Diagnostic accuracy of AI-Based models for pulmonary edema detection: A systematic review and Meta-Analysis

Pulmonary edema is a life-threatening condition caused by fluid accumulation in the lungs that impairs gas exchange. Machine learning models using clinical and imaging data can improve early detection and diagnostic accuracy.

This systematic review and meta-analysis evaluate ML-models performance and key features to detect pulmonary edema.

Following PRISMA guidelines, we searched five databases using MeSH and Emtree terms across Pubmed, Scopus, Web of science, Embase, Ebsco. Two reviewers screened studies, extracted data, and assessed quality using PROBAST+AI and GRADE. Pooled sensitivity, specificity, and AUC calculated using STATA 18.0, with heterogeneity explored via meta-regression, subgroup, and sensitivity analyses. Publication bias assessed using Deek's funnel plot.

Total of 14 studies met inclusion criteria and 10 studies included in Meta-analysis. Machine learning models demonstrated high pooled sensitivity (0.90, 95% CI: 0.78-0.96), specificity (0.89, 95% CI: 0.70-0.96), and AUROC (0.95, 95% CI: 0.93-0.97) for pulmonary edema detection, suggesting promising diagnostic potential. However, substantial heterogeneity (I² = 79.36%) observed, indicating variability across study populations, imaging modalities, and analytic approaches. Sensitivity analyses confirmed that no single study disproportionately influenced results, and no evidence of publication bias was detected (p = 0.41). Risk-of-bias assessments were moderate, though applicability concerns limited generalizability in some studies. Given the high heterogeneity, the certainty of evidence graded as low.

AI models demonstrate good diagnostic accuracy for pulmonary edema, particularly with imaging data, though heterogeneity limits generalizability. Low certainty of evidence underscores the need for standardized approaches.