Predicting Sepsis Onset in ICUs: A Graph Neural Network Framework with Uncertainty Estimation

Sepsis remains a leading cause of morbidity and mortality in intensive care units (ICUs), with early detection being a critical factor in improving patient outcomes. This paper presents a novel approach to sepsis prediction using Graph Neural Networks (GNNs) combined with Gaussian process-based uncertainty estimation to enhance both predictive accuracy and confidence calibration. By modeling patient data as temporal graphs, where nodes represent clinical events such as lab results and vital signs, and edges capture their temporal and relational dependencies, our method effectively learns complex patterns indicative of sepsis onset. The integration of a spectral-normalized neural Gaussian process into a heterogeneous graph transformer architecture provides robust, uncertainty-aware predictions, ensuring that the model captures the decision boundary between septic and non-septic patients without overconfidence in uncertain cases. Evaluating our method across multiple datasets, including those with significant distribution shifts in body mass index and age, we demonstrate both superior predictive performance and resilience to population variability, highlighting its capacity to extract meaningful clinical patterns while preserving calibrated uncertainty. These findings establish uncertainty-aware GNNs as a powerful tool for real-time, interpretable sepsis prediction, paving the way for more trustworthy AI-driven decision support in critical care.