Development of an explainable machine learning model for 3-year cardiovascular risk prediction in new-onset type 2 diabetes using the TyG index and ultrasound features

Figures

• 

  Fig. 1. Flow chart for patient screening. T2D, type 2 diabetes; CVD, cardiovascular disease

• 

  Fig. 2. Performance of the models in predicting CVD incidence. (A) AUCs of different models with different numbers of features. (B) The AUC, sensitivity, specificity, and F1 score of the LightGBM model with various numbers of features. ROC, receiver operating characteristic curve; AUC, area under the ROC curve; RF, random forest; XGBoost, extreme gradient boosting; LightGBM, light gradient boosting machine; CatBoost, categorical boosting; CVD, cardiovascular disease

• 

  Fig. 3. Global interpretability assessment through SHAP values. (A) Feature importance ranking via the SHAP bar plot. (B) Feature importance ranking via a SHAP dot plot. The horizontal axis represents the SHAP value magnitude, which quantifies each feature’s directional impact on CVD predictions. The features are ranked vertically by importance. (C) SHAP dependence plot. Dependence plots visualize the influence of individual features on CVD predictions. Threshold-exceeding SHAP values (> 0) indicate increased likelihood of CVD in the predictive algorithm. TyG, triglyceride–glucose; IMT, intima–media thickness; CVD, cardiovascular disease

• 

  Fig. 4. Performance comparison between the LightGBM model and the FRS for predicting 3-year cardiovascular risk in patients with new-onset type 2 diabetes. Receiver operating characteristic curves and classification metrics for the training cohort (A) and validation cohort (B). (C) Decision curve analysis comparing the net clinical benefit of the LightGBM model and FRS across a range of threshold probabilities. (D) Calibration plot of the LightGBM model in the validation cohort, demonstrating good agreement between the predicted and observed CVD risk probabilities (Hosmer–Lemeshow test: P = 0.959; mean absolute error = 0.02), supporting model reliability. FRS, Framingham Risk Score

• 

  Fig. 5. Explainable model interpretation and cardiovascular risk stratification. (A–C) SHAP force plots for three representative patients illustrating individualized prediction outputs. The blue and red bars indicate the direction and magnitude of each feature’s contribution to the model’s prediction. Patients A and B represent low- and medium-risk cases, respectively, with negative SHAP values driving the prediction below the baseline. Patient C shows high predicted risk, driven primarily by elevated carotid plaque burden, plaque thickness, IMT, and the TyG index. (D) Kaplan–Meier survival curves for risk stratification. Distinct difference in 3-year cardiovascular event–free survival was observed among the three groups (P < 0.0001), demonstrating the model’s clinical utility for risk stratification. SHAP, SHapley Additive ExPlanations; TyG, triglyceride–glucose; IMT, intima–media thickness; CVD, cardiovascular disease

• 

  Fig. 6. A screenshot of the web application. Input features, including age, hypertension status, IMT, TyG index, carotid plaque burden, and plaque thickness, are included. The model estimates a 50.2% probability of cardiovascular disease, categorizing the individual as high risk. A SHAP force plot provides model interpretability, visually illustrating each feature’s contribution to the prediction. A clinical recommendation is automatically generated, encouraging physician consultation and medical intervention. IMT, intima–media thickness; TyG, triglyceride–glucose; SHAP, SHapley Additive exPlanations