This study was approved by the Ethics Committee of Shaoxing Second Hospital. The data were anonymous, so informed consent was waived. From January 2018 to January 2021, data from 4327 hospitalized patients with abnormal blood glucose levels were retrospectively analysed. After reviewing clinical information, a total of 549 patients with newly diagnosed T2D were included in the study. The inclusion criteria were as follows: 1) fasting plasma glucose (FPG) ≥ 7.0 mmol/l, 2-hour postprandial glucose (2hPG) ≥ 11.1 mmol/l, and/or glycated haemoglobin (HbA1c) ≥ 6.5% (19); 2) abnormal blood glucose duration ≤ six months; and 3) age between 40 and 70 years old. The exclusion criteria were as follows: 1) history of antidiabetic drug use; 2) history of lipid-lowering treatment; 3) history of arteriosclerotic cardiovascular disease or severe renal dysfunction; 4) history of malignancy; and 5) lack of carotid ultrasonography or chest CT examination. Arteriosclerotic cardiovascular disease includes stroke, transient ischemic attack, coronary heart disease, heart failure, and arterial occlusion (3). We randomly divided the patients into a training set and a validation set at a ratio of 7:3. The details of patient selection are presented in Figure 1 .

Data on sex, age, height, weight, waist circumference, tobacco use, alcohol use and history of hypertension were collected. Body mass index (BMI) was calculated as weight (kg) divided by the squared value of height (m²). Hypertension was defined as blood pressure (BP) ≥140/90 mmHg on two different occasions or a history of antihypertensive treatment.

Clinical laboratory test data, such as triglycerides (TGs), total cholesterol (TC), triglycerides/high-density lipoprotein cholesterol (TG/HDL-C), FPG, and HbA1c, were also recorded. The TyG index was calculated by the equation log [(fasting TG (mg/dl) × FPG (mg/dl)/2] (20).

The extent of arteriosclerosis in the participants was assessed by carotid ultrasonography according to the literature (3). Briefly, for each participant, the unilateral carotid artery was divided into the common carotid region, internal carotid region, external carotid region, and bifurcation carotid region. The presence of plaques in each region was scored as 1 (single plaque), 2 (multiple plaques), or 3 (stenosis). The plaque score (PS) was calculated by summing the scores of both carotid arteries. Based on the PS, the severity of arteriosclerosis was classified into no plaque burden (PS = 0), low plaque burden (PS < 3), and high plaque burden (PS ≥ 3). The ultrasound images were independently reviewed by two experienced sonographers who had no knowledge of the clinical data. Any discrepancies were resolved by consensus.

Radiomics IMAT analysis was proceeded on chest CT images obtained by using a 64-row or 16-row multidetector CT scanner (SOMATOM Definition AS, Siemens Medical Solutions, and BrightSpeed, GE Healthcare). The following CT parameters were applied: rotation time, 500 milliseconds; voltage, 120 kVp; automatic exposure control, and 2.5 mm reconstructed section thickness.

The acquired CT images were then imported into ITK-SNAP software (www.itksnap.org) to delineate the regions of interest (ROIs) of IMAT in the section above the aortic arch by two independent experienced radiologists who had no knowledge of the arteriosclerosis degree (13). To ensure the consistency of the data, the same delineate criterion was applied, and another expert was responsible for conforming the segmentation. After image segmentation, Z score normalization was applied to standardize the CT images. To assess the reproducibility of image segmentation, two months later, 50 patients were randomly selected, and the IMAT ROIs were delineated again by radiologists to build a resegmentation set. Then, intraobserver and interobserver repeatability were evaluated by using intraclass and interclass correlation coefficients (ICCs).

