Identifying Subphenotypes for Sepsis with Acute Kidney Injury via Multimodal Graph State Space Models
Keywords: Subphenotype, Sepsis with Acute Kidney Injury, Graph Representation Learning, State Space Models
TL;DR: We propose MGSSM-AKI, a framework using multimodal fusion and graph selective state space model to identify subphenotypes for Sepsis with Acute Kidney Injury, outperforming existing models on the MIMIC-IV dataset.
Abstract: Sepsis with acute kidney injury (SAKI) is a heterogeneous clinical syndrome and a leading cause of mortality in intensive care units (ICUs). Identifying subphenotypes of SAKI can improve treatment precision, enabling more targeted clinical interventions. Recently, the analysis of sepsis subphenotypes using electronic health records (EHRs) has gained interest among healthcare researchers. However, current methods typically rely on unimodal features, overlooking intrinsic correlations among patients and struggling with the sparse and high-dimensional nature of EHR data. In this paper, we propose **MGSSM-SAKI**, a novel **M**ultimodal **G**raph **S**elective **S**tate **S**pace **M**odel for identifying subphenotypes of **SAKI**. First, we develop a multimodal fusion module that integrates demographic information, laboratory results, vital signs, and diagnostic data. Next, we introduce an adaptive latent graph inference module that captures latent graph structures and co-optimizes them with the identification model to reveal intrinsic patient connections. Inspired by the recent success of state space models (SSMs), such as Mamba, we incorporate a graph learning model that combines graph neural networks with selective SSMs. Finally, we design a spectral modularity maximization objective function with regularization terms to achieve differentiable patient subphenotype identification. Experiments conducted on the MIMIC-IV dataset demonstrate that our model outperforms baseline models, exhibiting strong performance and interpretability.
Submission Number: 24
Loading