Development and validation of an interpretable machine learning model for predicting multidrug-resistant infections in ICU patients

Abstract

Background Multidrug-resistant (MDR) infections in intensive care units (ICUs) are difficult to recognize early. This study developed and validated an interpretable machine-learning model for early MDR risk prediction in ICU patients. Methods A retrospective cohort was built from the Medical Information Mart for Intensive Care IV (MIMIC-IV, n=48,501) for model development with internal validation (7:3 split) and from MIMIC-III (n=38,583) for temporal validation. Candidate predictors at ICU admission included demographics, vital signs, laboratory indices, comorbidities, therapeutic interventions, and severity scores. Missing values were imputed using missForest. Highly correlated predictors were removed using correlation screening. Feature selection used the intersection of Least Absolute Shrinkage and Selection Operator regression and Boruta. Data imbalance was handled by Random Undersampling. Eight models were constructed including Logistic Regression, K-Nearest Neighbors, Decision Tree, Random Forest (RF), Extreme Gradient Boosting, Light Gradient Boosting Machine, Support Vector Machine, and Neural Network. Hyperparameters were optimized using grid search and cross-validation. Model performance, calibration, and clinical utility were evaluated. Model interpretability used SHapley Additive exPlanations (SHAP). A web calculator was implemented using Shiny in R. Results RF achieved the best overall performance in internal validation (AUC 0.784, 95% CI 0.769-0.798) and remained stable in temporal validation (AUC 0.778, 95% CI 0.771-0.786). The final model included 19 predictors. SHAP identified enteral nutrition, Acute Physiology Score III, stage of acute kidney injury, blood urea nitrogen, opioids use, hemoglobin, and platelet as the leading contributors. A web-based calculator was created to output individualized risk with SHAP-based explanations. Conclusions We developed and validated an interpretable RF model for early prediction of MDR infections in ICU patients and translated it into a web-based calculator with SHAP explanations, enabling rapid bedside risk stratification to support timely antimicrobial stewardship and resource allocation.