SepsiGuard: A Triple-Layer Agentic XAI Framework for Early Sepsis Prediction and Adaptive ICU Monitoring

Sepsis is a major contributor to morbidity and mortality in the Intensive Care Unit (ICU). There is a need to predict sepsis quickly and effectively to take appropriate measures. Machine learning models have shown excellent predictive performance for various adverse health conditions in ICUs. However, the lack of transparency in these models, often termed a "black box," makes them difficult to use. Traditional clinical scoring systems, such as SOFA and qSOFA, rely on fixed thresholds and fail to adapt dynamically to evolving patient conditions. Conventional Explainable AI techniques, such as feature importance plots, are static and often difficult to understand for clinicians. This paper introduces a new technique to predict sepsis in the ICU using "Agentic XAI," which is a new technique of using Large Language Models (LLMs) as autonomous agents to iteratively develop explanations. This technique combines complex physiological signals into a narrative, which is expected to bridge the gap between probability and clinical reasoning. This paper presents the methodological architecture that is expected to be used to develop this technique using recent advancements in predictive modelling in ICUs.