Multicenter development and validation of machine-learning risk models to predict procedural complete revascularization and in-hospital heart failure in STEMI patients treated with primary PCI

Figures

• 

  Figure 1. Flowchart showing the screening and selection of STEMI patients treated with PPCI from three participating centers, including application of the predefined inclusion and exclusion criteria and assignment to the training cohort (n = 734) and independent testing cohort (n = 352). Both endpoints—(in-hospital) HF prediction and procedural CR feasibility prediction—were modeled using the same training/testing framework.

• 

  Figure 2. Model performance comparison for (in-hospital) HF prediction in the training and testing cohorts. (A) Training-cohort performance metrics across machine-learning models, including accuracy, F1-score, NPV, PPV, sensitivity, specificity, and Youden index. (B) Testing-cohort performance metrics across models (the “validation set” in the figure corresponds to the independent testing cohort). (C) Training-cohort AUCs with 95% confidence intervals across models. (D) Testing-cohort AUCs with 95% confidence intervals across models. (E) Training-cohort ROC curves across models. (F) Testing-cohort ROC curves across models. The corresponding metric values are labeled in the plot for clarity. HF, heart failure; AUC, area under the receiver operating characteristic curve; ROC, receiver operating characteristic; PPV, positive predictive value; NPV, negative predictive value.

• 

  Figure 3. Clinical utility, classification performance, and calibration of the catBoost model for (in-hospital) HF prediction. (A) DCA of candidate models in the training cohort. (B) DCA of candidate models in the testing cohort. (C) Confusion matrix of the CatBoost model in the training cohort. (D) Confusion matrix of the CatBoost model in the testing cohort. (E) Clinical impact curve of the CatBoost model in the training cohort (predicted high-risk vs observed events across threshold probabilities). (F) Clinical impact curve of the CatBoost model in the testing cohort. (G) Calibration curve of the CatBoost model in the training cohort. (H) Calibration curve of the CatBoost model in the testing cohort. DCA, decision curve analysis; HF, heart failure.

• 

  Figure 4. SHAP-based interpretability of the catBoost model for (in-hospital) HF prediction. (A) Global feature importance ranked by mean absolute SHAP values. (B) SHAP beeswarm plot showing the distribution and directionality of feature effects across individuals (color indicates feature value). (C) SHAP dependence plots for selected key predictors illustrating potentially non-linear associations with the model output. (D) Prediction breakdown (waterfall plot) for one representative patient selected from the training cohort, showing how features cumulatively shift the prediction from the baseline (expected) value to the final output. (E) Feature contribution (force-like) plot for the same representative patient from the training cohort.

• 

  Figure 5. Model performance comparison for procedural CR prediction in the training and testing cohorts. (A) Training-cohort performance metrics across machine-learning models, including accuracy, F1-score, NPV, PPV, sensitivity, specificity, and Youden index. (B) Testing-cohort performance metrics across models (the “validation set” in the figure corresponds to the independent testing cohort). (C) Training-cohort AUCs with 95% confidence intervals across models. (D) Testing-cohort AUCs with 95% confidence intervals across models. (E) Training-cohort ROC curves across models. (F) Testing-cohort ROC curves across models. The corresponding metric values are labeled in the plot for clarity. CR, complete revascularization; AUC, area under the receiver operating characteristic curve; ROC, receiver operating characteristic.

• 

  Figure 6. Clinical utility, classification performance, and calibration of the catBoost model for procedural CR prediction. (A) DCA of candidate models in the training cohort. (B) DCA of candidate models in the testing cohort. (C) Confusion matrix of the CatBoost model in the training cohort. (D) Confusion matrix of the CatBoost model in the testing cohort. (E) Clinical impact curve of the CatBoost model in the training cohort. (F) Clinical impact curve of the CatBoost model in the testing cohort. (G) Calibration curve of the CatBoost model in the training cohort. (H) Calibration curve of the CatBoost model in the testing cohort. CR, complete revascularization; DCA, decision curve analysis.

• 

  Figure 7. SHAP-based interpretability of the catBoost model for procedural CR prediction. (A) Global feature importance ranked by mean absolute SHAP values. (B) SHAP beeswarm plot showing the distribution and directionality of feature effects across individuals (color indicates feature value). (C) SHAP dependence plots for selected key predictors illustrating potentially non-linear associations with the likelihood of achieving procedural CR. (D) Prediction breakdown (waterfall plot) for one representative patient selected from the training cohort. (E) Feature contribution (force-like) plot for the same representative patient from the training cohort. SHAP, Shapley Additive Explanations; CR, complete revascularization; Time-to-Guidewire, symptom-to-guidewire crossing time.