Gram-negative bloodstream infection (GN-BSI) can induce fatal septic shock. Moreover, the increasingly severe problem of antimicrobial resistance has led to high clinical mortality, particularly among intensive care unit (ICU) patients. The early identification of pathogens and timely antibiotic therapy are critical for patient outcomes. However, conventional diagnostic methods like blood culture are time-consuming and can delay treatment. Furthermore, the implementation of molecular detection techniques in routine laboratories is often hindered by high costs and technical complexity. Machine learning (ML) offers a promising alternative for the early prediction of GN-BSI. This study aims to develop an early prediction model for GN-BSI by integrating clinical and laboratory parameters from ICU patients using machine learning algorithms, thereby assisting in the early diagnosis and treatment of GN-BSI. This retrospective study utilized data from ICU patients admitted to the West District of the First Affiliated Hospital of Anhui Medical University from January and July 2025. Following data preprocessing and multiple imputation of missing values, the dataset was randomly divided into training and validation sets in a 7:3 ratio. Feature selection was performed using Lasso regression and multivariate logistic regression. Seven ML models were developed and evaluated using metrics including the area under the receiver operating characteristic curve (AUC), accuracy, sensitivity, specificity, F1 score, positive predictive value (PPV), and negative predictive value (NPV). Model interpretability was further assessed using Shapley Additive Explanations (SHAP) analysis. This study ultimately included 405 ICU patients. Following further feature selection, four variables were identified, including deep vein catheterization, continuous renal replacement therapy (CRRT), procalcitonin, and C-reactive protein (CRP). Early prediction models for GN-BSI in ICU patients were constructed using seven machine learning algorithms. Among these models, the XGBoost model demonstrated the best performance, with an AUC value of 0.898, accuracy of 88.43%, F1-score of 0.783, PPV of 85.00%, and NPV of 89.10%. SHAP bar and beeswarm plots illustrated the contribution of the four variables to the outcome. The SHAP dependency plot and force analysis provided model interpretability at the factor level and individual level, respectively. We developed, evaluated, and interpreted a machine learning model for predicting GN-BSI in ICU patients, which may facilitate timely intervention and treatment. The XGBoost model shows potential for clinical application following internal validation and further refinement.
Zhou et al. (Sat,) studied this question.