Open-source computational pipeline flags instances of acute respiratory distress syndrome in mechanically ventilated adult patients.

Summary

An open-source pipeline that operationalizes the Berlin Definition identified ARDS from radiology reports and clinician notes with 93.5% sensitivity and a 17.4% false positive rate on an external dataset. Performance far exceeded the cohort’s 22.6% ARDS documentation rate, highlighting under-recognition and the potential of automated adjudication to improve timely diagnosis.

Key Findings

• Automated ARDS adjudication achieved 93.5% sensitivity with a 17.4% false positive rate on an external, publicly available dataset.
• The pipeline operationalizes the Berlin Definition using interpretable classifiers applied to radiology reports and physician notes.
• Observed performance far exceeded the cohort’s 22.6% ARDS documentation rate, revealing extensive under-recognition.

Clinical Implications

Integration into EHRs could enable real-time ARDS flagging to prompt lung-protective ventilation, early proning, and timely consultation. Prospective evaluation is needed to confirm clinical impact and mitigate false positives.

Limitations

• Retrospective design without prospective clinical impact assessment
• Reliance on free-text radiology reports and notes can introduce misclassification
• Non-trivial false positive rate (17.4%) may cause alert fatigue if deployed

Future Directions

Prospective, multi-center trials integrating the tool into EHR workflows to assess effects on time-to-lung-protective ventilation, proning, and patient outcomes; domain adaptation and bias assessments across hospitals.