Identification of Acute Respiratory Failure Phenotypes With Electronic Health Record Data

Abstract

Background

Secondary analysis of clinical trial data and highly selected observational cohorts have identified 2 subphenotypes in acute respiratory failure, but have not been reported previously using only real-world electronic health record (EHR) data.

Research Question

Are subphenotypes of acute ventilator-dependent respiratory failure identifiable using readily available EHR data?

Study Design and Methods

This multicenter retrospective cohort study used patient encounters from the Medical University of South Carolina (n = 4,233 between 2016 and 2023) and the Medical Information Mart for Intensive Care III (n = 8,313 between 2001 and 2012) to train and validate K-means models with multiply imputed cluster analysis at 24 and 48 hours after intubation.

Results

Clustering models identified 2 clusters for 24-hour and 48-hour models in both training and test cohorts with clusters separating on variables related to pulmonary physiology, perfusion, organ dysfunction, and metabolic dysregulation. Cluster 2 showed higher 90-day mortality after discharge and more ventilator days compared with cluster 1 that persisted despite multivariable adjustment for age, illness severity, and comorbidities. Cluster models and clusters were stable in 0- to 24-hour and 25- to 48-hour models with crossover (29.2% and 25.9% of the test and training cohorts) from the higher-acuity cluster 2 to the lower-acuity cluster 1 subphenotype occurring by 48 hours after intubation.

Interpretation

Our results suggest that acute ventilator-dependent respiratory failure has 2 subphenotypes that are discernible using readily available data from EHRs with identifiable differences in pulmonary physiologic features, perfusion, organ dysfunction, and metabolic dysregulation at 24 and 48 hours after intubation. This may enable future EHR tools to identify particularly vulnerable patients.