Online Critical-State Detection of Sepsis Among ICU Patients using Jensen-Shannon Divergence

TL;DR

This paper introduces a real-time, interpretable method for early sepsis detection in ICU patients by measuring physiological divergence from a reference group using Jensen-Shannon divergence, validated on real-world data.

Contribution

It presents a novel non-parametric divergence-based approach for early sepsis detection that highlights contributing physiological factors in real time.

Findings

Successfully distinguishes septic from non-septic patients

Provides real-time divergence measurements

Identifies key physiological contributors to sepsis

Abstract

Sepsis is a severe medical condition caused by a dysregulated host response to infection that has a high incidence and mortality rate. Even with such a high-level occurrence rate, the detection and diagnosis of sepsis continues to pose a challenge. There is a crucial need to accurately forecast the onset of sepsis promptly while also identifying the specific physiologic anomalies that contribute to this prediction in an interpretable fashion. This study proposes a novel approach to quantitatively measure the difference between patients and a reference group using non-parametric probability distribution estimates and highlight when abnormalities emerge using a Jensen-Shannon divergence-based single sample analysis approach. We show that we can quantitatively distinguish between these two groups and offer a measurement of divergence in real time while simultaneously identifying specific physiologic factors contributing to patient outcomes. We demonstrate our approach on a real-world dataset of patients admitted to Atlanta, Georgia's Grady Hospital.