Prediction of stroke events in patients with type 2 diabetes mellitus by interpretable machine learning based on perivascular adipose tissue features: a multicenter cohort study

Introduction

Ischemic stroke (IS) is among the leading causes of death and disability worldwide (1). The annual incidence of IS is approximately 101.3 per 100,000 individuals, placing a significant economic burden on society and families while also causing considerable psychological stress (2). Approximately 20% of IS cases are caused by carotid atherosclerotic plaques (3). Type 2 diabetes mellitus (T2DM), a metabolic disorder primarily driven by insulin resistance, is a common condition strongly associated with cardiovascular and cerebrovascular complications (4-6). Studies have shown that compared to the general population, patients with T2DM are 2–5-fold more likely to experience an IS (7,8). Additionally, the mortality, disability, and recurrence rates after stroke are significantly higher in patients with T2DM than in those without it (9,10). Therefore, accurate assessment of the risk of carotid plaques in patients with T2DM is crucial for predicting the occurrence of IS and may facilitate the development of more effective management and prevention strategies.

The mechanisms linking carotid atherosclerosis to stroke in patients with T2DM are associated with various factors, including inflammation, age, obesity, smoking, dyslipidemia, and other stress markers (11,12). In recent years, an increasing abundance of research has identified inflammation as a key driver of atherosclerotic plaque progression and rupture (13,14). Inflammatory mediators within plaques can diffuse through the vascular wall into surrounding adipose tissue, promoting lipolysis and inhibiting the differentiation of preadipocytes, resulting in a higher water-to-lipid ratio in perivascular adipose tissue (PVAT) (15,16). Studies have shown that perivascular fat density (PFD) can be used to identify early subclinical carotid disease and vulnerable plaques, while the radiological parameters of PVAT can predict high-risk carotid plaques and differentiate patients at different stages of carotid atherosclerosis (17-19). However, in current clinical practice, the assessment of carotid atherosclerotic plaques is still largely limited to the measurement of the PFD around carotid plaques (20). This approach fails to fully capture the complex three-dimensional structure of adipose tissue and does not adequately reflect the finer spatial relationships between voxels (21,22). Therefore, identifying higher-dimensional PVAT features is crucial for assessing the risk of IS in patients with T2DM.

Radiomics-based machine learning (ML) is an emerging field that leverages imaging features to objectively and quantitatively identify lesion phenotypes (23). Initially developed for oncology research, this approach has recently been increasingly applied in studies examining cardiovascular and cerebrovascular events (24,25). ML holds significant potential for predicting adverse cardiovascular events, monitoring disease progression, and assessing prognosis (26,27). However, only a few studies have examined the application of PVAT in carotid plaques (28,29). Thus far, radiological models based on PVAT specifically designed for predicting IS risk in patients with T2DM are lacking.

Therefore, we conducted a multicenter study with the aim of developing and validating an ML-based model that integrates radiological features of PVAT with clinical and radiological data to construct an IS risk prediction model for patients with T2DM and evaluate its potential utility in stratifying patients’ risk levels during follow-up. This risk prediction model was designed to support effective personalized interventions and to reduce the risk of stroke onset through early intervention, ultimately alleviating the social and individual burden of stroke. We present this article in accordance with the TRIPOD+AI reporting checklist (available at https://qims.amegroups.com/article/view/10.21037/qims-2025-1760/rc).

Methods

Study population

This retrospective, multicenter study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments, and was approved by the Ethics Committee of The Fifth Affiliated Hospital of Wenzhou Medical University (No. 2025-136). The requirement for informed consent was waived due to the retrospective nature of the analysis. Computed tomography angiography (CTA) data related to carotid atherosclerosis were retrospectively collected for all patients with T2DM treated between January 2015 and January 2024 at The Fifth Affiliated Hospital of Wenzhou Medical University (Center 1), between January 2017 and January 2023 at The Second Affiliated Hospital of Wenzhou Medical University (Center 2), and between January 2016 and January 2024 at the Lishui People’s Hospital (Center 3). The inclusion criteria, exclusion criteria, and clinical data collected are detailed in Appendices 1,2. The patient selection process and the distribution of the training set and validation set are illustrated in Figure 1. The sample size estimation is described in Appendix 3.

Figure 1 Flowchart of patient selection. Center 1, The Fifth Affiliated Hospital of Wenzhou Medical University; Center 2, The Second Affiliated Hospital of Wenzhou Medical University; Center 3, Lishui People’s Hospital. CTA, computed tomography angiography; MRI, magnetic resonance imaging.

