Establishment and validation of predictive model of ARDS in critically ill patients.

Summary

In a Chinese prospective cohort (derivation n=400; external n=160), a logistic model combining sex, LIPS, hepatic disease, shock, and lung contusion predicted incident ARDS with AUC 0.836 internally and 0.799 externally, outperforming LIPS alone. SHAP aided interpretability and decision curve analysis demonstrated net clinical benefit.

Key Findings

• Prospective derivation cohort (n=400) with 117 ARDS events; external validation cohort (n=160) with 44 ARDS events.
• Final logistic model variables: sex, Lung Injury Prediction Score (LIPS), hepatic disease, shock, and lung contusion.
• Internal validation AUC 0.836 (95% CI 0.762–0.910); external validation AUC 0.799 (95% CI 0.723–0.875).
• Outperformed LIPS alone in discrimination and decision curve analysis; SHAP improved model interpretability.

Clinical Implications

Embedding this model (or its key variables) into ICU triage could prompt closer monitoring and early lung-protective strategies (e.g., conservative fluids, timely ventilatory support) for high-risk patients.

Limitations

• Single-country, two-center cohorts with modest sample size may limit generalizability
• Clinical impact of model-guided interventions not tested; potential calibration drift over time not assessed

Future Directions

Multicenter international validation, real-time EHR integration, impact trials testing model-triggered prevention strategies, and periodic recalibration.