An interpretable machine learning model for predicting in-hospital mortality in ICU patients with ventilator-associated pneumonia

Abstract

Background Ventilator-associated pneumonia (VAP) is a common nosocomial infection in ICU, significantly associated with poor outcomes. However, there is currently a lack of reliable and interpretable tools for assessing the risk of in-hospital mortality in VAP patients. This study aims to develop an interpretable machine learning (ML) prediction model to enhance the assessment of in-hospital mortality risk in VAP patients. Methods This study extracted VAP patient data from versions 2.2 and 3.1 of the MIMIC-IV database, using version 2.2 for model training and validation, and version 3.1 for external testing. Feature selection was conducted using the Boruta algorithm, and 14 ML models were constructed. The optimal model was identified based on the area under the receiver operating characteristic curve (AUROC), accuracy, sensitivity, and specificity across both validation and test cohorts. SHapley Additive exPlanations (SHAP) analysis was applied for global and local interpretability. Results A total of 1,894 VAP patients were included, with 12 features ultimately selected for model construction: 24-hour urine output, blood urea nitrogen, age, diastolic blood pressure, platelet count, anion gap, body temperature, bicarbonate level, sodium level, body mass index, and whether combined with congestive heart failure and cerebrovascular disease. The random forest (RF) model showed the best performance, achieving an AUC of 0.780 in internal validation and 0.724 in external testing, outperforming other ML models and common clinical scoring systems. Conclusion The RF model demonstrated robust and reliable performance in predicting in-hospital mortality risk for VAP patients. The developed online tool can assist clinicians in efficiently assessing VAP in-hospital mortality risk, supporting clinical decision-making.