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

Summary

Using structured EHR data plus radiology reports, a regularized logistic model achieved AUROC 0.91 (internal) and 0.88 (external) with good calibration (ICI 0.13). At a set threshold, sensitivity/specificity were both 80%, PPV 64%, and cases were identified a median 2.2 hours after Berlin criteria, enabling robust retrospective ARDS phenotyping across health systems.

Key Findings

• Regularized logistic EHR+radiology model achieved AUROC 0.91 (internal) and 0.88 (external), with external ICI 0.13.
• At a prespecified threshold, sensitivity and specificity were both 80%, with PPV 64%.
• Model identified ARDS a median 2.2 hours after meeting Berlin criteria, enabling timely retrospective capture.
• Physician-adjudicated labels and cross-system validation support generalizability.

Clinical Implications

Hospitals and researchers can apply this validated EHR+radiology model to consistently identify ARDS cases for quality improvement, surveillance, and research; prospective adaptation could support earlier recognition and trial enrollment.

Limitations

• Retrospective design; model identifies ARDS shortly after criteria are met rather than predicting pre-onset.
• External validation limited to two health systems; performance in broader settings and languages requires testing.

Future Directions

Prospective implementation for early ARDS recognition, integration with bedside alerts, and testing across diverse health systems and international datasets.