The study was conducted in accordance with the Declaration of Helsinki, and approved by Ethics Committee of Nanjing First Hospital (protocol code KY20220124-02 and date of approval 24 January 2022). A total of 190 individuals aged between 40 and 70 years, including 92 patients with T2DM and 98 healthy controls (HCs), were recruited from the Department of Endocrinology, Nanjing First Hospital (public hospital), between February 2022 and December 2023. The study sample size was calculated by G*Power software (version 3.1.9.7). All participants were right-handed and educated for at least 8 years; all participants have health insurance. All patients with T2DM were diagnosed according to the criteria of the American Diabetes Association [18]. The regimen for patients with T2DM included oral hypoglycemic drugs (n = 71), insulin (n = 48), glucagon-like peptide 1 receptor agonist (GLP-1RA) (n = 19). Sex, age, and education-matched HCs were recruited through community health inspections or online advertising during the same period. We measured the fasting glucose levels and postprandial glucose levels and excluded individuals with a fasting glucose level greater than 6.1 mmol/l or postprandial glucose level greater than 7.8 mmol/l. Participants were not excluded from the fMRI analysis because of excessive head motion during the scan.

In order to minimize the impact of potential confounding variables on the main measures of the study, participants were excluded from the study if they had head injury, Parkinson’s disease, epilepsy, stroke, alcoholism, major depression or other neurological or psychiatric illness, severe vision or hearing loss, or other major medical illness.

Demographic data were recorded including sex, age, education levels, weight, height, and body mass index (BMI) [weight in kg/(height in m)2]. Two separate assessments were employed to determine the average value. Blood samples were obtained through venipuncture at 8 A.M. following an overnight fast of a minimum of 10 h. Fasting blood glucose, fasting serum C-peptide, HbA1c, triglycerides, total cholesterol, low-density lipoprotein cholesterol (LDL-C), and high-density lipoprotein cholesterol (HDL-C) levels were measured.

Classical neuropsychological assessments were executed to evaluate affective and cognitive symptoms in all participants. The Montreal Cognitive Assessment (MoCA) [19] and Mini Mental State Exam (MMSE) [20] are routine cognitive screening tests that assess possible dementia and MCI. The Auditory Verbal Learning Test (AVLT) and AVLT-delay [21] are commonly used methods for assessing episodic verbal learning. The Complex Figure Test (CFT) and CFT-delay [22] assess visual perception, visuomotor integration, and visuospatial recall. The Digit Symbol Substitution Test (DSST) [23] is commonly used to assess the ability to switch attention and information processing. The Digital Span Test (DST) [24] is a measurement of attention and short-term memory. The Trail-Making Test A and B (TMT-A and B) are employed to evaluate executive control, which includes structured visual search, focused attention, and cognitive adaptability. The Clock Drawing Test (CDT) is utilized to appraise executive function and visual-spatial ability. All neuropsychological tests were conducted by two trained and experienced physicians.

All subjects were scanned using a 3.0-T MRI scanner (Ingenia, Philips Medical Systems, Netherlands) with an 8-channel phased-array head coil. Foam padding and earplugs were used to minimize involuntary head motion and reduce noise. Participants were directed to remain motionless with their eyes shut, endeavoring not to focus on any specific thoughts and to prevent head movement throughout the scanning process.

Functional images were obtained with a gradient echo-planar imaging (EPI) sequence with the following parameters: TR/TE, 2000 ms/30 ms; slices, 36; flip angle (FA), 90°; thickness, 4 mm; gap, 0 mm; field of view (FOV), 240 mm × 240 mm; matrix size, 64 × 64; and voxel size, 3.75 mm × 3.75 mm × 4.0 mm. The scan lasted for 8 min and 8 s. Structural images were obtained with a high-resolution T1-weighted (T1W) sequence, namely three-dimensional turbo fast echo (3D-TFE), with the following scan parameters: TR/TE, 8.1 ms/3.7 ms; FA, 8°; FOV, 256 mm × 256 mm; slice, 170; thickness, 1 mm; gap, 0 mm; and matrix size, 256 × 256. In addition, both scans were acquired utilizing sensitivity encoding (SENSE) technology for parallel imaging with a SENSE factor of 2.

The rs-fMRI data were pre-processed using GRETNA software on the MATLAB 2013b platform. The DICOM images were first converted to NIFTI files, and the first ten time points were removed for signal equilibrium. In the subsequent step, 220 remaining volumes underwent correction to calculate the time delay in data acquisition between slices (slice timing) and to adjust for any head movement during scanning (realignment).

The spDCM analysis was conducted using the DCM toolbox integrated within SPM12 to identify the effective connectivity among brain regions during the resting state. First, a first-level analysis was performed, and a general linear model with six rigid body motion parameters, WM signal, and CSF signal as interference covariates was applied to the time series. In order to calculated the DCM model from the preprocessed fMRI data, the next we defined the region of interest (ROI) and extract the time series.

Eight ROIs were selected as seeds matching the 6-mm spherical ROIs used in prior research examining the seed-to-voxel functional connectivity (FC) of these networks [25,26,27]. The DMN hubs were the PCC (1, − 61, 38) and MPFC (1, 55, − 3); the SN hubs were the left anterior insula (l-AINS) (− 44, 13, 1), right anterior insula (r-AINS) (47, 14, 0), left supramarginal gyrus (l-SMG) (− 60, − 39, 31), and right supramarginal gyrus (r-SMG) (62, − 39, 31); and the ECN hubs were the left lateral prefrontal cortex (l-LPFC) (− 43, 33, 28) and right lateral prefrontal cortex (l-LPFC) (41, 38, 30).

The primary eigenvariates of the time series were derived from the aforementioned ROIs. A gray matter (GM) mask, set at a threshold of 0.5, was applied to the time series data within the standard MNI space in order to exclude any extraneous contributions from white matter (WM) and cerebrospinal fluid (CSF).

Dynamic causal modeling analysis was conducted on the specified ROIs, wherein each subject had bidirectional connections established between every pair of ROIs, resulting in a fully interconnected model with each node linked to all others. Subsequently, the parameters for this comprehensively connected model were estimated. The process of model estimation relied on the parametric empirical Bayes (PEB) framework [28]. Instead of estimating each individually, a single spDCM model was jointly estimated and then underwent Bayesian model reduction. Finally, the inherent connectivity strength was extracted from the estimation results, i.e., the effective connectivity value.

Differences in demographic characteristics, clinical data, and neuropsychological variables between HCs and patients with T2DM were analyzed with two-sample t tests for means and χ2 tests for proportions (statistically significant difference was set at p < 0.05). For the spDCM analysis, to compare the difference in mean connection strength between the two groups, an independent two-sample t test was performed after adjusting for the effects of age, sex, and education, with the significance level set at p < 0.05 (FDR correction). Additionally, to examine the correlation between abnormal causal connectivity strength and clinical variables in individuals with T2DM, Pearson’s or Spearman’s correlation coefficients were calculated utilizing the SPSS 26.0 software package. Furthermore, correlation analyses using Spearman’s or Pearson’s methods were conducted to assess the relationships between neuropsychological assessment scores and effective connectivity strength, with a p value less than 0.05 deemed to signify statistical significance.