Machine Learning in the ICU: Predicting Mortality in Patients with Carbapenem-Resistant Gram-Negative Bacilli Bloodstream Infections

Background and ObjectiveMultidrug and carbapenem resistant gram-negative bacilli bloodstream infections cause high mortality in intensive care units (ICUs). Predicting mortality can improve treatment and support end-of-life decisions. This study aimed to develop a machine learning model to predict mortality in ICU patients with these infections.MethodsThis retrospective cohort study was conducted at a tertiary care medical center between 2017 and 2023. Adult ICU patients with bloodstream infections caused by multidrug and carbapenem resistant Klebsiella pneumoniae, Pseudomonas aeruginosa, and Acinetobacter baumannii were included. Demographic, clinical, and laboratory data were collected. Mann-Whitney U and Chi-square tests were used to compare the groups. Multivariable analysis with binary logistic regression was used to identify mortality risk factors. Ten machine learning classifiers were evaluated using stratified 5-fold cross-validation, and model predictions were interpreted with SHapley Additive exPlanations (SHAP).Results197 patients were included, with a 15-day mortality rate of 48%. The Light Gradient Boosting Machine (LightGBM) classifier showed the best performance, with an AUROC of 0.94, AUPRC of 0.952, accuracy of 0.868, precision of 0.906, recall of 0.822, F1 score of 0.855, Matthews Correlation Coefficient (MCC) of 0.744, and Brier score of 0.131. SHAP analysis revealed coagulopathy, rapid access to antibiotics, septic shock, SOFA score, platelet count, C-reactive protein (CRP) level, and time-related parameters as the most important predictive features.ConclusionThe LightGBM model showed promising results in predicting mortality in ICU patients. This model may support early intervention and assist in complex end-of-life decisions. This study was registered at ClinicalTrials.gov(https://clinicaltrials.gov/ct2/show/NCT06167083).