A contextualized universal vital assessment score better predicts in-hospital mortality

Abstract

Rationale: Illness severity scores guide triage, characterize patient populations, and predict mortality. The Universal Vital Assessment (UVA) Score employs data available at presentation, but its ability to predict mortality in patients with acute hypoxemic respiratory failure (AHRF) in resource-constrained settings is only modest, with an area under the receiver-operating characteristic (AUROC) curve of 0.69 in a prior study. The UVA score includes temperature, heart rate, respiratory rate, blood pressure, Glasgow Coma Score, HIV infection, and oxygen saturation. Contextualizing oxygenation to include level of respiratory support using the imputed partial pressure of arterial oxygen to fraction of inspired oxygen ratio (iP/F) could improve prognostication.Methods: We analyzed data from five East African referral hospitals in the Before Building Respiratory support in East Africa Through High flow versus standard flow oxygen Evaluation study (Before BREATHE), a prospective cohort study of adults with AHRF, and BREATHE (NCT05754034), a randomized controlled trial comparing high-flow oxygen support to standard-flow oxygen support. We analyzed a derivation cohort (n=2131) composed of 70% of the full cohort, reserving 30% of the cohort for validation after the contextualized UVA score is fully determined. We assessed the association of clinical characteristics with mortality using the Chi-square and Wilcoxon rank-sum tests. To identify a threshold in the prognostic association of iP/F and mortality, we employed decision tree analysis (rpart package in R) and pruned the tree using the largest complexity parameter whose cross-validation error was within one standard error of the minimum cross-validation error. We then built a logistic regression model using the original UVA score components, with iP/F replacing peripheral oxygen saturation. We assigned point values to iP/F severity categories by indexing their beta coefficient to the beta coefficient of the original UVA variable that performed best in this population. Results: In this population, optimal discernment of in-hospital mortality occurred at iP/F ≤100, which was associated with 54.2% absolute in-hospital mortality compared to 34.4% in-hospital mortality in patients with iP/F>100. Based on logistic regression, severe impairment of oxygenation was assigned 2 UVA points. The area under the receiver-operating characteristic (AUROC) curve for the contextualized UVA score was 0.730 (95% confidence interval [CI] 0.698 - 0.762), compared to 0.715 (95%CI 0.682-0.748) for the original UVA score in this population. The p-value calculated by the DeLong test for receiver-operating curve comparison was 0.03. Inability to ambulate independently was also significantly associated with mortality (p<0.001). Conclusions: In this multinational cohort of adults with AHRF in East Africa, contextualization of the UVA score to incorporate the degree of respiratory support was modestly more predictive of in-hospital mortality compared to the original score. Future studies should evaluate whether adding a point for mild-to-moderate hypoxemia improves the score in a heterogeneous cohort, whether adding an ability-to-walk variable further improves discernment, and whether this result can be replicated in the validation cohort and in other cohorts.