Machine learning for predicting acute lung injury after trauma: a systematic review and meta-analysis.

BACKGROUND

Acute lung injury (ALI) and acute respiratory distress syndrome (ARDS) remain major complications in trauma and other high-risk acute-care populations. Numerous statistical and machine-learning models have been developed for risk prediction, but their reported performance may vary because of differences in diagnostic criteria, predictor timing, validation strategies, and clinical settings. This systematic review and meta-analysis evaluated the performance of prediction models for ALI/ARDS and related acute pulmonary complications and explored potential sources of heterogeneity.

METHODS

The Cochrane Library, PubMed, EMBASE, and Web of Science were searched through August 13, 2025. Eligible studies developed or validated prediction models for ALI, ARDS, or related acute pulmonary complications in high-risk acute-care populations. Two reviewers independently screened studies, extracted data, and assessed risk of bias using PROBAST. Pooled C-index and accuracy were calculated when comparable estimates were available. Exploratory subgroup analyses examined diagnostic criteria, study design, and predictor timing. Sensitivity analyses included restriction to trauma- or burn-related studies and leave-one-out analyses.

RESULTS

Fourteen studies involving 5,608 participants were included. Outcome definitions varied and were based on the American-European Consensus Conference criteria, the Berlin definition, or other criteria. Logistic regression was the most commonly evaluated approach. In training datasets, pooled accuracy and C-index were 0.732 (95% CI, 0.694-0.768) and 0.80 (95% CI, 0.76-0.83), respectively. In test datasets, the corresponding estimates were 0.803 (95% CI, 0.756-0.843) and 0.81 (95% CI, 0.77-0.86). Penalized regression models, particularly LASSO, showed high discrimination in individual studies but were less frequently evaluated. Subgroup analyses suggested a lower pooled training C-index in studies using AECC criteria than in those using the Berlin definition or other criteria. Sensitivity analyses yielded results consistent with the primary findings.

CONCLUSION

Prediction models for ALI/ARDS and related acute pulmonary complications showed generally good discrimination, with logistic regression remaining the most widely used approach. However, heterogeneous outcome definitions, inconsistent predictor timing, limited external validation, and incomplete reporting continue to hinder clinical translation. Future studies should standardize outcome definitions, clearly specify prediction time points and horizons, and prioritize transparent multicenter external validation.

TRIAL REGISTRATION: PROSPERO (Registration No. CRD42024576566).