Two radiologists independently collected clinical and imaging data based on the inclusion and exclusion criteria and grouped patients according to their magnetic resonance imaging (MRI) findings. IS was defined according to imaging evidence of ischemia in the carotid artery territory and the presence of clinical symptoms lasting at least 24 hours. Patients with positive MRI findings (indicating acute or subacute IS) within 2 weeks before the head and neck CTA examination were assigned to the stroke group. Patients with negative brain MRI results were assigned to the non-stroke group. Ultimately, 494 patients were included, with 167 patients in the stroke group and 327 in the non-stroke group. Patients from Center 1 were randomly divided into an internal training set (n=193) and an internal validation set (n=84) in a 7:3 ratio, while those from Centers 2 and 3 were placed into external validation set 1 (n=105) and external validation set 2 (n=102), respectively.

Image acquisition, preprocessing, and PVAT segmentation

CTA examinations in Centers 1 and 2 were performed with a SOMATOM Force computed tomography (CT) scanner (Siemens Healthineers, Erlangen, Germany), while those in Center 3 were performed with an Aquilion One CT scanner (Toshiba Medical Systems, Otawara, Japan). Detailed information on the CTA acquisition parameters is provided in Table S1. The preprocessing of CTA images, delineation of the PVAT, and measurement of PFD are described in Appendices 4,5. The degree of stenosis was assessed via CTA according to the North American Symptomatic Carotid Endarterectomy Trial criteria (30).

Radiomics feature extraction and selection

A total of 1,688 radiomic features were extracted from the PVAT through the use of RadCloud version 7.1 (Huiying Medical Technology Co., Ltd., Beijing, China; http://radcloud.cn) and categorized into four groups (Appendix 6). Subsequently, all feature values were standardized before feature selection, and the intraclass correlation coefficient (ICC; Appendix 7) was used to evaluate the reproducibility of the features obtained from the PVAT segmentation. Next, in the internal training set, features were sequentially screened via variance thresholding, SelectKBest, and least absolute shrinkage and selection operator (LASSO) regression to identify radiomics features associated with stroke (Appendix 8). Finally, radiomics scores (Radscores) were calculated based on the features with nonzero coefficients in the LASSO model.

Model development, evaluation, and visualization

The clinical variables underwent univariate logistic regression (LR) analysis, and those with a P value of <0.05 were included in the multivariate LR analysis to identify independent clinical risk factors for developing the clinical model. For PFD, we directly developed an LR model due to the smaller number of input features. In order to identify the optimal ML model, six classifiers—k-nearest neighbors (KNN), support vector machine (SVM), LR, linear discriminant analysis (LDA), naive Bayes (NB), and extreme gradient boosting (XGBoost)—were employed in the training set by linearly combining features with nonzero coefficients, which were weighted by their respective coefficients. The optimal combinations of hyperparameters for each classifier in the internal training set are listed in Appendix 9.

The models’ performance was evaluated through various metrics, including the area under the receiver operating characteristic curve (AUC), sensitivity, specificity, and accuracy. The ML model that had the highest average AUC in the training set was selected as the optimal radiomics model, and the results were converted into a Radscore. Following this, the Radscore, clinical variables, and PFD were integrated to construct the combined model. Calibration curves were constructed to assess the agreement between observed outcomes and predicted risks, and decision curve analysis (DCA) was used to calculate the models’ net benefit. Additionally, the SHapley Additive exPlanation (SHAP) method was used to visually and analytically interpret the models’ prediction process (Appendix 10). The workflow of the radiomics methodology used in this study is illustrated in Figure 2.

Figure 2 Schematic representation of the radiomics analysis steps. AI, artificial intelligence; AUC, area under the receiver operating characteristic curve; CTA, computed tomography angiography; FAI, fat attenuation index; PFD, perivascular fat density; T2DM, type 2 diabetes mellitus; TC, total cholesterol; VOI, volume of interest.

Follow-up and survival analysis

After discharge, 494 patients with carotid atherosclerotic plaques were followed up through a review of medical records or telephone interviews with the patients or their relatives. The follow-up assessments preferentially used CTA as the imaging modality, with follow-ups conducted every 3–6 months. The median follow-up period was 25 months (range, 0.5–36 months). Follow-up outcomes were recorded as the time from discharge to the first occurrence of a new ischemic event (including transient ischemic attack or stroke). Kaplan-Meier curves were used to visualize differences in the follow-up event rates between groups. In the validation set, we performed Cox proportional hazards regression to assess the association of the model-predicted risk groups with follow-up events.

Statistical analysis

The data were statistically analyzed with R statistical software v4.3.1 (The R Foundation for Statistical Computing, Vienna, Austria) and Python v3.7.6 (Python Software Foundation, Wilmington, DE, USA). The normality of continuous variables was evaluated with the Kolmogorov-Smirnov test. Normally distributed continuous variables are reported as the mean ± standard deviation and were compared between groups via the independent samples t-test. Nonnormally distributed continuous variables are reported as the median and interquartile range and were compared with the Mann-Whitney test. Categorical variables are reported as frequencies and were compared between groups via the Chi-squared test. The DeLong test was used to compare the AUCs of the different ML models, and the corresponding P values were calculated. A P value of <0.05 was considered statistically significant. Univariate and multivariate analyses were conducted via the Cox proportional hazards regression model to identify predictors of stroke events during the follow-up period, with statistical significance set at P<0.05. Survival curves were plotted through the use of the Kaplan-Meier method and were compared with the log-rank test.

