Machine Learning Models for Predicting Bleeding Risk in Anticoagulated Patients with Atrial Fibrillation and Venous Thromboembolism: A Comparative Evidence Synthesis

TL;DR

Machine learning models predict bleeding risk better than traditional methods in patients with heart rhythm issues and blood clots, but more research is needed to confirm their usefulness in real-world settings.

Contribution

This study compares machine learning models with conventional clinical scores for predicting bleeding risk in anticoagulated patients with atrial fibrillation and venous thromboembolism.

Findings

Machine learning models like random forest and XGBoost outperformed traditional tools in predicting bleeding events.

Deep learning ensembles achieved the highest area under the curve (AUC) values in both atrial fibrillation and venous thromboembolism populations.

The improvement in accuracy over traditional scores was modest, with AUC differences ranging from 0.05 to 0.15.

Abstract

Background: Accurate prediction of bleeding events in patients receiving oral anticoagulants remains a key challenge in the management of atrial fibrillation (AF) and venous thromboembolism (VTE). Machine learning (ML) algorithms have emerged as powerful tools that capture complex, nonlinear interactions among risk factors, potentially offering superior accuracy. Objectives: To synthesize evidence comparing ML-based bleeding risk models with conventional clinical scores in anticoagulated AF and VTE populations. Methods: We conducted a systematic review with narrative synthesis of studies published between 2015 and 2025 applying ML algorithms to predict bleeding events in anticoagulated AF or VTE patients. Results: Thirteen studies were identified (seven AF and six VTE), including 464,523 participants in total. ML algorithms such as random forest (RF), extreme gradient boosting (XGBoost), and neural networks consistently outperformed traditional tools. In AF, AUCs ranged from 0.64 to 0.76 compared to 0.52–0.61 for HAS-BLED. In VTE, ML models achieved 0.59–0.91 versus 0.61–0.65 for RIETE or VTE-BLEED. Deep learning ensembles reached the highest AUCs (>0.8). Conclusions: ML-based bleeding risk models demonstrated statistically superior discrimination compared to established scores in both AF and VTE contexts, but effect sizes were modest (ΔAUC 0.05–0.15) and clinical utility remains uncertain. Broader validation, calibration assessment, and demonstration of impact on clinical outcomes are necessary before routine adoption.

Figures