After the image segmentation, the radiomics features of IMAT were extracted with the pyradiomic platform (https://keyan.deepwise.com) (21). In the preprocessing stage, Z score normalization was used to process the images with a normalization scale of 100, and the B-spline interpolation sampling method was used to resample CT images to the same resolution. Eight kinds of filters (wavelet, Laplacian of Gaussian, square, square root, logarithm, exponential, gradient transform, and local binary pattern transform) were applied to preprocess the CT images. In total, 1,316 radiomics features of IMAT were extracted from the ROIs of each CT image: 252 first-order features, 14 shape features, 336 greyscale co-occurrence matrix (GLCM), 224 grey-level size zone matrix (GLSZM), 224 grey-level run length matrix (GLRLM), 196 grey-level distance-zone matrix (GLDM), and 70 neighbourhood grey-tone difference matrix (NGTDM).

Least absolute shrinkage and selection operator (LASSO) regression was used to select the extracted radiomic features of IMAT that were highly correlated with arteriosclerosis. First, 260 features with unique values were excluded. Pearson correlation analysis was used to estimate the correlation between the remaining features. Features with a correlation coefficient under 0.90 were excluded. After dimensionality reduction of features, the 144 features were included in subsequent modelling.

Both clinical data and radiomics features were applied for model construction in indicating arteriosclerosis in patients with newly diagnosed T2D. In terms of clinical elements, univariate and multivariate logistic regression were conducted to explore the relationship between clinical elements and the carotid plaque burden, which represented arteriosclerosis. Then, clinical elements with P < 0.05 in multivariate analysis were included to establish model 1. With regard to the radiomics model, the L1-based method was used for IMAT radiomics feature selection. By summing the included features weighted by their coefficients, a radiomic score (Rad-score) formula (model 2) could be constructed. A combined clinical-radiomics model (model 3) was developed by integrating the clinical elements and the radiomics signature and presented in the form of a nomogram.

A confusion matrix was used to quantify the performances of model 1, model 2, and model 3. Furthermore, receiver operating characteristic (ROC) curves were generated to quantify the clinical usefulness of the three models. Calibration curves were plotted to determine the discrimination ability of the radiomic-clinical nomogram for the training and validation sets. The usefulness of the radiomic-clinical nomogram was assessed by the net benefits in different threshold probabilities by decision curve analysis (DCA). Figure 2 presents the flowchart of the study.

All statistical analyses were performed with SPSS (version 26.0), MedCalc (vision 19.5.6), and R software (version 4.0.2). Normality was evaluated by the Kolmogorov-Smirnov test. Data with a nonnormal distribution were expressed as medians (interquartile ranges) for continuous variables and as percentages for categorical variables. Group differences were evaluated by the Wilcoxon rank sum test for continuous variables and by the chi-square test for categorical variables. Pearson’s chi-squared test was used to identify the difference between the training and validation sets. The performance of the clinical model, radiomics model, and combined clinical-radiomics model was quantified by the area under the curve (AUC) with 95% confidence intervals (95% CIs). The DeLong test was employed to determine whether significant differences existed in the AUC values of the three models. We also calculated the ICCs to evaluate the agreement of extracted radiomics features by two radiologists. Kappa test analyses were used to determine the intra- and interobserver agreement. P values ≤ 0.05 were considered statistically significant.

The baseline characteristics of the patients with or without plaque burden are compared in Table 1 . Patients with plaque burden were older, more likely to be male, had more hypertension, and were more likely to be current smokers (57.00 [52.00-63.00] vs. 52.00 [46.00-58.00)] years, P < 0.001; 70.60% vs. 61.80%, P =0.029; 51.70% vs. 36.80%, P < 0.001; 49.80% vs. 37.10%, P =0.003, respectively). However, there were no differences in the physical examination results (such as BMI and waist circumference) and laboratory findings (such as TGs, TC, FPG, and the TyG index) between the groups.

In terms of clinical elements, age and smoking history were independent risk factors associated with carotid plaque burden in patients with newly diagnosed diabetes, while age, smoking history, and TyG index were independent risk factors associated with high carotid plaque burden (PS ≥ 3) in patients with newly diagnosed diabetes. The results of the logistic regression analysis are shown in Table 2 .

