An interpretable machine learning approach for predicting clinically important gastrointestinal bleeding in critically ill patients

Background

Clinically important gastrointestinal bleeding (CIGIB) is a serious complication in critically ill patients, contributing to prolonged ICU stays and increased mortality. Despite efforts to identify high-risk patients, no previous studies have employed machine learning models to predict CIGIB during ICU stay or identify key predictors in this context.

Methods

This single-center retrospective study included ICU patients aged 18 years or older admitted between 2017 and 2024. Patients with ICU stays of less than 24 h or GIB within 24 h of admission were excluded. Machine learning models, including XGBoost, Random Forest, and L1-regularized logistic regression, were trained using patient data from the first 24 h of ICU admission. Model performance was assessed using AUROC, precision, recall, and F1 scores. Shapley Additive Explanations (SHAP) were employed to evaluate key predictors.

Results

A total of 7357 ICU patients were included, of whom 171 (2.3%) experienced CIGIB. The XGBoost model demonstrated the highest predictive performance with an AUROC of 0.84. Key predictors included APACHE III scores, hematocrit levels, APTT, creatinine and respiratory rate, while invasive mechanical ventilation and stress ulcer prophylaxis within the first 24 h of ICU admission did not rank among the top 20 predictors based on SHAP values.

Conclusions

This study represents the first application of machine learning for predicting CIGIB in ICU patients, providing valuable insights into risk stratification. The model demonstrated high predictive accuracy and interpretability, highlighting its potential to guide early intervention and prophylaxis. Further multi-center studies and interventional trials are needed to validate these findings and refine clinical risk prediction strategies.