Development and External Validation of a Detection Model to Retrospectively Identify Patients With Acute Respiratory Distress Syndrome.

Summary

A regularized logistic regression model combining structured EHR data and radiology reports retrospectively identified ARDS with AUROC 0.91 internally and 0.88 externally, calibrated with external ICI 0.13. At a chosen threshold, sensitivity and specificity were both 80% and PPV 64%, identifying cases a median of 2.2 hours after meeting Berlin criteria across two health systems.

Key Findings

• Training, internal validation, and external validation cohorts included 1,845, 556, and 199 patients; ARDS prevalence was 19%, 17%, and 31%, respectively.
• EHR-radiology model achieved AUROC 0.91 (internal) and 0.88 (external) with external ICI 0.13.
• At a set threshold, sensitivity and specificity were 80% and PPV 64%; cases were identified a median 2.2 hours after Berlin criteria.

Clinical Implications

Hospitals can deploy retrospective ARDS detection for case ascertainment, benchmarking, and research screening; future prospective integration could support earlier recognition and standardized phenotyping.

Limitations

• Retrospective design; performance depends on data availability and documentation quality
• Generalizability beyond two health systems and portability of NLP for radiology may vary

Future Directions

Prospective, real-time deployment to assess impact on early recognition and outcomes; transportability studies; open sharing of code and definitions to enhance reproducibility.