The value of local validation of a predictive model. A nomogram for predicting failure of non-invasive ventilation in patients with SARS-COV-2 pneumonia.

Abstract

Objective: We aimed to determine predictors of non-invasive ventilation (NIV) failure and validate a nomogram to identify patients at risk of NIV failure.

Design: Observational, analytical study of a retrospective cohort from a single center, compared with an external cohort (March 2020 to August 2021).

Setting: Two intensive care units (ICUs).

Patients: Patients with pneumonia due to severe acute respiratory syndrome (SARS-CoV-2) and NIV > 24 h (154 and 229 in each cohort).

Interventions: The training cohort identified NIV failure predictors. A nomogram, created via logistic regression, underwent validation with the Hosmer-Lemeshow (HL), calibration curve and test and area under the curve (AUC). Its external validity was tested using AUC.

Main variables of interest: Demographics, comorbidities, severity scores, NIV settings, vital signs, blood gases, and oxygenation at the start and 24 h after NIV, NIV failure.

Results: NIV failure was 37.6% and 18% in the training and validation cohorts, respectively. Risk factors for NIV failure inluded age, obesity, sequential organ failure assessment (SOFA) score at admission, and heart rate (HR) and heart rate, acidosis, consciousness, oxygenation, respiratory rate (HACOR) 24 h post-NIV. The model's HL test result was 0.861, with an AUC of 0.89 (confidence interval [CI] 0.839-0.942); validation AUC was 0.547 (CI 0.449-0.645).

Conclusions: A predictive model using age, obesity, SOFA score, HR, and HACOR at 24 h predicts NIV failure in our COVID-19 patients but may not apply to other ICUs.