(1444) Machine Learning Models for Predicting Mortality in Sepsis: A Systematic Review

Introduction: Sepsis is the leading cause of death among patients in the intensive care units (ICUs). Various Machine Learning (ML) models have been used to predict sepsis-associated mortality. Therefore, we conducted a systematic review to evaluate the reporting and methods utilized in ML algorithms and traditional scoring systems and assess their performance in predicting in-hospital mortality among patients with sepsis.

Methods: A comprehensive search of databases (Ovid MEDLINE and EMBASE, Ovid Cochrane Database of Systematic Reviews, Scopus, and Web of Science) was performed from inception to May 10th, 2023. We included all full text peer-reviewed articles reporting the prediction of clinical outcomes in critically ill adult patients with sepsis. Two authors screened the papers, extracted the data independently and evaluated the risk of bias using the Prediction model Risk of Bias Assessment (PROBAST) Tool.

Results: Out of 786 studies, 24 were included, involving 1,123,799 adult patients with sepsis. 17 studies (71%) were found to have a high risk of bias. The prevalence of mortality ranged between 6.5% and 51%. A diverse range of predictive ML models was utilized to predict in-hospital or in-ICU mortality in sepsis patients. Among these models, Random Forest (RF) was the most frequently mentioned, appearing in 14 studies, followed by Support Vector Machine (SVM) in 9 studies, extreme Gradient Boosting (XGBoost) in 8 studies, and Naive Bayes (NB) in 6 studies. Internal validation of ML models was conducted in 20 studies, while only five studies performed external validation. Notably, when comparing ML models with logistic regression (LR) and traditional scoring systems, the ML models consistently outperformed, highlighting their superiority in predicting mortality for sepsis patients (AUC ranges: 0.62 to 0.99 vs 0.55 to 0.87).

Conclusions: ML models generally exhibited a favorable performance in predicting sepsis associated mortality. However, due to the significant heterogeneity and risk of bias in the populations, methodologies, and predictors employed, conducting meta-analyses was not feasible. Future prospective, multicenter studies are essential to validate the effectiveness of ML algorithms in predicting sepsis- associated mortality.