Arch Acad Emerg Med. 2025 Dec 30;13(1):e86. doi: 10.22037/aaem.v13i1.2720. eCollection 2025.
ABSTRACT
INTRODUCTION: Missed or delayed diagnosis of pulmonary embolism (PE) is associated with increased morbidity, mortality, and longer hospitalizations. This study aimed to evaluate the diagnostic accuracy of Artificial Intelligence (AI) models in detecting PE across imaging.
METHODS: We systematically searched PubMed/MEDLINE, Scopus, Embase and Web of Science from inception to 1 January 2025 without language or regional limits. After removing duplicate results, the remaining records were screened through titles/abstracts, and two reviewers independently assessed full texts. Risk of bias was evaluated in duplicate with the QUADAS-2 tool. Pooled sensitivity, specificity, positive and negative likelihood ratios, diagnostic odds ratio and area under the ROC curve were calculated with random-effects models in STATA 17. Heterogeneity was quantified with Cochran’s Q and I², while we explored its sources using subgroup analyses (for categorical moderators) and meta-regression (for continuous moderators). Publication bias was assessed with Deeks’ funnel plot and trim-and-fill, and we examined robustness through leave-one-out sensitivity analyses.
RESULTS: A total of 1,432 records were identified through database searches, with 654 duplicates removed. After screening titles and abstracts of 787 articles, 256 full-text articles were assessed for eligibility, and 28 studies met the inclusion criteria. Internal validation phases included 43,330 participants (4,866 PE-positive, 38,463 PE-negative), while external validation phases comprised 3,588 participants (1,699 PE-positive, 1,889 PE-negative). In the internal validation phase, the pooled sensitivity and specificity of AI in PE diagnosis across imaging were 0.91 (95% confidence interval (CI): 0.88-0.95; I²=9%) and 0.94 (95% CI: 0.86-0.98; I²=99.78%), respectively. The positive likelihood ratio (PLR) was 16.08, and the negative likelihood ratio (NLR) was 0.09, both statistically significant (P < 0.001). The pooled diagnostic odds ratio (DOR) was 163.55 (95% CI: 71.30-375.14, I2: 96.1), and the area under the curve (AUC) was 0.95 (95% CI: 0.93 to 0.97), indicating excellent accuracy. In external validation, the pooled sensitivity and specificity were slightly lower at 0.89 (95% CI: 0.79-0.95; I²=95.60%) and 0.88 (95% CI: 0.80-0.93; I²=91.48%), respectively. The DOR was 59.65 (95% CI: 23.53 to 151.17, I2: 89.6) and AUC was 0.94 (95% CI: 0.92 to 0.96, I2: 89.6). There was no significant publication bias detected.
CONCLUSION: AI models achieved high diagnostic accuracy in detecting PE through imaging. However, this performance tends to decrease from internal to external validation, highlighting limitations in generalizability. Additionally, substantial heterogeneity was observed across studies, as indicated by high I² values, which should be considered when interpreting the pooled estimates.
PMID:41664805 | PMC:PMC12883040 | DOI:10.22037/aaem.v13i1.2720
