naviDCN: Navigator-Guided Multi-Modal Deep Clustering for Sepsis Phenotyping in Early ICU Admission

Area of Science:

• Computational biology
• Medical informatics
• Artificial intelligence in medicine

Background:

• Sepsis is a complex, life-threatening condition with heterogeneous patient responses.
• Current unsupervised clustering methods for sepsis phenotypes lack clinical interpretability.
• Integrating clinical knowledge into clustering is crucial for meaningful sepsis subtyping.

Purpose of the Study:

• To develop a novel deep learning framework, naviDCN, for identifying clinically relevant sepsis phenotypes.
• To enhance the interpretability of sepsis clustering by incorporating clinical knowledge.
• To discover distinct sepsis phenotypes with unique clinical trajectories and outcomes.

Main Methods:

• Developed naviDCN, a framework with multi-modal encoders, a deep clustering network (DCN), and a navigator module.
• Utilized an attention mechanism on multi-modal electronic health record data for embedding.
• Employed iterative optimization of network weights and cluster centroids, guided by a navigator module incorporating clinical knowledge.

Main Results:

• Discovered four distinct sepsis phenotypes with unique clinical characteristics, SOFA score trajectories, and mortality patterns.
• naviDCN successfully differentiated between phenotypes with differing prognoses (e.g., potential improvement vs. deterioration).
• The navigator module enhanced phenotype interpretability without sacrificing clustering performance.

Conclusions:

• naviDCN provides a framework for discovering clinically meaningful sepsis phenotypes.
• The identified phenotypes offer insights into sepsis heterogeneity, aiding in targeted treatment strategies.
• This approach improves the interpretability of sepsis subtyping for clinical relevance.