Abstract
Sepsis, a life-threatening medical condition characterized by the body’s extreme response to infection, remains a critical challenge in intensive care units, responsible for significant morbidity and mortality worldwide. Traditional and deep learning methods for sepsis outcome prediction have primarily treated patient data as independent observations, overlooking inherent structural and relational dependencies between patients. This project introduces Sepsis-GNN, a novel Graph Neural Network (GNN)-based framework explicitly designed to capture complex inter-patient relationships within clinical datasets. By converting tabular patient data into graph structures—where patients are nodes connected by edges representing clinical similarity, Sepsis-GNN enables richer representation and improved predictive performance. Utilizing the comprehensive MIMIC-IV dataset comprising 18,803 ICU patient records and 91 clinical features, our proposed framework demonstrates superior predictive accuracy and interpretability over established baseline models including FT-Transformer, Tab-Transformer, COMPOSER, Random Forest, and XGBoost. Experimental validation underscores Sepsis-GNN’s effectiveness in capturing structural patterns vital for accurate mortality prediction, highlighting its potential to support clinical decision-making with reliable, actionable insights.