Results

Clinical model and PFD model construction

Based on the PFD surrounding the carotid plaque, a PFD model was established through LR, with its coefficients being used as weights. The patients’ clinical characteristics are summarized in Table 1 and did not differ significantly between the training and validation sets (Table S2). Univariate and multivariate LR analyses, hypertension, and total cholesterol were identified as independent risk factors and were subsequently used to develop the clinical model (Table 2).

Feature selection and radiomics model construction

The intra- and interobserver agreement for the radiomics features selected from the PVAT of the same lesion was satisfactory, with ICCs ranging from 0.876 to 0.982 and from 0.831 to 0.929, respectively. Therefore, all extracted radiomic features were retained for further analysis. In the training cohort, seven significantly correlated features were obtained through variance threshold, SelectKBest, and LASSO: three wavelet features, three first-order features, and one texture feature (Figure S1). Details regarding the PVAT features are provided in Table S3 and Figure S2. Pairwise correlations between the selected features were assessed with the Pearson correlation coefficient (r), and the results were visualized as a correlation matrix heatmap (Figure S3). Finally, all selected features were combined and used to construct the radiomics model.

Construction and evaluation of the ML classifiers

The predictive performance of the radiomics models developed with the six ML classifiers is summarized in Table 3 and Figure S4. The DeLong test was applied to the six ML classifiers (Figure S5). As shown in Figure S6, the comprehensive predictive metrics of the XGBoost classifier were superior to those of the other classifiers. Therefore, XGBoost was identified as the most suitable ML classifier for constructing the radiomics model, and the model outputs were used to calculate the Radscore. Figure S7 shows the distribution of Radscores in the stroke and non-stroke groups in each cohort based on the combined model. In all cohorts, the Radscores were significantly higher in the stroke group than in the non-stroke group (all P values <0.001).

Combined model evaluation

Table 4 presents the AUC, accuracy, sensitivity, and specificity of the different models for predicting IS in patients with T2DM. Figure 3A provides a comparison of the prediction performance of the combined model and the clinical, PFD, and radiomics models, showing that the combined model exhibited superior predictive performance. The combined model demonstrated superior accuracy and robustness across all cohorts (Figure 3B). Although the combined model performed slightly better than did the radiomics model, the difference was not significant (detailed results from the DeLong test are shown in Figure S8).

Figure 3 Performance evaluation of different models in the four sets. (A) AUC curves for the clinical, PFD, radiomics, and combined models. (B) Radar chart of the predictive performance of the clinical, PFD, radiomics, and combined models in the four cohorts. (C) Calibration curves for the clinical, PFD, radiomics, and combined models in the four cohorts. (D) DCA of the clinical, PFD, radiomics, and combined models in the four cohorts. AUC, area under the receiver operating characteristic curve; DCA, decision curve analysis; PFD, perivascular fat density.

There was a high degree of agreement between the predicted and actual outcomes across sets for the combined model according to the calibration curves, and there was good consistency between clinical observations and model predictions (Figure 3C). The DCA showed that the combined model has a significant advantage in clinical decision-making compared to the other models (Figure 3D).

SHAP analysis of the combined XGBoost model

The SHAP bar chart in Figure 4A illustrates the magnitude of each feature’s contribution to the model’s predictive performance. The SHAP summary plot in Figure 4B provides an intuitive visualization of each feature’s impact on the model’s predictions, with positive contributions shown in purple and negative contributions shown in yellow. For example, the purple dots with higher feature values for Exponential_glszm_SizeZoneNonUniformity indicate that the variation in the size zones of grayscale areas in the image is more uneven, suggesting that heterogeneous tissue characteristics may be associated with an increased risk of IS. In contrast, the yellow dots with lower feature values imply a reduced risk of IS. These low-value features decrease the forward prediction in the model, as reflected by the negative SHAP values on the X-axis, indicating a prediction of non-stroke.

Figure 4 SHAP visualization of the overall model. (A) A bar chart of the importance of all features in the model based on SHAP values. (B) A SHAP summary plot showing each feature’s positive (purple) and negative (yellow) contributions to the predicted probability. (C) A SHAP heatmap showing the direction and magnitude of the impact of each feature across all cases in the model. (D) A SHAP decision plot showing each important feature’s effect on the final predicted probability for a single case. PFD, perivascular fat density; SHAP, SHapley Additive exPlanation; TC, total cholesterol.

The next feature, Wavelet-HLL_firstorder_Range, represents the degree of intensity variation within the image. Higher range values typically indicate greater differences in pixel intensity, providing additional insights into the correlation with disease states. When the value of this feature increases, the SHAP value is positive, indicating that higher values of this feature contribute positively to the model’s prediction, potentially because greater variability suggests an unstable condition within the vascular structure.

