Biomarker-Based Risk Stratification Tool in Pediatric Acute Respiratory Distress Syndrome: Single-Center, Longitudinal Validation in a 2014–2019 Cohort

Abstract

Objectives: The Pediatric Acute Respiratory Distress Syndrome Biomarker Risk Model (PARDSEVERE) used age and three plasma biomarkers measured within 24 hours of pediatric acute respiratory distress syndrome (ARDS) onset to predict mortality in a pilot cohort of 152 patients. However, longitudinal performance of PARDSEVERE has not been evaluated, and it is unclear whether the risk model can be used to prognosticate after day 0. We, therefore, sought to determine the test characteristics of PARDSEVERE model and population over the first 7 days after ARDS onset. Design: Secondary unplanned post hoc analysis of data from a prospective observational cohort study carried out 2014–2019. Setting: University-affiliated PICU. Patients: Mechanically ventilated children with ARDS. Interventions: None. Measurements and Main Results: Between July 2014 and December 2019, 279 patients with ARDS had plasma collected at day 0, 266 at day 3 (11 nonsurvivors, two discharged between days 0 and 3), and 207 at day 7 (27 nonsurvivors, 45 discharged between days 3 and 7). The actual prevalence of mortality on days 0, 3, and 7, was 23% (64/279), 14% (38/266), and 13% (27/207), respectively. The PARDSEVERE risk model for mortality on days 0, 3, and 7 had area under the receiver operating characteristic curve (AUROC [95% CI]) of 0.76 (0.69–0.82), 0.68 (0.60–0.76), and 0.74 (0.65–0.83), respectively. The AUROC data translate into prevalence thresholds for the PARDSEVERE model for mortality (i.e., using the sensitivity and specificity values) of 37%, 27%, and 24% on days 0, 3, and 7, respectively. Negative predictive value (NPV) was high throughout (0.87–0.90 for all three-time points). Conclusions: In this exploratory analysis of the PARDSEVERE model of mortality risk prediction in a population longitudinal series of data from days 0, 3, and 7 after ARDS diagnosis, the diagnostic performance is in the “acceptable” category. NPV was good. A major limitation is that actual mortality is far below the prevalence threshold for such testing. The model may, therefore, be more useful in cohorts with higher mortality rates (e.g., immunocompromised, other countries), and future enhancements to the model should be explored.