Prediction models for the occurrence and mortality of sepsis-associated lung injury: a systematic review and meta-analysis
-
By
-
Chen Liu
-
Jian Huo
-
Yan-Song Li
-
An-Min Hu
-
Ting-Ting Ao
-
June 9, 2026
-
Objective:
To synthesize and quantitatively evaluate prediction models for sepsis-associated lung injury and short-term mortality.
Key Findings:
- Nine studies included, primarily from China, reporting 68 model phase units.
- Pooled test-phase AUC for ARDS occurrence was 0.749.
- Pooled AUCs for short-term mortality were 0.800 (training), 0.778 (validation), and 0.815 (testing).
- High heterogeneity observed, particularly for ARDS occurrence and mortality training models.
- Machine learning models did not consistently outperform logistic regression.
Interpretation:
Current models showed moderate discrimination but are limited by bias, weak methods, and heterogeneity.
Limitations:
- High overall risk of bias in 4 studies and unclear risk in 6.
- Certainty of evidence was low for all outcome families and modeling phases.
- Heterogeneity in model performance across studies.
Conclusion:
Models for predicting ARDS occurrence and mortality in sepsis need separate development and validation.
