Machine learning prediction of physician-assigned treatment categories from pre-procedural electronic medical record
1Department of Engineering, Ilam University, Ilam, Iran.
Insights
Machine learning models can predict physician treatment choices for coronary artery disease (CAD) using electronic health records. The model accurately predicts medical therapy, percutaneous coronary intervention (PCI), and coronary artery bypass grafting (CABG) assignments.
Area of Science:
- Cardiology
- Medical Informatics
- Machine Learning
Background:
- Cardiovascular diseases (CVDs) are a major global health burden.
- Treatment decisions for coronary artery disease (CAD) involving medical therapy, percutaneous coronary intervention (PCI), and coronary artery bypass grafting (CABG) are complex.
- Predicting these physician-assigned treatment categories is crucial for understanding clinical practice patterns.
Purpose of the Study:
- To develop and evaluate a machine learning model to predict physician-assigned treatment categories for CAD patients.
- To identify key pre-procedural factors influencing treatment decisions.
- To assess the model's performance using a retrospective multicenter registry.
Main Methods:
- A retrospective analysis of 2,682 patient records from the BioArc Clinical Registry.
- Development of a multiclass machine learning model (XGBoost) using pre-decision demographic, clinical, and risk-factor variables.
- Implementation of data preprocessing techniques including handling missing data, encoding, SMOTE, recursive feature elimination (RFE), and nested cross-validation.
Main Results:
- The XGBoost model, utilizing 18 predictors identified by RFE, achieved an overall accuracy of 82.3% and a macro F1-score of 0.775.
- One-vs-rest AUC values were 0.78 for medical therapy, 0.75 for PCI, and 0.80 for surgery.
- SHAP and partial dependence plots highlighted age, systolic blood pressure, symptom-related features, and opium consumption as influential predictors.
Conclusions:
- Pre-procedural electronic medical record (EMR) data contain significant information predictive of observed treatment assignment patterns in CAD.
- The developed model predicts physician treatment choices but does not identify the clinically optimal treatment or guarantee improved outcomes.
- Further external validation, calibration, and prospective outcome-based evaluation are necessary before clinical implementation.
Abstract:
Cardiovascular diseases (CVDs) remain a leading source of morbidity, mortality, and healthcare burden worldwide. In patients with coronary artery disease (CAD), choosing among medical therapy, percutaneous coronary intervention (PCI), and coronary artery bypass grafting (CABG) is clinically complex and depends on multiple pre-procedural factors. In this retrospective multicenter registry study, we developed a multiclass machine learning model to predict physician-assigned treatment categories rather than clinically optimal treatment. The analytic cohort included 2,682 unique patient records from the BioArc Clinical Registry, with class imbalance across medical therapy (n = 1,634), PCI (n = 817), and surgery (n = 231). Semi-structured JSON records were parsed to extract pre-decision demographic, clinical, symptom-related, and risk-factor variables. Missing data handling, encoding, SMOTE, recursive feature elimination (RFE), and hyperparameter tuning were performed within a nested stratified cross-validation workflow to reduce dependence on a single data split and to limit information leakage. The final XGBoost workflow retained 18 predictors after RFE. Across the outer folds of nested cross-validation, the model achieved an overall accuracy of 82.3% ± 1.5% (95% CI: 80.8%-83.8%) and a macro F1-score of 0.775 ± 0.018 (95% CI: 0.757-0.793), with one-vs-rest AUC values of 0.78 ± 0.02, 0.75 ± 0.02, and 0.80 ± 0.03 for medical therapy, PCI, and surgery, respectively. Class-specific performance and confusion patterns were summarized from outer-fold predictions. Model interpretation using SHAP and partial dependence plots identified age, systolic blood pressure, symptom-related features, and opium consumption as influential predictors in this dataset. These findings indicate that pre-procedural EMR data contain signal related to observed treatment-assignment patterns; however, the model does not identify the optimal treatment or demonstrate improved clinical outcomes. External validation, external calibration, and prospective outcome-based evaluation are required before clinical deployment.
