PREDICTING SEPSIS-INDUCED HYPOTENSION PATIENT ATTRIBUTES FOR RESTRICTIVE VERSUS LIBERAL FLUID STRATEGY.

Area of Science:

• Critical Care Medicine
• Health Informatics
• Machine Learning in Medicine

Background:

• Sepsis-induced hypotension management typically involves intravenous fluids and vasopressors.
• Patient characteristics influencing liberal versus restrictive fluid strategies require further elucidation.
• Machine learning can uncover predictors and patient phenotypes for fluid management in sepsis.

Purpose of the Study:

• To identify key predictors differentiating restrictive from liberal fluid strategies in sepsis-induced hypotension.
• To determine the likelihood of receiving each fluid strategy within distinct patient phenotypes.
• To characterize patient phenotypes associated with sepsis-induced hypotension.

Main Methods:

• Retrospective observational study of 15,292 patients with sepsis-induced hypotension (2014-2021).
• Supervised machine learning (XGBoost) to predict fluid strategies; unsupervised learning to identify patient phenotypes.
• Subset analyses included patients with pneumonia, congestive heart failure (CHF), or chronic kidney disease (CKD).

Main Results:

• XGBoost model achieved an AUC of 0.84 for predicting fluid strategies.
• Worse oxygenation predicted restrictive fluids; higher pulse and BUN predicted liberal fluids.
• Congestive heart failure, chronic kidney disease, and pneumonia predicted restrictive fluid strategy. Three phenotypes identified: mild organ injury, severe hypoxemia, renal dysfunction.

Conclusions:

• Key predictors for restrictive versus liberal fluid strategies in sepsis-induced hypotension were identified.
• Distinct patient phenotypes associated with sepsis-induced hypotension were characterized.
• Findings can inform personalized fluid management strategies for sepsis patients.