A logistic regression classifier was established according to the selected clinical characteristics. In all subjects included in the study, the AUC of the training set was 0.687 (95% CI: 0.634-0.730), the accuracy rate was 0.656, the sensitivity was 0.617, and the specificity was 0.694. The AUC of the validation set was 0.721 (95% CI: 0.642-0.799), the accuracy rate was 0.685, the sensitivity was 0.691, and the specificity was 0.679. In subjects with a high plaque burden, the AUC of the training set was 0.755 (95% CI: 0.683-0.826), the accuracy rate was 0.697, the sensitivity was 0.629, and the specificity was 0.737. The AUC of the validation set was 0.620 (95% CI: 0.490-0.750), the accuracy rate was 0.654, the sensitivity was 0.633, and the specificity was 0.667.

ICCs that represented the intraobserver and interobserver consistency of the feature extraction were calculated. A total of 1270 stable features with ICCs greater than 0.75 were retained for subsequent analysis. After applying the LASSO algorithm in the training set, ten features associated with the present of arteriosclerosis and ten features associated with the severity of arteriosclerosis were selected from the extracted features. These twenty features included eight first-order features and twelve texture features, which can be seen in Table 3 and Supplementary Figure S1 . Based on the selected features, the linear regression (LR) algorithm was used to construct the radiomics model and to calculate the Rad-score (Supplementary Appendix S1).

For all subjects included in the study, the AUC of the training set was 0.932 (0.907-0.957), the accuracy rate was 0.862, the sensitivity was 0.872, and the specificity was 0.852. The AUC of the validation set was 0.927 (0.890-0.964), the accuracy rate was 0.842, the sensitivity was 0.840, and the specificity was 0.845. In subjects with a high plaque burden, the AUC of the training set was 0.734 (0.663-0.805), the accuracy rate was 0.670, the sensitivity was 0.771, and the specificity was 0.610. The AUC of the validation set was 0.698 (0.582-0.814), the accuracy rate was 0.642, the sensitivity was 0.567, and the specificity was 0.686.

Nomograms including clinical factors and Rad-score are shown in Figure 3 . The formulas to assess the risk probability for atherosclerosis are available in Supplementary Appendix S2. The performances of the three models are presented by a confusion matrix. The false-positive and false-negative rates in model 3 were lower than model 1 in both the training sets and validation sets ( Figure 4 , Supplementary Table S1 ). In all subjects included in the study, the AUC of model 3 was higher than that of model 1 [0.934 (0.909, 0.959) vs. 0.687 (0.634, 0.730), P < 0.001 in the training set, 0.933 (0.898, 0.969) vs. 0.721 (0.642, 0.799), P < 0.001 in the validation set]. Similar indicative efficacy was found between model 3 and model 2 (P = 0.5694). In subjects with high plaque burden, the AUC of model 3 was higher than that of both model 1 and model 2 [0.824 (0.765, 0.882) vs. 0.755 (0.683, 0.826) and 0.734 (0.663, 0.805), P < 0.001 in the training set, 0.717 (0.604, 0.830) vs. 0.620 (0.490, 0.750) and 0.698 (0.582, 0.814), P < 0.001 in the validation set, respectively]. Table 4 lists the AUC, accuracy, sensitivity and specificity of the three models. The comparison of the ROC curves of the three models is presented in Figure 5 .

The calibration curve along with the H-L test demonstrated good consistency between the observed carotid plaque burden and indicated arteriosclerosis in both the training and validation sets ( Figures 3C–F ). DCA showed that model 3 and model 2 had better performance than model 1 for all subjects included in the study in both training set and validation set, as shown in Figure 6 . However, for subjects with a high plaque burden, model 3 had higher efficacy than the other two models in the training set and showed no significant differences in the validation set.