Next, Wavelet-LLH_glszm_SizeZoneNonUniformity was found to be similar to the first feature, representing the nonuniformity of size zones. Higher values may indicate a more complex spatial arrangement of gray-level regions, which could be related to abnormal tissue characteristics. Lower negative SHAP values reinforce the notion that more uniform PVAT tissue features may protect against IS events.

The SHAP heatmap in Figure 4C further demonstrates the direction and magnitude of each feature’s influence across all cases within the model. The SHAP decision plot in Figure 4D depicts how each feature contributes to the final predicted probability for individual cases. Figure 5 presents two specific examples to better illustrate the interpretability of the model. These SHAP visualizations highlight the explainability of the model, emphasizing the importance of radiomics features in predicting symptomatic and asymptomatic carotid plaques while offering a transparent interpretation of the model’s decision-making process. These features collectively support the importance of texture and intensity variations in imaging as predictive markers for stroke risk, with their interactions providing valuable insights into underlying pathological mechanisms.

Figure 5 SHAP visualization of a single model. (A) Patient A is an example of a correctly predicted stroke case, and (B) patient B is an example of a correctly predicted non-stroke case. FAI, fat attenuation index; HU, Hounsfield units; PFD, perivascular fat density; SHAP, SHapley Additive exPlanation; TC, total cholesterol.

Follow-up and survival analysis

We further analyzed the patients’ follow-up data after discharge. During this period, the incidence of ischemic events in the non-stroke group was 10.8%, 13.2%, 11.5%, and 8.9% in the training, internal testing, external validation 1, and external validation 2 sets, respectively, while the incidence rates of ischemic events in the stroke group were 31.2%, 32.2%, 33.3%, and 28.8%, respectively. Event-free survival (EFS) differed significantly between the stroke and non-stroke groups in both the training and validation cohorts (Figure 6A-6D). The patients were divided into two groups based on the combined model: the predicted non-stroke group (XGBoost score <0.837) and the predicted stroke group (XGBoost score ≥0.837); this threshold was determined according to the maximization of the Youden index. Kaplan-Meier analysis indicated that the combined model effectively stratified patients into groups with significantly different follow-up event rates, showing close agreement with observed outcomes. The multivariate Cox regression analysis identified the independent predictors of follow-up ischemic events to be stroke status [hazard ratio (HR) =3.916; 95% confidence interval (CI): 1.792–6.558; P<0.001] and predicted stroke status (HR =1.352; 95% CI: 1.317–4.777; P=0.030) (Table S4).

Figure 6 The Kaplan-Meier survival curves based on patient grouping according to stroke occurrence and patient grouping predicted by the nomogram in the (A) training set, (B) internal validation set, (C) external validation set 1, and (D) external validation set 2. CI, confidence interval; HR, hazard ratio.

Discussion

This study developed and validated a model based on PVAT features derived from CTA images to predict acute stroke events in patients with T2DM. The integration of clinical and radiological features further enhanced the model. A comparative analysis was conducted on six commonly used ML algorithms, and their predictive performance was evaluated across multiple dimensions. Ultimately, XGBoost was selected as the optimal model, and a combined model incorporating XGBoost with independent risk factors was constructed. This model demonstrated excellent predictive performance, achieving an AUC of 0.847, 0.856, and 0.876 in the internal validation, external validation 1, and external validation 2 sets, respectively. The XGBoost-based combined model offered a robust and interpretable method for IS risk assessment and prognostic stratification in patients with T2DM, potentially aiding in personalized clinical decision-making and management.

This study integrated and extracted various types of radiomics features, and ultimately, seven key features were selected via LASSO regression. Among these, wavelet features were the most prominent, followed by the first-order features. This finding aligns with several studies that have identified the substantial prognostic value of wavelet features, emphasizing their critical role in constructing robust radiomics models (22,31). In addition, first-order features, which characterize the distribution of gray-level values in the region of interest, play a vital role in distinguishing and classifying lesions (32). These features are analogous to the mean voxel intensity of PVAT, reflecting changes in the water-lipid balance within PVAT (33). Therefore, these features can serve as effective imaging biomarkers for predicting symptomatic carotid plaques, highlighting the potential of radiomics in enhancing our understanding of the microstructural changes in PVAT and improving risk stratification in clinical practice.

Six ML models were developed based on the aforementioned radiomics features. Among them, the XGBoost classifier demonstrated outstanding performance and was selected as the best classifier. Its superiority may be related to its robust ability to handle nonlinear features and effectively deal with missing values. In addition, through the adoption of ensemble learning, it can construct multiple weak classifiers and combine them into a strong classifier, which significantly improves the model’s prediction accuracy and stability. Moreover, its built-in feature importance assessment function allows the model to excel in selecting the most relevant features, further enhancing predictive performance (34,35).

