Accuracy of artificial intelligence algorithms in predicting acute respiratory distress syndrome: a systematic review and meta-analysis.

Summary

Synthesizing 33 studies, AI models achieved pooled sensitivity 0.81, specificity 0.88, and AUC 0.91 for ARDS prediction, with CNN/SVM/XGB performing best and image-plus-multimodal inputs yielding the highest accuracy. The PROSPERO-registered review underscores AI’s clinical promise while highlighting model and predictor heterogeneity.

Key Findings

• Pooled diagnostic performance for ARDS prediction: sensitivity 0.81, specificity 0.88, AUC 0.91 across 33 studies.
• CNN, SVM, and XGB algorithms outperformed others; models using imaging plus other predictors achieved the highest AUC.
• Quality assessed with QUADAS-2; PROSPERO registered (CRD42023491546), supporting methodological transparency.

Clinical Implications

AI models, especially CNN/SVM/XGB with multimodal inputs, can support early ARDS identification and triage. Implementation should include external validation, calibration, and workflow integration to minimize false alarms and bias.

Limitations

• Heterogeneity in model types, predictors, and ARDS definitions across studies
• Predominantly retrospective model development with limited external validation, risking overfitting and bias

Future Directions

Prospective, multi-center impact studies integrating AI into ICU workflows; standardized ARDS definitions; fairness, calibration, and explainability evaluations.