Development and Validation of an Electronic Health Record-Based, Pediatric Acute Respiratory Distress Syndrome Subphenotype Classifier Model.

Summary

Using two temporally distinct PARDS cohorts, the authors trained an EHR-only machine learning model that accurately classified biomarker-defined hyperinflammatory vs hypoinflammatory subphenotypes. A parsimonious 5-laboratory-variable model maintained high performance (AUC ~0.92) at 24 hours after diagnosis, suggesting feasibility for early precision stratification without biomarker assays.

Key Findings

• EHR-only classifier achieved AUC 0.93 (95% CI 0.87–0.98) with 88% sensitivity and 83% specificity for hyperinflammatory PARDS.
• A parsimonious model using five laboratory variables achieved AUC 0.92 (95% CI 0.86–0.98) with 76% sensitivity and 87% specificity.
• Subphenotypes were defined via biomarker-guided latent class analysis, providing a biological reference standard.
• Classification feasible at 24 hours after PARDS diagnosis, supporting early risk stratification.

Clinical Implications

Clinicians could use EHR-derived lab values within 24 hours to identify hyperinflammatory PARDS, informing escalation, adjunctive therapies, and enrollment into phenotype-specific trials.

Limitations

• Single-center study limits generalizability
• Retrospective design with potential unmeasured confounding
• External validation was temporal but not multicenter
• Clinical impact on outcomes not tested prospectively

Future Directions

Multicenter, prospective validation and testing of phenotype-guided therapy or trial stratification using the parsimonious model.