In addition, we established a PFD model, evaluated baseline clinical data, and identified hypertension and total cholesterol as independent predictors of IS in patients with T2DM. Previous research has established a strong association between hypertension and the occurrence of IS (36). Prolonged hypertension damages the vascular walls, accelerates the process of atherosclerosis, and makes blood vessels more vulnerable and prone to rupture, which can impair blood flow, potentially leading to IS (37,38). Similarly, studies such as those by Lewington et al. (39) have demonstrated that elevated total cholesterol levels increase the risk of stroke. Cholesterol deposits in the arterial walls form plaques, which can eventually obstruct blood flow and cause vascular blockages (40,41).

A clinical model was established based on these independent predictors. However, the PFD and clinical models performed suboptimally. In order to improve predictive accuracy, a combined model was developed that integrated the Radscore derived from the XGBoost classifier with independent clinical predictors and PFD. The XGBoost-based combined model demonstrated good predictive performance.

SHAP has become a widely used tool for visualizing each feature’s contribution to both overall and individual predictions (42). SHAP analysis in our study highlighted the importance and impact of each feature within the combined model. Case-specific analyses demonstrated the contribution of these features in certain scenarios. The final SHAP values can be used to calculate predictive probabilities, enabling more personalized and precise IS predictions.

In this study, a history of stroke was identified as an independent risk factor for EFS, which is consistent with previous research findings (43). Moreover, stroke status predicted by the XGBoost-based combined model was confirmed to be an independent risk factor for IS. The differences observed in the EFS curves stratified by XGBoost scores closely mirrored those based on the history of stroke, indicating that the XGBoost-based combined model can effectively guide the stratification of patients into different ischemic risk categories in T2DM. Therefore, this model could play a crucial role in assessing personalized recurrence risk and providing valuable insights for clinicians in developing appropriate and effective treatment strategies.

This study was the first to combine the radiological characteristics of PVAT with clinical risk factors to predict IS risk in patients with T2DM. Our proposed model, using PVAT features derived from CTA imaging, performs better than do previous ultrasound-based models (44) in identifying patients at risk of IS and can be used to support patient risk stratification based on predicted ischemic event likelihood. The XGBoost-based combined model, integrated with the SHAP method, can assist clinicians in evaluating the risk of individual IS symptoms. For patients with T2DM at risk of IS, this model offers guidance on whether more proactive preventive measures in patients at higher predicted risk are warranted. These patients may also require more frequent follow-up assessments. However, in real-world clinical settings, final treatment decisions must still be based on a comprehensive evaluation of each patient.

This study involved several limitations that should be acknowledged. First, given the retrospective design, it may be subject to selection bias. Moreover, only patients who could recover and be discharged were included, resulting in a relatively high proportion of negative cases, which may limit the clinical applicability of the developed model. Second, the data used were derived exclusively from Chinese patients, potentially restricting the generalizability of its findings to other geographic regions with different distributions of clinical characteristics. Third, our study focused on predicting the risk of initial IS in patients with T2Ds. The purpose of the follow-up event analysis was solely to evaluate the model’s ability to classify patients into different risk categories in a real-world timeline rather than to establish a prognostic or survival prediction model. Therefore, to assess the predictive ability of the model in terms of patient prognosis, future studies should be large-sample and prospective in design. Finally, the model incorporated the radiomics features of PVAT but not plaque-specific characteristics, which may reduce the model’s predictive performance.

Conclusions

This study developed and validated a novel predictive model based on PVAT radiomics features to predict acute stroke events in patients with T2DM. The purpose of this model is to identify symptomatic carotid plaques and stratify patients into distinct risk categories during follow-up. When applied in conjunction with SHAP, this model enhances personalized prediction by clarifying the contribution of individual features. Therefore, this model provides a foundation for multidimensional management and personalized risk assessment in patients with T2DM, potentially enabling the administration of preventive strategies at an earlier period among high-risk patients with T2DM.

Acknowledgments

None.

Footnote

Reporting Checklist: The authors have completed the TRIPOD+AI reporting checklist. Available at https://qims.amegroups.com/article/view/10.21037/qims-2025-1760/rc

Data Sharing Statement: Available at https://qims.amegroups.com/article/view/10.21037/qims-2025-1760/dss

Funding: This research was funded by the General Research Project of Zhejiang Provincial Education Department (No. Y202457302 to T.Z.) and the Medical Science and Technology Project of Zhejiang Province (No. 2025KY495 to W.H.).

Conflicts of Interest: All authors have completed the ICMJE uniform disclosure form (available at https://qims.amegroups.com/article/view/10.21037/qims-2025-1760/coif). T.Z. reports a funding from the General Research Project of Zhejiang Provincial Education Department (No. Y202457302). W.H. reports a funding from the Medical Science and Technology Project of Zhejiang Province (No. 2025KY495). Y.J. is an employee of Huiying Medical Technology Co., Ltd. The other authors have no conflicts of interest to declare.

