Background. Early identification of sepsis in patients with prolonged ICU stay requires predictive approaches considering dynamic clinical changes. Objective. To identify the risk factors and compare two data alignment strategies (left-aligned and right-aligned) for developing sepsis prediction models in critically ill patients, predominantly those in prolonged or chronic critical illness. Material and methods. A single-center retrospective study was conducted using data from the Russian Intensive Care Dataset (RICD v2.0) between December 2017 and September 2024. Two data alignment approaches were applied for model development: left-aligned (based on ICU admission) and right-aligned (based on time of sepsis onset). Predictive models were constructed using logistic regression and the XGBoost machine learning algorithm. Sepsis was identified according to Sepsis-3 criteria. Model performance was evaluated using the area under the receiver operating characteristic curve (AUROC). Results. The study included 575 patients (336 ones with sepsis; median age 64 years; median ICU stay 42 days; in-hospital mortality 11.3%). In case of left-aligned approach, significant risk factors included age ≥67 y.o., community-acquired pneumonia, elevated lactate and creatinine, ischemic stroke and chronic comorbidities. Predictive efficacy of logistic regression model was trivial (AUROC 0.661). The right-aligned approach allowed detection of significant changes in vital signs up to 6 hours before sepsis onset. The highest predictive performance was achieved using the XGBoost algorithm (AUROC 0.734). Conclusion. This study first identified specific risk factors for sepsis in patients with prolonged ICU stay. The right-aligned approach combined with XGBoost demonstrated superior predictive accuracy. These findings warrant further assessment in studies with internal and external validation.
Yadgarov et al. (Thu,) studied this question.