Scholarlink

<p>Recurrent venous thromboembolism is a major cause of morbidity and mortality, and current risk scores provide limited predictive accuracy. Machine learning (ML) models may improve risk stratification by capturing complex, nonlinear associations, but their comparative performance remains uncertain. We systematically searched PubMed, Embase, Web of Science, Scopus, and EBSCO to July 31, 2025, for studies developing or validating ML models for Recurrent VTE prediction. Eligible studies reported discrimination, calibration, or classification metrics. Risk of bias was assessed with PROBAST + AI, and evidence certainty with GRADE. Pooled estimates of discrimination were calculated using bivariate random-effects meta-analysis, with subgroup and meta-regression analyses to explore heterogeneity. Eight studies (<i>n</i> = 235–21,227) were included. ML models achieved strong pooled discrimination (AUC 0.89, 95% CI 0.86–0.91), sensitivity 0.80 (95% CI 0.56–0.93), and specificity 0.83 (95% CI 0.66–0.92), with a diagnostic odds ratio of 20 (95% CI 7–52). Heterogeneity was high (I² &gt; 79%) but mostly explained by model type and sample size. Neural networks outperformed other ML approaches (AUC &gt; 0.90). No significant publication bias was detected. Risk of bias was low in model development but high during evaluation due to reliance on internal validation. Certainty of evidence was graded as moderate. ML models demonstrate superior discrimination compared with traditional Recurrent VTE risk scores and have potential to inform individualized anticoagulation decisions. However, limited external validation, reliance on specialized predictors, and substantial heterogeneity restrict immediate clinical application. Large, prospective, multicenter studies using routinely available data are needed before widespread implementation.</p> Graphical abstract <p>Recurrent venous thromboembolism (RVTE) remains a major clinical challenge, and existing risk scores offer limited predictive accuracy. We systematically reviewed and meta-analyzed eight studies (n = 235–21,227) evaluating machine learning (ML) models for RVTE prediction. ML approaches demonstrated strong pooled performance (AUC 0.89, 95% CI 0.86–0.91; sensitivity 0.80; specificity 0.83; diagnostic odds ratio 20), with neural networks consistently outperforming other algorithms. High heterogeneity (I² &gt; 79%) was explained by model type and sample size. Risk of bias was low in development but high in evaluation due to limited external validation. Overall certainty of evidence was moderate. </p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Accuracy of machine learning models in predicting recurrent venous thromboembolism: a systematic review and meta-analysis

  • 期刊论文

  • 发表日期：

    2026年2月20日

查看全文
  • Pooya Eini,
  • Peyman Eini,
  • Homa Serpoush,
  • Mohammad Rezayee,
  • Jason Tremblay

摘要

Recurrent venous thromboembolism is a major cause of morbidity and mortality, and current risk scores provide limited predictive accuracy. Machine learning (ML) models may improve risk stratification by capturing complex, nonlinear associations, but their comparative performance remains uncertain. We systematically searched PubMed, Embase, Web of Science, Scopus, and EBSCO to July 31, 2025, for studies developing or validating ML models for Recurrent VTE prediction. Eligible studies reported discrimination, calibration, or classification metrics. Risk of bias was assessed with PROBAST + AI, and evidence certainty with GRADE. Pooled estimates of discrimination were calculated using bivariate random-effects meta-analysis, with subgroup and meta-regression analyses to explore heterogeneity. Eight studies (n = 235–21,227) were included. ML models achieved strong pooled discrimination (AUC 0.89, 95% CI 0.86–0.91), sensitivity 0.80 (95% CI 0.56–0.93), and specificity 0.83 (95% CI 0.66–0.92), with a diagnostic odds ratio of 20 (95% CI 7–52). Heterogeneity was high (I² > 79%) but mostly explained by model type and sample size. Neural networks outperformed other ML approaches (AUC > 0.90). No significant publication bias was detected. Risk of bias was low in model development but high during evaluation due to reliance on internal validation. Certainty of evidence was graded as moderate. ML models demonstrate superior discrimination compared with traditional Recurrent VTE risk scores and have potential to inform individualized anticoagulation decisions. However, limited external validation, reliance on specialized predictors, and substantial heterogeneity restrict immediate clinical application. Large, prospective, multicenter studies using routinely available data are needed before widespread implementation.

Graphical abstract

Recurrent venous thromboembolism (RVTE) remains a major clinical challenge, and existing risk scores offer limited predictive accuracy. We systematically reviewed and meta-analyzed eight studies (n = 235–21,227) evaluating machine learning (ML) models for RVTE prediction. ML approaches demonstrated strong pooled performance (AUC 0.89, 95% CI 0.86–0.91; sensitivity 0.80; specificity 0.83; diagnostic odds ratio 20), with neural networks consistently outperforming other algorithms. High heterogeneity (I² > 79%) was explained by model type and sample size. Risk of bias was low in development but high in evaluation due to limited external validation. Overall certainty of evidence was moderate.