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

Background: With the increasing use of machine learning (ML)-based risk prediction models for venous thromboembolism (VTE) in patients, the quality and applicability of these models in practice and future research remain unknown. The prediction mechanism of ML and the number of selected factors have been research hotspots in VTE prediction. Objective: This study aimed to systematically review the literature on the predictive value of ML for VTE. Methods: PubMed, Web of Science, MEDLINE, Embase, CINAHL, and Cochrane Library databases were searched for studies published up to March 26, 2025. Studies that developed and validated an ML model for VTE prediction in the patient population and were published in English were eligible, and studies with duplicate data were excluded. The Prediction Model Risk of Bias Assessment Tool was used to assess the risk of bias in the included studies. Meta-analyses were performed to evaluate the C-index, sensitivity, and specificity. Results: A total of 27 studies with 596,092 patients reported the assessment value of ML models for predicting VTE. The risk of bias assessment yielded 18 (67%) studies with a high risk of bias, 8 (30%) with an unclear risk of bias, and 1 (4%) with a low risk of bias. The pooled sensitivity and specificity were 0.79 (95% CI 0.78-0.80) and 0.82 (95% CI 0.81-0.82), respectively. The positive likelihood ratio was 5.02 (95% CI 3.81-6.60), the negative likelihood ratio was 0.27 (95% CI 0.22-0.33), and the diagnostic odds ratio was 20.14 (95% CI 13.69-29.63; P<.001). A random-effects model was leveraged for meta-analysis of the C-index, which was 0.84 (95% CI 0.80-0.88). The most significant predictors for VTE were age, D-dimer level, and VTE history. Conclusions: ML has been shown to effectively predict VTE in patients. However, a high risk of bias was identified in most of the included studies (18/27, 67%), primarily due to shortcomings in handling missing data and reporting the study design. Consequently, future research must prioritize external validation and address methodological rigor to facilitate the translation of these models into routine clinical practice.