Abstract:

BACKGROUND Sepsis is one of the main causes of mortality in intensive care units (ICUs). Early prediction is critical for reducing injury. As approximately 36% of sepsis occur within 24 h after emergency department (ED) admission in Medical Information Mart for Intensive Care (MIMIC-IV), a prediction system for the ED triage stage would be helpful. Previous methods such as the quick Sequential Organ Failure Assessment (qSOFA) are more suitable for screening than for prediction in the ED, and we aimed to find a light-weight, convenient prediction method through machine learning.

METHODS We accessed the MIMIC-IV for sepsis patient data in the EDs. Our dataset comprised demographic information, vital signs, and synthetic features. Extreme Gradient Boosting (XGBoost) was used to predict the risk of developing sepsis within 24 h after ED admission. Additionally, SHapley Additive exPlanations (SHAP) was employed to provide a comprehensive interpretation of the model's results. Ten percent of the patients were randomly selected as the testing set, while the remaining patients were used for training with 10-fold cross-validation.

RESULTS For 10-fold cross-validation on 14,957 samples, we reached an accuracy of 84.1%±0.3% and an area under the receiver operating characteristic (ROC) curve of 0.92±0.02. The model achieved similar performance on the testing set of 1,662 patients. SHAP values showed that the five most important features were acuity, arrival transportation, age, shock index, and respiratory rate.

CONCLUSION Machine learning models such as XGBoost may be used for sepsis prediction using only a small amount of data conveniently collected in the ED triage stage. This may help reduce workload in the ED and warn medical workers against the risk of sepsis in advance.

Key words: Sepsis, Machine learning, Emergency department, Triage, Informatics