Objectives: Patients with severe, persistent hypoxemic respiratory failure have a higher mortality. Early identification is critical for informing clinical decisions, using rescue strategies, and enrollment in clinical trials. The objective of this investigation was to develop and validate a prediction model to accurately and timely identify patients with severe hypoxemic respiratory failure at high risk of death, in whom novel rescue strategies can be efficiently evaluated.
Design: Electronic medical record analysis.
Setting: Medical, surgical, and mixed ICU setting at a tertiary care institution.
Patients: Mechanically-ventilated ICU patients.
Measurements and main results: Mechanically ventilated ICU patients were screened for severe hypoxemic respiratory failure (Murray lung injury score of ≥ 3). Survival to hospital discharge was the dependent variable. Clinical predictors within 24 hours of onset of severe hypoxemia were considered as the independent variables. An area under the curve and a Hosmer-Lemeshow goodness-of-fit test were used to assess discrimination and calibration. A logistic regression model was developed in the derivation cohort (2005-2007). The model was validated in an independent cohort (2008-2010). Among 79,341 screened patients, 1,032 met inclusion criteria. Mortality was 41% in the derivation cohort (n = 464) and 35% in the validation cohort (n = 568). The final model included hematologic malignancy, cirrhosis, aspiration, estimated dead space, oxygenation index, pH, and vasopressor use. The area under the curve of the model was 0.85 (0.82-0.89) and 0.79 (0.75-0.82) in the derivation and validation cohorts, respectively, and showed good calibration. A modified model, including only physiologic variables, performed similarly. It had comparable performance in patients with acute respiratory distress syndrome and outperformed previous prognostic models.
Conclusions: A model using comorbid conditions and physiologic variables on the day of developing severe hypoxemic respiratory failure can predict hospital mortality.