Abstract
Sepsis remains a life-threatening condition requiring early recognition and intervention to improve patient outcomes, particularly in the high-acuity environment of the Emergency Department (ED). Traditional diagnostic criteria often lack sensitivity and specificity for early detection. This retrospective cohort study aimed to evaluate the utility of various machine learning (ML) algorithms for predicting sepsis onset in ED patients using routinely collected clinical and laboratory data. Data from 8,572 adult patients presenting to a tertiary ED between January 2018 and December 2020 were analyzed, with 583 (6.8%) developing sepsis within 72 hours of admission based on Sepsis-3 criteria. We trained and tested Logistic Regression, Random Forest, Support Vector Machine, and Gradient Boosting (XGBoost) models on a comprehensive dataset including demographics, vital signs, and initial laboratory results. XGBoost demonstrated superior performance, achieving an Area Under the Receiver Operating Characteristic Curve (AUC-ROC) of 0.89 (95% CI: 0.87-0.91) on the unseen test set, with a sensitivity of 0.82 and a specificity of 0.80 at its optimal threshold. Key predictors identified included lactate levels, C-reactive protein, white blood cell count, mean arterial pressure, and age. These findings suggest that ML models, particularly XGBoost, can effectively leverage routine ED data to provide timely and accurate predictions of sepsis onset, potentially enabling earlier clinical intervention and improving patient management.
