Machine learning to predict major bleeding during anticoagulation for venous thromboembolism: possibilities and...

Area of Science:

• Cardiology
• Medical Informatics
• Data Science

Background:

• Predicting major bleeding (MB) in venous thromboembolism (VTE) patients is crucial for safe anticoagulation.
• Traditional risk scores have limitations in accurately identifying high-risk individuals.
• Machine learning (ML) offers a promising avenue for developing more precise predictive tools.

Purpose of the Study:

• To develop and validate ML algorithms for predicting MB risk in VTE patients during the initial three months of anticoagulation.
• To compare the performance of ML models against established risk scores like RIETE and VTE-BLEED.

Main Methods:

• Utilized data from the Registro Informatizado de Enfermedad TromboEmbólica (RIETE) database, including 55 baseline variables.
• Trained and validated ML algorithms, with XGBoost identified as the best-performing method.
• Conducted prospective validation using new RIETE data and external validation with the COMMAND-VTE database.

Main Results:

• The XGBoost model demonstrated superior performance (F1 score: 15.4%) compared to RIETE (8.6%) and VTE-BLEED (6.4%) in the prospective validation cohort.
• In external validation, XGBoost's F1 score (5.2%) was lower than the RIETE score (17.3%), indicating variable performance across different datasets.
• The ML model showed high specificity (93%) but limited sensitivity (33.2%) in prospective validation.

Conclusions:

• ML algorithms, particularly XGBoost, show promise in predicting MB in VTE patients, outperforming traditional scores in certain validation settings.
• The performance variability across prospective and external validation cohorts suggests challenges in generalizability.
• Further research and model refinement are necessary to enhance the clinical utility of ML-based bleeding risk prediction in VTE management.