Machine learning prediction of physician-assigned treatment categories from pre-procedural electronic medical record data in cardiac patients: A multiclass classification study

Figures

• 

  Fig 1. Schematic representation of the proposed methodology. The workflow starts with data extraction from the BioArc Registry, followed by preprocessing of semi-structured records and temporal feature engineering. Class imbalance is addressed within training folds using SMOTE. Feature selection, hyperparameter tuning, and model evaluation are performed within a nested stratified cross-validation framework. Finally, model behavior is examined using SHAP analysis and complementary partial dependence plots.

• 

  Fig 2. Summarized clustered structure of the BioArc Registry [16], showing main sections with feature counts.

• 

  Fig 3. Different levels of the nested fields for symptoms, dates, severity, and notes depicting the detailed anatomy of the Signs and Symptoms part (code 7190) from the BioArc Registry.

• 

  Fig 4. Model performance as a function of the number of features retained during recursive feature elimination. The line plot shows the mean inner cross-validated macro F1-score across candidate retained-feature counts. Error bars or shaded intervals represent variability across inner folds. The selected feature count of 18 retained predictors is indicated on the plot.

• 

  Fig 5. Partial dependence plots derived from the trained XGBoost model. These plots summarize the marginal association between selected predictors and model-predicted probabilities. The displayed trends reflect model behavior within this dataset and should not be interpreted causally.

• 

  Fig 6. SHAP summary bar plots for the final XGBoost model. Separate class-specific SHAP feature importance charts are shown for medical therapy, PCI, and surgery. The x-axis shows the mean absolute SHAP value, indicating the overall magnitude of each feature’s contribution to the model output for that class.

• 

  Fig 7. One-vs-rest ROC curves for the final XGBoost workflow based on outer-fold predictions. Receiver operating characteristic curves are shown for medical therapy, PCI, and surgery using aggregated predictions from the outer folds of nested cross-validation.

• 

  Fig 8. One-vs-rest precision–recall curves for the final XGBoost workflow based on outer-fold predictions. Precision–recall curves are shown for medical therapy, PCI, and surgery using aggregated predictions from the outer folds of nested cross-validation.

• 

  Fig 9. One-vs-rest calibration curves for the final XGBoost workflow based on outer-fold predictions. Calibration curves are shown for medical therapy, PCI, and surgery using aggregated predictions from the outer folds of nested cross-validation.

• 

  Fig 10. Aggregated confusion matrix for the final XGBoost workflow across outer folds. Misclassifications are interpreted in relation to the clinical proximity of the treatment categories. The dominant confusion pattern shows residual confusion between the medical therapy and PCI classes.