Ethical Statement: The authors are accountable for all aspects of the work in ensuring that questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved. The study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. This multicenter retrospective study was approved by the Ethics Committee of The Fifth Affiliated Hospital of Wenzhou Medical University (No. 2025-136), and the requirement of informed consent from patients was waived due to its retrospective design.

Open Access Statement: This is an Open Access article distributed in accordance with the Creative Commons Attribution-NonCommercial-NoDerivs 4.0 International License (CC BY-NC-ND 4.0), which permits the non-commercial replication and distribution of the article with the strict proviso that no changes or edits are made and the original work is properly cited (including links to both the formal publication through the relevant DOI and the license). See: https://creativecommons.org/licenses/by-nc-nd/4.0/.

References

1. Global, regional, and national burden of stroke and its risk factors, 1990-2019: a systematic analysis for the Global Burden of Disease Study 2019. Lancet Neurol 2021;20:795-820. [Crossref] [PubMed]
2. Saini V, Guada L, Yavagal DR. Global Epidemiology of Stroke and Access to Acute Ischemic Stroke Interventions. Neurology 2021;97:S6-S16. [Crossref] [PubMed]
3. Zhang S, Yu X, Gu H, Kang B, Guo N, Wang X. Identification of high-risk carotid plaque by using carotid perivascular fat density on computed tomography angiography. Eur J Radiol 2022;150:110269. [Crossref] [PubMed]
4. Glovaci D, Fan W, Wong ND. Epidemiology of Diabetes Mellitus and Cardiovascular Disease. Curr Cardiol Rep 2019;21:21. [Crossref] [PubMed]
5. Li W, Chen D, Tao Y, Lu Z, Wang D. Association between triglyceride-glucose index and carotid atherosclerosis detected by ultrasonography. Cardiovasc Diabetol 2022;21:137. [Crossref] [PubMed]
6. Sarwar N, Gao P, Seshasai SR, Gobin R, Kaptoge S, Di Angelantonio E, Ingelsson E, Lawlor DA, Selvin E, Stampfer M, Stehouwer CD, Lewington S, Pennells L, Thompson A, Sattar N, White IR, Ray KK, Danesh J. Diabetes mellitus, fasting blood glucose concentration, and risk of vascular disease: a collaborative meta-analysis of 102 prospective studies. Lancet 2010;375:2215-22. [Crossref] [PubMed]
7. O'Donnell MJ, Chin SL, Rangarajan S, Xavier D, Liu L, Zhang H, et al. Global and regional effects of potentially modifiable risk factors associated with acute stroke in 32 countries (INTERSTROKE): a case-control study. Lancet 2016;388:761-75. [Crossref] [PubMed]
8. Bragg F, Holmes MV, Iona A, Guo Y, Du H, Chen Y, Bian Z, Yang L, Herrington W, Bennett D, Turnbull I, Liu Y, Feng S, Chen J, Clarke R, Collins R, Peto R, Li L, Chen ZChina Kadoorie Biobank Collaborative Group. Association Between Diabetes and Cause-Specific Mortality in Rural and Urban Areas of China. JAMA 2017;317:280-9. [Crossref] [PubMed]
9. Stitziel NO, Kanter JE, Bornfeldt KE. Emerging Targets for Cardiovascular Disease Prevention in Diabetes. Trends Mol Med 2020;26:744-57. [Crossref] [PubMed]
10. Pan L, Zou H, Meng X, Li D, Li W, Chen X, Yang Y, Yu X. Predictive values of metabolic score for insulin resistance on risk of major adverse cardiovascular events and comparison with other insulin resistance indices among Chinese with and without diabetes mellitus: Results from the 4C cohort study. J Diabetes Investig 2023;14:961-72. [Crossref] [PubMed]
11. Ménégaut L, Laubriet A, Crespy V, Leleu D, Pilot T, Van Dongen K, de Barros JP, Gautier T, Petit JM, Thomas C, Nguyen M, Steinmetz E, Masson D. Inflammation and oxidative stress markers in type 2 diabetes patients with Advanced Carotid atherosclerosis. Cardiovasc Diabetol 2023;22:248. [Crossref] [PubMed]
12. Gyldenkerne C, Mortensen MB, Kahlert J, Thrane PG, Warnakula Olesen KK, Sørensen HT, Thomsen RW, Maeng M. 10-Year Cardiovascular Risk in Patients With Newly Diagnosed Type 2 Diabetes Mellitus. J Am Coll Cardiol 2023;82:1583-94. [Crossref] [PubMed]
