Abstract
Physicians, particularly intensivists, face information overload and decision fatigue, underscoring the need for automated diagnostic tools. Acute Respiratory Distress Syndrome (ARDS) affects over 10% of critical care patients, with over 40% mortality rate, yet is only recognized in 30-70% of cases in clinical settings. We present a reproducible computational pipeline that automates ARDS adjudication in retrospective datasets of mechanically ventilated adults, implementing the Berlin Definition via natural language processing and classification algorithms. We used labeled chest imaging reports from two hospitals to train an XGBoost model to detect bilateral infiltrates, and a labeled subset of attending physician notes from one hospital to train another XGBoost model to detect a pneumonia diagnosis. Both models achieve high discriminative performance on test sets-an area under the receiver operating characteristic curve (AUROC) of 0.88 for adjudicating bilateral infiltrates on chest imaging reports, and an AUROC of 0.87 for detecting pneumonia on attending physician notes. We integrated these models with rule-based components and validated the entire pipeline on a subset of healthcare encounters from a third hospital (MIMIC-III). We find a sensitivity of 93.5% in adjudicating ARDS - far surpassing the 22.6% ARDS documentation rate we found for this cohort - along with a false positive rate of 17.4%. We conclude that our reproducible, automated pipeline holds promise for improving ARDS recognition and could aid clinical practice through real-time EHR integration.