Development and Validation of a Machine learning Model for the Prediction of Short-Term Mortality among Patients with Sepsis: A Retrospective Cohort Study

Abstract Background: Sepsis, a severe condition arising from the body's response to infection, leading to organ damage, is a major contributor to mortality in intensive care units. This study assesses and compares the predictive performance of six machine learning algorithms in forecasting in-patient mortality among sepsis patients. Results: There were 6,455 sepsis patients (5,165 for training and 1,290 for internal validation) in Medical Information Mart for Intensive Care IV v2.0 and 411 sepsis patients in eICU Collaborative Research Database v2.0 who met the inclusion criteria. The median age of the patients was 64 years and 61 years respectively. 42 features were retained for model building. The most important features for predicting mortality were urine output, Glasgow Coma Scale score, urine output, average body temperature, age, and lactate. The random forest algorithm performed the best on both the imbalanced and balanced datasets, achieving an AUC of 78.6% and 78.5%, respectively, on the external test set. Conclusions: This study showcases the applicability of machine learning for outcome prediction in sepsis patients, employing internal and multicenter external validation datasets. The application of the SHAP method enhances model interpretability by providing insights into global explanations, individual explanations, and the interactions between features.