Subphenotyping sepsis based on organ interaction trajectory using a deep temporal graph clustering model: a...

Area of Science:

• Computational biology
• Medical informatics
• Systems biology

Background:

• Sepsis is a complex syndrome characterized by multi-organ dysfunction.
• Understanding dynamic inter-organ interactions is crucial for sepsis subphenotyping and targeted therapy.
• Current approaches lack exploration of dynamic organ interaction trajectories.

Purpose of the Study:

• To quantify dynamic organ interaction trajectories for sepsis subphenotyping.
• To define distinct sepsis phenotypes based on these dynamic patterns.
• To support personalized treatment and clinical decision-making in sepsis management.

Main Methods:

• Developed a deep temporal graph clustering model to analyze organ interactions within 48 hours post-diagnosis.
• Trained and validated the model on MIMIC-III and eICU datasets.
• Compared phenotypes' characteristics, organ system coupling, and outcomes; used XGBoost for early classification and evaluated fluid management effects.

Main Results:

• Identified three distinct sepsis phenotypes (A, B, C) with significant differences in characteristics, organ coupling, and outcomes.
• Phenotype A showed lowest mortality (5.59%) with synchronous organ improvement; Phenotype C had highest mortality (38.27%) with rapid asynchrony-to-synchrony transition.
• The model achieved robust external validation and high predictive performance (AUROC 0.84) for early phenotype classification.

Conclusions:

• Characterized sepsis phenotypes using organ interaction dynamics, revealing three heterogeneous progression patterns.
• These patterns offer novel insights into sepsis pathophysiology, aiding clinical trial design and resource allocation.
• Findings underscore the need for phenotype-specific treatment strategies, including fluid management.