Interpretable prediction of hospital mortality in bleeding critically ill patients based on machine learning and SHAP.

Abstract

Background: Hemorrhage is a prevalent and critical condition in the intensive care unit (ICU), characterized by high incidence, elevated mortality rates, and substantial therapeutic challenges. Accurate prediction of mortality in patients with hemorrhage is essential for developing personalized prevention and treatment strategies. Nevertheless, the implementation of effective predictive models in clinical practice remains limited, primarily due to the lack of robust and interpretable tools.

Objective: This study aimed to develop an interpretable model for predicting mortality risk in critically ill patients with hemorrhage admitted to ICUs. The SHapley Additive exPlanations (SHAP) method was applied to interpret the eXtreme Gradient Boosting (XGBoost)model, identifying key prognostic factors in this population.

Methods: In this retrospective cohort study, we derived data from the eICU Collaborative Research Database (eICU-CRD) to develop and evaluate a predictive model. ​Clinical data from the first 24 h of ICU admission were extracted, and the dataset was randomly split into training (80%) and validation (20%) sets. Model performance was compared​ to four other machine learning algorithms using the area under the curve (AUC). ​SHAP was utilized to interpret the XGBoost model. External validation was subsequently performed using data from the ​Chinese REFRAIN cohort, which focuses on hemorrhage and coagulopathy in critically ill patients.​​.

Trial registration: The study protocol was retrospectively registered in the Chinese Clinical Trial Registry (ChiCTR) on December 17, 2024 (Registration number ChiCTR2400094140).

Results: A total of 10,306 eligible patients with hemorrhage were included. The observed in-hospital mortality rate was 11.5%.Among the five models compared, XGBoost demonstrated the highest predictive performance (AUC = 0.81), whereas logistic regression (LR) showed the lowest generalizability(AUC = 0.726). Decision curve analysis revealed that the XGBoost model provided a greater net benefit than other models at threshold probabilities of 10-30%. SHAP analysis identified the top 15 predictors of mortality, with bilirubin level ranked as the most influential variable.​​ External validation using the REFRAIN cohort confirmed the robustness of model(AUC = 0.776).

Conclusions: The interpretable predictive model improves mortality risk stratification in ICU patients with hemorrhage, supporting clinicians in optimizing treatment plans and resource allocation. Enhanced model transparency through SHAP explanations may facilitate clinical adoption by improving trust in model reliability.