13. Ridker PM, Everett BM, Thuren T, MacFadyen JG, Chang WH, Ballantyne C, et al. Antiinflammatory Therapy with Canakinumab for Atherosclerotic Disease. N Engl J Med 2017;377:1119-31. [Crossref] [PubMed]
14. Ridker PM, Libby P, MacFadyen JG, Thuren T, Ballantyne C, Fonseca F, Koenig W, Shimokawa H, Everett BM, Glynn RJ. Modulation of the interleukin-6 signalling pathway and incidence rates of atherosclerotic events and all-cause mortality: analyses from the Canakinumab Anti-Inflammatory Thrombosis Outcomes Study (CANTOS). Eur Heart J 2018;39:3499-507. [Crossref] [PubMed]
15. Antonopoulos AS, Sanna F, Sabharwal N, Thomas S, Oikonomou EK, Herdman L, et al. Detecting human coronary inflammation by imaging perivascular fat. Sci Transl Med 2017;9:eaal2658. [Crossref] [PubMed]
16. Oikonomou EK, Marwan M, Desai MY, Mancio J, Alashi A, Hutt Centeno E, et al. Non-invasive detection of coronary inflammation using computed tomography and prediction of residual cardiovascular risk (the CRISP CT study): a post-hoc analysis of prospective outcome data. Lancet 2018;392:929-39. [Crossref] [PubMed]
17. Zhang S, Gu H, Yu X, Kang B, Yuan X, Wang X. Association Between Carotid Artery Perivascular Fat Density and Intraplaque Hemorrhage. Front Cardiovasc Med 2021;8:735794. [Crossref] [PubMed]
18. Mao L, Chen L, Qu M, He X. Pericarotid Adipose Tissue is Associated with Circulatory Markers of Inflammation and Carotid Atherosclerosis. Angiology 2025;76:857-63. [Crossref] [PubMed]
19. Dai A, Yao W, Lei J, Yan L, Dang L, Zhao H, Gu J, Li J, Nie Y, Zheng M, Wang D, Wang Q. Relationship Between Pericarotid Fat Density and Pathology-Based Carotid Plaque Risk Characteristics. J Craniofac Surg 2024; Epub ahead of print. [Crossref]
20. Tang C, Wang QH, Cai J, Wang YM, Chen ZP, Fan LZ, Zheng L, Qiao T. Correlation analysis of carotid adipose tissue density and plaque stability in patients with carotid artery stenosis. Zhonghua Yi Xue Za Zhi 2024;104:3123-9. [Crossref] [PubMed]
21. Marcelin G, Silveira ALM, Martins LB, Ferreira AV, Clément K. Deciphering the cellular interplays underlying obesity-induced adipose tissue fibrosis. J Clin Invest 2019;129:4032-40. [Crossref] [PubMed]
22. Chen C, Tang W, Chen Y, Xu W, Yu N, Liu C, Li Z, Tang Z, Zhang X. Computed tomography angiography-based radiomics model to identify high-risk carotid plaques. Quant Imaging Med Surg 2023;13:6089-104. [Crossref] [PubMed]
23. Gillies RJ, Kinahan PE, Hricak H. Radiomics: Images Are More than Pictures, They Are Data. Radiology 2016;278:563-77. [Crossref] [PubMed]
24. Kolossváry M, Karády J, Szilveszter B, Kitslaar P, Hoffmann U, Merkely B, Maurovich-Horvat P. Radiomic Features Are Superior to Conventional Quantitative Computed Tomographic Metrics to Identify Coronary Plaques With Napkin-Ring Sign. Circ Cardiovasc Imaging 2017;10:e006843. [Crossref] [PubMed]
25. Lin A, Kolossváry M, Yuvaraj J, Cadet S, McElhinney PA, Jiang C, Nerlekar N, Nicholls SJ, Slomka PJ, Maurovich-Horvat P, Wong DTL, Dey D. Myocardial Infarction Associates With a Distinct Pericoronary Adipose Tissue Radiomic Phenotype: A Prospective Case-Control Study. JACC Cardiovasc Imaging 2020;13:2371-83. [Crossref] [PubMed]
26. Dong X, Li N, Zhu C, Wang Y, Shi K, Pan H, Wang S, Shi Z, Geng Y, Wang W, Zhang T. Diagnosis of coronary artery disease in patients with type 2 diabetes mellitus based on computed tomography and pericoronary adipose tissue radiomics: a retrospective cross-sectional study. Cardiovasc Diabetol 2023;22:14. [Crossref] [PubMed]
27. Miao S, Yu F, Sheng R, Zhang X, Li Y, Qi Y, Lu S, Ji P, Fan J, Zhang X, Xu T, Wang Z, Liu Y, Yang G. Radiomics of pericoronary adipose tissue on computed tomography angiography predicts coronary heart disease in patients with type 2 diabetes mellitus. BMC Cardiovasc Disord 2024;24:300. [Crossref] [PubMed]
