High-fidelity discrimination of ARDS versus other causes of respiratory failure using natural language processing and iterative machine learning

Despite the high morbidity and mortality associated with Acute Respiratory Distress Syndrome (ARDS), discrimination of ARDS from other causes of acute respiratory failure remains challenging, particularly in the first 24 hours of mechanical ventilation. Delay in ARDS identification prevents lung protective strategies from being initiated and delays clinical trial enrolment and quality improvement interventions. Medical records from 1,263 ICU-admitted, mechanically ventilated patients at Northwell Health were retrospectively examined by a clinical team who assigned each patient a diagnosis of "ARDS" or "non-ARDS" (e.g., pulmonary edema). We then applied an iterative pre-processing and machine learning framework to construct a model that would discriminate ARDS versus non-ARDS, and examined features informative in the patient classification process. Data made available to the model included patient demographics, laboratory test results from before the initiation of mechanical ventilation, and features extracted by natural language processing of radiology reports. The resulting model discriminated well between ARDS and non-ARDS causes of respiratory failure (AUC=0.85, 89% precision at 20% recall), and highlighted features unique among ARDS patients, and among and the subset of ARDS patients who would not recover. Importantly, models built using both clinical notes and laboratory test results out-performed models built using either data source alone, akin to the retrospective clinician-based diagnostic process. This work demonstrates the feasibility of using readily available EHR data to discriminate ARDS patients prospectively in a real-world setting at a critical time in their care and highlights novel patient characteristics indicative of ARDS.