Subphenotype Identification for Sepsis-Associated Acute Kidney Injury Using Graph Bidirectional Mamba Networks.

Area of Science:

• Biomedical Informatics
• Computational Biology
• Critical Care Medicine

Background:

• Sepsis-associated acute kidney injury (SA-AKI) is a complex condition with high mortality in ICUs.
• Current methods for SA-AKI subtyping using EHR data are limited by static features and data sparsity.
• There is a need for advanced models to capture intricate patient correlations for precise SA-AKI subtyping.

Purpose of the Study:

• To develop a novel Graph Bidirectional Mamba Network (GBMN) for identifying SA-AKI subphenotypes.
• To improve the precision of SA-AKI subtyping by leveraging multi-modal EHR data and latent patient graph structures.
• To enhance targeted clinical interventions for SA-AKI through accurate subphenotype identification.

Main Methods:

• Developed a multi-modal fusion module integrating demographic, laboratory, vital signs, and diagnostic data.
• Introduced an adaptive latent graph inference module to capture and optimize patient graph structures.
• Incorporated a graph learning model combining Graph Neural Networks with Mamba (a state space model).
• Designed a spectral modularity maximization objective for differentiable subphenotype identification.

Main Results:

• The proposed GBMN model demonstrated superior performance compared to baseline models on the MIMIC-IV dataset.
• The model achieved strong performance in identifying SA-AKI subphenotypes.
• The identified subphenotypes and their contributing factors showed significant interpretability.

Conclusions:

• GBMN offers a powerful and interpretable approach for SA-AKI subphenotype identification using EHR data.
• The model's ability to uncover patient connections and key factors can guide personalized treatment strategies.
• This research advances precision medicine in critical care by enabling more targeted interventions for SA-AKI.