AI models accurately detected pulmonary embolism on imaging, though accuracy dropped when tested in external datasets overall, too.
Artificial Intelligence and Pulmonary Embolism Diagnosis
Missed or delayed pulmonary embolism diagnosis is linked with worse outcomes and longer hospital stays, making reliable imaging interpretation essential. In a new systematic review and meta-analysis, researchers evaluated how well AI models detect pulmonary embolism across imaging-based workflows, focusing on pooled diagnostic accuracy and whether results held up beyond the original development settings.
The team searched major databases from inception to January 1, 2025, then screened records and assessed eligible full texts in duplicate. Study quality and risk of bias were evaluated using QUADAS-2. The analysis pooled sensitivity and specificity and summarized overall discrimination using the area under the ROC curve, with random-effects models to account for between-study differences.
Diagnostic Accuracy Across Imaging Validation Settings
Across 28 included studies, internal validation phases covered 43,330 participants, including 4,866 pulmonary embolism positive cases. In this setting, AI achieved a pooled sensitivity of 0.91 and pooled specificity of 0.94, with an AUC of 0.95, suggesting excellent overall accuracy for pulmonary embolism detection.
External validation phases included 3,588 participants, with 1,699 pulmonary embolism positive cases. Here, performance was slightly lower, with pooled sensitivity of 0.89 and pooled specificity of 0.88, while the AUC remained high at 0.94. The authors interpreted this drop from internal to external validation as an important signal that model generalizability may be limited when tools are applied to new populations or settings.
Interpreting Heterogeneity and Generalizability
Despite strong pooled estimates, the review found substantial heterogeneity across studies, reflected by high I² values in several metrics. Subgroup analyses and meta-regression were used to explore potential sources of variation, and leave-one-out analyses were performed to test robustness. No significant publication bias was detected.
Overall, the findings suggest AI can support pulmonary embolism detection through imaging, but clinicians and health systems should weigh the observed heterogeneity and the consistent performance decrease in external validation when interpreting real-world readiness.
Reference: Farzaei A et al. The Role of Artificial Intelligence in Diagnosing Pulmonary Embolism: A Systematic Review and Meta-analysis. Arch Acad Emerg Med. 2025;13(1):e86.
