Abstract

Purpose: Early mortality probability prediction in intensive care units using clinical decision support approaches is important for timely patient risk stratification, improving prognosis, and optimizing hospital resources. We developed an all-cause mortality risk prediction model based on the Machine Learning (ML) algorithm eXtreme Gradient Boosting (XGBoost). Additionally, we included an explainability layer based on Shapley Additive exPlanations (SHAP) that allow the user to understand the model's predictions and individual feature contribution. Material and Methods:We trained and tested our model with a curated database assembled after performing a retrospective study of ICU admissions from the public Medical Information Mart for Intensive Care (MIMIC-III, version 1.4) and eICU Collaborative Research Database (eICU-CRD, version 2.0) databases. Results and Conclusions: We showed that our XGBoost-based ML model accurately predicts mortality better than the most commonly used scores in ICUs and other ML-based models. Additionally, we showed that our model could predict all-cause mortality using just 20 or eight clinical variables measured within the first 24 hours of ICU admission. Finally, we demonstrated that race/ethnicity is an important factor influencing the performance of our ML-based models and that cohorts used to do so need to be diverse and representative to increase their global applicability.