Dynamic prognostic prediction in sepsis using longitudinal blood gas trajectories: development and external...

Area of Science:

• Critical Care Medicine
• Physiological Monitoring
• Data Science in Healthcare

Background:

• Sepsis prognosis relies on timely and accurate prediction of patient outcomes.
• Arterial blood gas (ABG) parameters reflect critical physiological status during sepsis.
• Conventional clinical variables may not fully capture the dynamic nature of sepsis progression.

Purpose of the Study:

• To develop and validate a dynamic prognostic framework for sepsis using longitudinal ABG data.
• To assess the incremental prognostic information provided by ABG-derived dynamic features compared to clinical variables.

Main Methods:

• Retrospective observational study using MIMIC-IV and eICU-CRD databases.
• Analysis of eight ABG variables (PaO2, PaCO2, FiO2, PaO2/FiO2, pH, base excess, lactate, glucose) in 12-h windows.
• Latent class growth analysis and growth mixture modeling to identify ABG trajectories; elastic-net logistic regression for dynamic prediction models.

Main Results:

• Identified clinically interpretable longitudinal ABG trajectories reflecting sepsis heterogeneity.
• Models integrating ABG features showed moderate discrimination for in-hospital mortality prediction.
• External validation demonstrated small-to-moderate improvements (ROC-AUC increase of 0.036-0.055) with ABG features, but gains were inconsistent.

Conclusions:

• Longitudinal ABG trajectories capture dynamic physiological heterogeneity in sepsis.
• ABG-derived dynamic features offer limited complementary prognostic information beyond conventional variables.
• Clinical utility requires prospective validation due to modest incremental benefit and inconsistent performance.