Reproducible clinical archetypes in acute respiratory failure: a multi-cohort trajectory analysis

Area of Science:

• Critical Care Medicine
• Pulmonary Medicine
• Data Science in Healthcare

Background:

• Acute hypoxemic respiratory failure (AHRF) is a common critical illness with significant biological heterogeneity.
• Static phenotyping fails to capture the dynamic evolution of AHRF, potentially leading to misclassification and suboptimal treatment.
• Identifying distinct clinical trajectories may improve understanding and management of AHRF.

Purpose of the Study:

• To identify and externally validate reproducible clinical course archetypes, termed trajectory classes (TCs), for persistent AHRF.
• To develop models for early prediction of these TCs.
• To explore the association of TCs with ARDS and hyperinflammatory subphenotypes.

Main Methods:

• Analysis of MIMIC-IV and two external cohorts (n=10,418) of adults with AHRF.
• Joint modeling of daily oxygenation (PaO2/FiO2) and time to ICU discharge/death using a competing-risk latent class mixed model.
• Development of an XGBoost model for early TC prediction using 12 variables.

Main Results:

• A four-class model optimally described AHRF trajectories: early recovery, stable persistence, biphasic deterioration, and rapid decline (100% mortality).
• These archetypes generalized well to external cohorts with high certainty.
• Early prediction models achieved AUCs >= 0.78 by day 3.
• TC4 was enriched for hyperinflammatory subphenotypes, and TC2 was common in ARDS patients.

Conclusions:

• Four reproducible oxygenation archetypes effectively characterize the 14-day course of persistent AHRF.
• These trajectories offer early prognostic value beyond baseline severity.
• The identified TCs can potentially guide therapeutic strategies, reduce heterogeneity in clinical trials, and inform biological phenotyping efforts.