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

Area of Science:

• Critical Care Medicine
• Pulmonary Medicine
• Biostatistics

Background:

• Acute lung injury (ALI) and acute respiratory distress syndrome (ARDS) are significant complications in acute care.
• Existing risk prediction models for ALI/ARDS show performance variability due to differing criteria and methods.
• This review assesses prediction model performance for ALI/ARDS and related pulmonary issues.

Purpose of the Study:

• To systematically review and meta-analyze the performance of prediction models for ALI/ARDS.
• To identify sources of heterogeneity in model performance.
• To evaluate models for acute pulmonary complications in high-risk populations.

Main Methods:

• Searched Cochrane Library, PubMed, EMBASE, and Web of Science.
• Included studies developing or validating ALI/ARDS prediction models.
• Extracted data and assessed bias using PROBAST; calculated pooled C-index and accuracy.

Main Results:

• Included 14 studies with 5,608 participants; outcome definitions varied.
• Pooled training accuracy was 0.732, C-index 0.80; test set accuracy 0.803, C-index 0.81.
• Logistic regression was common; penalized models like LASSO showed promise but were less evaluated.

Conclusions:

• Prediction models for ALI/ARDS demonstrate good discrimination, with logistic regression widely used.
• Heterogeneity in definitions, timing, validation, and reporting impedes clinical use.
• Standardized definitions, clear time points, and transparent external validation are needed for future research.