Development of an interpretable machine learning model to predict short-term bleeding risk in patients receiving dual
Haolong Han1,2, Jifan Zhang3, Xia Wang2
1Nanjing Drum Tower Hospital Clinical College of Nanjing University of Chinese Medicine, Nanjing, 210008, China.
Insights
This study developed a machine learning model to predict bleeding risk in cardiac surgery patients on dual antithrombotic therapy. The random forest model shows strong performance for individualized risk assessment.
Area of Science:
- Cardiology
- Medical Informatics
- Machine Learning
Background:
- Bleeding is a significant complication in cardiac surgery, particularly for patients on combined anticoagulant and antiplatelet therapy.
- Existing prediction models often exclude patients undergoing combined coronary artery bypass grafting (CABG) and valve surgery or fail to incorporate postoperative factors and dual antithrombotic therapy.
Purpose of the Study:
- To develop and validate machine learning (ML) models for predicting short-term bleeding events after discharge.
- To enable individualized risk assessment in patients receiving warfarin plus aspirin after combined CABG and valve surgery.
Main Methods:
- Retrospective analysis of 499 adult patients receiving dual antithrombotic therapy post-cardiac surgery.
- Selection of 11 key bleeding predictors using LASSO regression.
- Training and validation of seven ML algorithms, including random forest (RF), with external validation on 93 patients.
- Performance evaluation using AUC, accuracy, sensitivity, specificity, PPV, NPV, F1-score, and Brier score.
- Model interpretability assessed using Shapley Additive Explanations (SHAP).
Main Results:
- Eleven clinical variables were identified as significant bleeding predictors.
- The RF model achieved the highest performance in internal validation (AUC=0.85) and maintained strong performance in external validation (AUC=0.82).
- SHAP analysis provided transparent interpretation of global and individual bleeding risk contributions.
Conclusions:
- Machine learning models can effectively predict short-term bleeding risk in cardiac surgery patients on dual antithrombotic therapy.
- The developed RF model offers clinical utility for individualized risk assessment.
- SHAP interpretation enhances the clinical applicability of the prediction model.
Introduction:
Bleeding is a serious complication in cardiac surgery, especially among patients receiving combined anticoagulant and antiplatelet therapy. Current prediction models rarely include those undergoing both valve surgery and coronary artery bypass grafting (CABG) or account for postoperative factors and dual antithrombotic therapy, limiting their clinical utility.
Aim:
This study aimed to develop and validate machine learning models for the individualized prediction of bleeding events within three months after discharge in patients receiving warfarin plus aspirin following combined CABG and valve surgery.
Method:
Data from 499 adult patients who underwent cardiac surgery and received combined anticoagulant and antiplatelet therapy between June 2019 and December 2023 were retrospectively analyzed. Patients were randomly assigned to training (70%) and internal validation (30%) cohorts. Eleven key bleeding predictors were selected using the least absolute shrinkage and selection operator (LASSO) method. Seven machine learning algorithms, logistic regression, decision tree, random forest (RF), support vector machine, eXtreme Gradient Boosting (XGBoost), Light Gradient Boosting Machine (LightGBM) and k-nearest neighbors, were subsequently trained using these predictors. The model performance was evaluated based on the area under the receiver operating characteristic curve (AUC), accuracy, sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), F1 score, and Brier score. External validation was performed in a temporally independent cohort of 93 patients. Model interpretability was assessed through Shapley Additive Explanations (SHAP), which were used to visualize both global and individual risk contributions.
Results:
Eleven clinical variables, including anemia, diabetes, heart failure, atrial fibrillation, age, postoperative drainage, previous bleeding, stroke, body mass index, estimated glomerular filtration rates and intraoperative bleeding, were identified as predictors. Among the models, the RF algorithm demonstrated the best performance (internal validation: AUC = 0.85, accuracy = 0.80, sensitivity = 0.75, specificity = 0.82, PPV = 0.64, NPV = 0.89 and F1-score = 0.69). In the external validation, the RF model maintained strong performance (AUC = 0.82, accuracy = 0.73, sensitivity = 0.76, specificity = 0.72, PPV = 0.44, NPV = 0.92 and F1-score = 0.56). Decision curve analysis confirmed the clinical utility of the model, and SHAP visualizations provided a transparent interpretation of individualized bleeding risks.
Conclusion:
This study demonstrates the feasibility and clinical utility of ML-based prediction of short-term bleeding risk in cardiac surgery patients on dual antithrombotic therapy. The RF model enables individualized risk assessment, with SHAP interpretation supporting clinical applicability.
More Related Videos
Related Concept Videos
01:29Venous Thrombosis III: Interprofessional Care
01:30Anticoagulant Drugs: Low-Molecular-Weight Heparins
01:30Venous Thrombosis IV: Nursing Management
01:18Anticoagulant Drugs: Vitamin K Antagonists and Direct Oral Anticoagulants
Warfarin, a prominent vitamin K antagonist family member, exerts its effect by inhibiting the enzyme VKORC1 (vitamin K epoxide reductase complex 1). By hindering this enzyme, warfarin...