Development and validation of a biomarker-based model to predict persistent hypoxemic respiratory failure among mechanically ventilated adults

Neha Sathe | 2021

Advisor: Alison E. Fohner

Research Area(s): Clinical Epidemiology

Full Text

Background. Acute hypoxemic respiratory failure (HRF) requiring invasive mechanical ventilation (IMV) is a common and morbid condition, but treatment is largely supportive and most pharmacologic trials have been negative. The early clinical trajectory is heterogeneous, with most patients resolving within days while others require prolonged IMV, which impairs both clinical resource allocation and the detection of therapeutic effects in trials. Our objectives were to (1) compare mortality between those with persistent HRF at day 3 to those with resolving HRF, and (2) develop and validate a model to predict persistent HRF.
Methods. I performed a secondary analysis of patients with acute HRF in the intensive care unit (ICU), enrolled in 2 independent prospective cohorts (discovery cohort n = 632, validation cohort n = 906). I used generalized linear models to estimate the relative risk (RR) of persistent HRF compared to resolving HRF in each cohort, adjusting for age, sex and the acute physiology and chronic health evaluation score (APACHE-II). For development of the prediction model, I split the entire discovery cohort into training (n=474) and testing (n=158) sets, and used the second cohort for external validation. I applied LASSO to 33 candidate clinical predictors (spanning demographics, comorbidities, vitals, standard laboratory values, and illness severity scores) for parsimonious model selection. Then I examined whether a combination of log2 transformed research biomarkers previously predictive of ICU outcomes (interleukin-6 [IL-6], interleukin-8 [IL-8], angiopoietin-2 [Ang-2], soluble tumor necrosis factor receptor 1 [sTNFR-1], measured in a subset) could improve model performance. I calculated area under the curve (AUC) for model performance.
Results. Approximately half of patients in the discovery and validation cohorts had persistent HRF at day 3 (298/632 in discovery and 514/906 in validation). At cohort enrollment, patients with persistent HRF had higher proportion of shock, pneumonia, as well as higher ICU illness severity scores. In both cohorts, persistent HRF was associated with an over 2-fold higher risk of death compared to resolving HRF (aRR 2.33, 95% CI 1.42, 3.82 in discovery; aRR 2.05, 95% CI 1.51, 2.78) after adjusting for age, sex, and a measure of baseline arterial oxygenation (P/F). In models adjusting for an overall illness severity score (APACHE-II) instead of P/F, this effect was attenuated in the discovery cohort but remained significant in the validation cohort (aRR 1.49, 95% CI 0.92, 2.41 in discovery; aRR 1.81, 95% CI 1.32, 2.44 in validation). A LASSO derived model to predict persistent HRF had moderate performance in the training (AUC 0.79, 95% CI 0.75, 0.83) and test (AUC 0.76, 95% CI 0.68, 0.83) cohorts, but had limited performance in external validation (AUC 0.69, 85% CI 0.65, 0.72). A model adding IL-6 and Ang-2 significantly improved performance across all 3 cohorts (training AUC 0.82, 95% CI 0.77, 0.86; test AUC 0.78, 95% CI 0.69, 0.87; validation AUC 0.72, 95% CI 0.65, 0.79).
Conclusions. Persistent HRF is associated with a high risk for mortality compared to resolving HRF even when adjusting for initial illness severity. A parsimonious model of P/F, APACHE-II, IL-6, and Ang-2 has moderate performance for predicting persistent HRF. This can be implemented to direct care to high-risk patients early, and to improve prognostic enrichment in trials for acute HRF.