Construction of diagnostic and prognostic models for premature coronary artery disease based on multiple machine algorithms

Figures

• 

  Figure 1. Model development and validation flowchart. PCAD, premature coronary artery disease; PIV, pan-immune inflammatory value; TyG, triglyceride-glucose index; WBC, white blood cell count; PSM, propensity score matching; GBM, gradient boosting machine; XGboost, extreme gradient boosting; SVM, support vector machine; RF, random forest; RSF, random survival forest; ROC, receiver operating characteristic.

• 

  Figure 2. Multiple machine learning results. (A) Gradient Boosting Machine (GBM): The top five feature variables based on relative importance ranking are CCI, WBC, Gender, PIV, and TyG. (B) Lasso regression: at the optimal lambda value (lambda = −7.48), 32 variables were retained, including TyG, PIV, and WBC. (C) Random forest: based on Gini importance evaluation, the five most important features are CCI, WBC, PIV, TyG, and PLT. (D) Support vector machines (SVM): The model achieved the lowest Root Mean Square Error (RMSE) of 0.3027, indicating optimal performance, with the top five contributing variables being CCI, WBC, RBC, PIV, and TyG. (E) Extreme Gradient boosting (XGBoost): Based on the feature importance scores, the variables are ranked as follows: PIV (0.80), WBC (0.18), and TyG (0.02).

• 

  Figure 3. Decision tree. The accuracy of the model is 0.88 and the confidence level is 0.86. WBC, white blood cells; TyG, triglyceride-glucose index; PIV, pan-immunoinflammatory value.

• 

  Figure 4. An ensemble model for the prognosis of PCAD. The StepCox[forward]-RSF ensemble model exhibited the highest mean C-index (0.99) across both training and validation cohorts. RSF, random survival forest.

• 

  Figure 5. Random survival forest. (A) Show that the error rate of the model stabilizes when the number of classification trees exceeds 200. (B) RSF: Based on Gini importance evaluation, the five most important features are LDL, TyG, PIV, CRI1, and CRI2. RSF, random survival forest; LDL, low-density lipoprotein cholesterol; TyG, triglyceride-glucose index; PIV, pan-immunoinflammatory value; CRI1and CRI2, castelli risk index 1 and 2.

• 

  Figure 6. ROC curve. (A) The optimal threshold for the PIV index predicted by the PCAD prognostic model was determined to be 229.5, at which the specificity and sensitivity of the PCAD prognostic model were 65.9% and 68.9%. (B) The optimal critical value of the TyG index predicts that the PCAD prognostic model is 4.9, at which the specificity and sensitivity of the PCAD prognostic model are 92.1% and 52.2%. TyG, triglyceride-glucose index; PIV, pan-immunoinflammatory value.

• 

  Figure 7. Kaplan–Meier survival analysis. (A) The follow-up of patients with different PIV index groups for 36 months. (B) The follow-up of patients with different TyG index group for 36 months. (C) Men and women were followed up for 36 months in different PIV + TyG groups. TyG, triglyceride-glucose index; PIV, pan-immunoinflammatory value.