Comparison of artificial intelligence and logistic regression models for mortality prediction in acute respiratory distress syndrome: a systematic review and meta-analysis.

Summary

Across 8 studies, AI models achieved pooled sensitivity 0.89, specificity 0.72, and SROC 0.84 for ARDS mortality prediction, outperforming logistic regression (SROC 0.81). Performance was stronger in moderate-to-severe ARDS, highlighting severity-dependent accuracy.

Key Findings

• AI models showed pooled sensitivity 0.89, specificity 0.72, and SROC 0.84 in validation sets
• Logistic regression models had lower performance (SROC 0.81; sensitivity 0.78; specificity 0.68)
• Prediction accuracy was higher in moderate-to-severe ARDS (SAUC 0.84 vs 0.81)
• Heterogeneity and disease severity influenced model accuracy; QUADAS-2 used to assess bias

Clinical Implications

AI-based prognostic models may improve early risk stratification, triage, and resource allocation for ARDS, particularly in moderate-to-severe cases; external validation, calibration, and workflow integration are prerequisites for clinical adoption.

Limitations

• Only eight studies with potential heterogeneity and varying definitions
• Limited external validation and reporting standards across included studies

Future Directions

Prospective multicenter external validation, standardized reporting (e.g., TRIPOD-AI), calibration and impact analyses, and evaluation of deployment across care settings.