Machine Learning in the Prediction of Venous Thromboembolism: Systematic Review and Meta-Analysis

Area of Science:

• Medical Informatics
• Biostatistics
• Clinical Prediction Models

Background:

• Machine learning (ML) models are increasingly used for venous thromboembolism (VTE) risk prediction.
• The quality and real-world applicability of these ML models require thorough investigation.
• Key research areas include ML prediction mechanisms and the identification of critical predictive factors for VTE.

Purpose of the Study:

• To systematically review and evaluate the predictive performance of ML models for VTE.
• To assess the overall diagnostic accuracy and identify key predictors in ML-based VTE risk models.

Main Methods:

• A comprehensive literature search was conducted across major databases up to March 26, 2025.
• Studies developing and validating ML models for VTE prediction in patient populations were included.
• Meta-analyses were performed to determine pooled sensitivity, specificity, and C-index, with risk of bias assessed using a standardized tool.

Main Results:

• Twenty-seven studies involving 596,092 patients were analyzed.
• Pooled sensitivity was 0.79 and pooled specificity was 0.82, with a C-index of 0.84.
• High risk of bias was noted in 67% of studies; age, D-dimer, and VTE history were significant predictors.

Conclusions:

• ML models demonstrate significant potential for predicting VTE in patients.
• A high prevalence of bias in existing studies, particularly in data handling and design reporting, limits current applicability.
• Future research should focus on external validation and enhanced methodological rigor to enable clinical translation.