Development and validation of a machine learning model for post-PCI exercise intolerance in patients with coronary artery disease via electronic medical records

Figures

• 

  Figure 1. The graphical abstract for the study.

• 

  Figure 2. Flowchart of the study design.

• 

  Figure 3. Feature variable selection via LASSO regression and the Boruta algorithm. (A) LASSO coefficient trajectories. Each curve represents a variable’s coefficient path along the sequence of log(λ). (B) Cross-validation for optimal λ selection. The left dashed line indicates λ corresponding to the minimum binomial deviance (λmin), and the right dashed line represents the one-standard-error criterion (λ1SE). (C) Feature importance according to the Boruta algorithm. The variables are categorized as “Confirmed” (teal), “Tentative” (yellow), or “Rejected” (red) according to their Z scores of importance.

• 

  Figure 4. Performance comparison of seven machine learning models in the testing dataset: (A) ROC curves; (B) calibration curves.

• 

  Figure 5. Decision curve analysis of the MLP model. The blue curve shows the net clinical benefit of the MLP model across threshold probabilities. The dashed black and gray lines represent the “treat all” and “treat none” strategies, respectively. The red dotted line indicates the F1-optimal threshold (0.30), where the model achieves the best balance between precision and recall.

• 

  Figure 6. SHAP-based interpretation of the MLP model. (A) Feature importance ranking derived from the mean absolute SHAP values, reflecting each variable’s average contribution to the model output. (B) SHAP beeswarm visualization illustrating how feature magnitude and direction (color-coded) influence the model’s predicted probability for the positive outcome.

• 

  Figure 7. Operation page of the web risk calculator. The calculator allows users to input categorical and numerical features derived from EMRs to estimate the probability of exercise intolerance. The risk output is visualized through a color-coded risk meter and SHAP-based feature explanations, where red bars indicate positive contributions to risk. In this example, the predicted probability of exercise intolerance was 69%, with age, hemoglobin, and smoking being the strongest risk contributors.