Window-Controlled Sepsis Prediction Using a Model Selection Approach

Shiyan Su, Su Lan, Zhicheng Zhang, Anjie Zhu

Abstract: Sepsis is a major risk to patient in the Intensive Care Unit (ICU) and is associated with substantial treatment expenditure. As most cases of sepsis are acquired during the ICU stay, timely identification and intervention play a crucial role in enhancing the survival rate of septic patients and reducing the financial burden of treatment. Prior research has established that machine learning approaches surpass conventional scoring systems in predicting sepsis. However, these existing machine learning methodologies exhibit limitations when predicting sepsis with flexible window settings. Their performance is heavily reliant on the selection of prediction and feature windows, which restricts their practical applicability in clinical settings. This paper aims to overcome this challenge by introducing a model selection approach for sepsis prediction, utilizing a window-controlled strategy. Experimental results demonstrate that our proposed model outperforms existing models and exhibits enhanced stability across various prediction and feature windows.