Predicting mortality in intensive care unit patients with CAUTI using an interpretable machine learning model: a retrospective cohort study from MIMIC-IV database.

Objective: The aim of this study was to develop a reliable model for predicting mortality in patients with catheter-associated urinary tract infection (CAUTI) in intensive care unit (ICU). Methods: The MIMIC-IV database was used for model development and validation in this study. Data from the first 24 h of ICU admission were collected, and 70% of the data were used to train the model and 30% to validate the model. Four machine learning models, including XGBoost, DecisionTree (DT), Logistic Regression (LR) and Random Forest (RF), were used to construct the prediction model. The SHAP method was used to explain the best performance model. Results: A total of 545 patients with CAUTI were finally included. The mortality of ICU patients with CAUTI was 7.89% (43/545). The area under the curve (AUC) of the Logistic regression model was 0.871, which showed better prediction performance among the four models. The DecisionTree machine had limited generalization ability, with an AUC of 0.542 and relatively poor prediction accuracy. The SHAP technique revealed 13 most important predictors of CAUTI in order of importance, among which use of vasoactive drugs,shock index,APSIII score, and concomitant malignancy were identified as variables with high predictive significance. Conclusion: The interpretable prediction model used in this study can help medical staff improve their ability to predict the risk of death in patients with CAUTI in ICU.