28. Yu S, Huo R, Qiao H, Ning Z, Xu H, Yang D, Shen R, Xu N, Han H, Chen S, Liu Y, Zhao X. Carotid artery perivascular adipose tissue on magnetic resonance imaging: a potential indicator for carotid vulnerable atherosclerotic plaque. Quant Imaging Med Surg 2023;13:7695-705. [Crossref] [PubMed]
29. Nie JY, Chen WX, Zhu Z, Zhang MY, Zheng YJ, Wu QD. Initial experience with radiomics of carotid perivascular adipose tissue in identifying symptomatic plaque. Front Neurol 2024;15:1340202. [Crossref] [PubMed]
30. Barnett HJM, Taylor DW, Haynes RB, Sackett DL, Peerless SJ, Ferguson GG, Fox AJ, Rankin RN, Hachinski VC, Wiebers DO, Eliasziw M. Beneficial effect of carotid endarterectomy in symptomatic patients with high-grade carotid stenosis. N Engl J Med 1991;325:445-53. [Crossref] [PubMed]
31. Oikonomou EK, Williams MC, Kotanidis CP, Desai MY, Marwan M, Antonopoulos AS, et al. A novel machine learning-derived radiotranscriptomic signature of perivascular fat improves cardiac risk prediction using coronary CT angiography. Eur Heart J 2019;40:3529-43. [Crossref] [PubMed]
32. Varriano G, Guerriero P, Santone A, Mercaldo F, Brunese L. Explainability of radiomics through formal methods. Comput Methods Programs Biomed 2022;220:106824. [Crossref] [PubMed]
33. Hu X, Chen J, Fu H, Chen Y, Fan D, Chen Y, Shen C. Association Between Carotid Artery Perivascular Fat Density and Embolic Stroke of Undetermined Source. Front Neurol 2021;12:765962. [Crossref] [PubMed]
34. Natekin A, Knoll A. Gradient boosting machines, a tutorial. Front Neurorobot 2013;7:21. [Crossref] [PubMed]
35. Borstelmann SMMachine Learning Principles for Radiology Investigators. Acad Radiol 2020;27:13-25. [Crossref] [PubMed]
36. Said MA, Verweij N, van der Harst P. Associations of Combined Genetic and Lifestyle Risks With Incident Cardiovascular Disease and Diabetes in the UK Biobank Study. JAMA Cardiol 2018;3:693-702. [Crossref] [PubMed]
37. Flint AC, Conell C, Ren X, Banki NM, Chan SL, Rao VA, Melles RB, Bhatt DL. Effect of Systolic and Diastolic Blood Pressure on Cardiovascular Outcomes. N Engl J Med 2019;381:243-51. [Crossref] [PubMed]
38. Liu FD, Shen XL, Zhao R, Tao XX, Wang S, Zhou JJ, Zheng B, Zhang QT, Yao Q, Zhao Y, Zhang X, Wang XM, Liu HQ, Shu L, Liu JR. Pulse pressure as an independent predictor of stroke: a systematic review and a meta-analysis. Clin Res Cardiol 2016;105:677-86. [Crossref] [PubMed]
39. Lewington S, Whitlock G, Clarke R, Sherliker P, Emberson J, Halsey J, Qizilbash N, Peto R, Collins R. Blood cholesterol and vascular mortality by age, sex, and blood pressure: a meta-analysis of individual data from 61 prospective studies with 55,000 vascular deaths. Lancet 2007;370:1829-39. [Crossref] [PubMed]
40. Liu Z, Zhang L, Wang L, Li K, Fan F, Jia J, Li J, Zhang Y. The predictive value of cumulative atherogenic index of plasma (AIP) for cardiovascular outcomes: a prospective community-based cohort study. Cardiovasc Diabetol 2024;23:264. [Crossref] [PubMed]
41. Kim MK, Han K, Kim HS, Park YM, Kwon HS, Yoon KH, Lee SH. Cholesterol variability and the risk of mortality, myocardial infarction, and stroke: a nationwide population-based study. Eur Heart J 2017;38:3560-6. [Crossref] [PubMed]
42. Ma J, Bo Z, Zhao Z, Yang J, Yang Y, Li H, Yang Y, Wang J, Su Q, Wang J, Chen K, Yu Z, Wang Y, Chen G. Machine Learning to Predict the Response to Lenvatinib Combined with Transarterial Chemoembolization for Unresectable Hepatocellular Carcinoma. Cancers (Basel) 2023;15:625. [Crossref] [PubMed]
43. Li J, Han M, Chen Y, Wu B, Wu Y, Jia W, Liu J, Luo H, Yu P, Tu J, Kuang J, Yi Y. Prediction of recurrent ischaemic stroke using radiomics data and machine learning methods in patients with acute ischaemic stroke: protocol for a multicentre, large sample, prospective observational cohort study in China. BMJ Open 2023;13:e076406. [Crossref] [PubMed]
44. Liu Y, Kong Y, Yan Y, Hui P. Explore the value of carotid ultrasound radiomics nomogram in predicting ischemic stroke risk in patients with type 2 diabetes mellitus. Front Endocrinol (Lausanne) 2024;15:1357580. [Crossref] [PubMed]