Advancing Mortality Prediction in Pulmonary Embolism Using Machine Learning Algorithms—Systematic Review and Meta-Analysis
Pooya Eini, Peyman Eini, Homa Serpoush, Mohammad Rezayee, Jason Tremblay
https://doi.org/10.1002/pul2.70166
Abstract
This systematic review and meta-analysis evaluated the performance of machine learning (ML) models in predicting mortality among pulmonary embolism (PE) patients, synthesizing data from 17 studies encompassing 844,071 cases. Logistic Regression was the most commonly used algorithm, followed by advanced models like Random Forests, Support Vector Machines, XGBoost, and Neural Networks. Pooled performance metrics from 12 studies demonstrated a sensitivity of 0.88 (95% CI: 0.78–0.94, I2 = 90.43%), specificity of 0.79 (95% CI: 0.62–0.89, I2 = 99.53%), positive likelihood ratio of 4.1 (95% CI: 2.2–7.7), negative likelihood ratio of 0.16 (95% CI: 0.08–0.29), diagnostic odds ratio of 26 (95% CI: 10–71), and an AUROC of 0.91 (95% CI: 0.88–0.93), indicating excellent discriminative ability. Subgroup analyses revealed higher sensitivity in advanced ML models (89.7%) and non-USA studies (97.2%), with advanced ML showing lower specificity heterogeneity (I2 = 0%). Significant heterogeneity was observed, particularly in specificity (I2 = 99%), driven by traditional ML and USA-based studies. Minimal publication bias was noted for sensitivity (Egger's p = 0.942), but specificity showed potential bias (Egger's p = 0.038 after outlier exclusion). These findings suggest that ML models outperform traditional risk stratification tools in predicting PE mortality, offering robust potential for clinical decision-making, though heterogeneity and retrospective study designs warrant cautious interpretation.