonstrated the longitudinal changes of the connectivity of the left hippocampus with the left IFG and the left IPL differed significantly between diabetic group and controls. Compared with controls, patients with diabetes had a 20.7% greater decremental rate in the connectivity of the left hippocampus with the left IFG (β = -0.207, 95% CI: -0.334 – -0.080) and a 16.0% greater decremental rate in the connectivity of the left hippocampus with the left IPL (β = -0.160, 95% CI: -0.271 – -0.048) from baseline to follow-up visit ( Table 2 ).

Considering that the antidiabetic drugs might have a direct effect on brain functional connectivity (7), we further divided the patients into two subgroups according to whether they took antidiabetic drugs at baseline (Medicated subgroup, N = 11; non-Medicated subgroup, N = 15) to obtain a comprehensive understanding of the brain neurophysiology in young patients with type 2 diabetes. Between-subgroup analyze demonstrated that the functional connectivity between PCC and left IPL was lower in medicated subgroup compared with that in non-medicated subgroup at baseline. No between-subgroup differences were observed in cognitive performance at baseline, the changes of cognitive performance and the changes of functional connectivity from baseline to the follow-up visit ( Supplemental Table 2 ). The decreased intrinsic functional connectivity in medicated subgroup at baseline provided some clues to the further research on the effect of antidiabetic drugs on brain function.

No significant between-group differences were observed in brain morphometric analyses (both voxel-based and global brain volume analyses) at both time-points. ( Supplemental Table 3 ).

In patients with type 2 diabetes, the change in Stroop Accuracy was negatively correlated with the change in functional connectivity between the left hippocampus and the left IPL after adjusting for age at baseline, sex, education level, and follow-up interval (β = -0.441, P = 0.045). This indicated that the greater the decline in the hippocampal connectivity, the greater the improvement of executive function. Moreover, the change in MoCA was negatively correlated with the change in functional connectivity between the PCC and left IPL after adjusting for the aforementioned potential confounders (β = −0.486, P = 0.037), indicating that the greater the decline in PCC connectivity, the greater the improvement of global cognition. In contrast, no association was observed between the changes in functional connectivity and the changes in cognitive performance in the controls ( Figure 3 ).

The associations between the changes in the three functional connectivities and the changes in the HbA1c level as well as other hyperglycemia-related variables (fasting plasma glucose level, fasting serum C-peptide level, HOMA-%β, and HOMA-IR) were also investigated in patients with type 2 diabetes. However, no significant associations were found after adjusting for potential confounders ( Supplemental Table 4 ).