Machine learning models outperformed traditional risk scores in predicting bleeding (AUC 0.880 vs 0.730 for HAS-BLED) and thromboembolic risk (AUC 0.792 vs 0.628 for CHA₂DS₂-VASc).
Case-Control (n=1,472)
No
Do machine learning models improve the prediction of bleeding and thromboembolic risks compared to traditional risk scores in patients on oral anticoagulants?
Machine learning models incorporating diverse clinical indicators and drug metabolism features significantly outperformed traditional risk scores in predicting bleeding and thromboembolic risks in patients on oral anticoagulants.
Balancing thromboembolic prevention against bleeding risk remains a key challenge during oral anticoagulant (OAC) therapy. CHA₂DS₂-VASc cannot predict residual thromboembolic risk, and HAS-BLED has insufficient predictive ability for direct oral anticoagulants (DOACs).While machine learning offers a transformative paradigm for risk assessment, its clinical application is often hindered by three critical challenges: data imbalance caused by low incidence of embolism and hemorrhage, the omission of drug metabolism-related features, and limited generalizability across diverse DOAC regimens. We retrospectively collected clinical data from patients receiving OAC in the First Affiliated Hospital of Soochow University from 2018 to 2024. To address data imbalance, we recruited patients in case-control manner, then implemented a non-boundary oversampling strategy. To investigate more valuable predictors, we incorporated drug metabolism-related features as potential predictors to develop robust models. Shapley Additive exPlanations (SHAP) analysis supports the global and local interpretation for prediction, validating the contribution of predictors and enhancing the credibility of models in clinic. 281 patients with bleeding events, 213 patients with thromboembolic events, and 978 as negatvie control were recruited. The overall dataset for bleeding risk prediction included 1,259 patients (positive-to-negative ratio ≈ 1:3.48), and that for thromboembolism risk prediction included 1,191 patients (positive-to-negative ratio ≈ 1:4.59). The Light Gradient Boosting Machine (LGBM) achieved an AUC of 0.880 for predicting bleeding risk, outperformed the HAS-BLED score (AUC = 0.730). The Logistic Regression (LR) for predicting thromboembolic risk achieved an AUC of 0.792, outperformed the CHA₂DS₂-VASc score (AUC = 0.628). Decision curve analysis further suggested that these models provided meaningful clinical net benefit within reasonable threshold ranges. SHAP identified pulmonary artery pressure (PAP), left atrial volume index (LAVI), platelet count, and left atrial appendage (LAA) volume as key predictors of thromboembolic risk, while estimated glomerular filtration rate (eGFR), body mass index (BMI), and direct bilirubin were key predictors of bleeding risk, consistent with clinical expectations. By integrating diverse clinical indicators, prioritizing collection of positive events, and applying a new data augmentation, we identified several novel predictors, including PAP, LAVI, platelet count, LAA volume, eGFR, and BMI. Compared with traditional risk scores, the new models suggested superior predictive performance, providing robust evidence-based support for personalized clinical decision-making.
Leng et al. (Tue,) conducted a case-control in Patients on oral anticoagulation therapy (n=1,472). Machine learning risk prediction models vs. Traditional risk scores (HAS-BLED and CHA₂DS₂-VASc) was evaluated on Bleeding and thromboembolic risk prediction (Area Under the Curve). Machine learning models outperformed traditional risk scores in predicting bleeding (AUC 0.880 vs 0.730 for HAS-BLED) and thromboembolic risk (AUC 0.792 vs 0.628 for CHA₂DS₂-VASc).
Synapse has enriched 5 closely related papers on similar clinical questions. Consider them for comparative context: