An interpretable machine-learning model for post-admission reassessment of bloodstream infection risk in ICU patients with pneumonia.

The aim of this study is to develop and validate a machine learning-based predictive model to assess the risk of acquired bloodstream infection (BSI) in ICU pneumonia patients. Data were obtained from the MIMIC-IV database and Dazhou Central Hospital. The MIMIC-IV cohort was randomly divided into a training set and an internal test set, and the Dazhou cohort served as an external validation set. Candidate predictors included demographic variables, comorbidities, severity scores, microbiological results, and ICU-course variables. Boruta feature selection followed by univariable and multivariable logistic regression was used to identify predictors for model development. To address class imbalance, SMOTE was applied during model training, and model discrimination was evaluated by the area under the receiver operating characteristic curve (AUC). SHAP was used to interpret the final model. Because several leading predictors in the primary model were only available after ICU admission, we additionally performed an admission-only sensitivity analysis after excluding post-admission variables such as ICU length of stay and sputum culture results. The final post-admission model incorporated six predictors: ICU length of stay, sputum culture for Gram-negative bacteria, sputum culture for Gram-positive cocci, APSIII score, SOFA score, and liver disease. Among the evaluated algorithms, the SMOTE-GBM model showed the best discrimination, with an AUC of 0.753 (95% CI 0.719-0.788) in the internal test set and 0.703 (95% CI 0.576-0.830) in the external validation set. SHAP analysis showed that ICU length of stay and sputum culture results were the major contributors to the post-admission model predictions. In the admission-only sensitivity analysis, model discrimination decreased, with the best AUC reaching 0.751 (95% CI 0.716-0.785) in the internal test set and 0.632 (95% CI 0.493-0.770) in the external validation set, indicating that early prediction using admission-available variables alone remained challenging. An interpretable post-admission machine-learning model showed moderate discrimination for reassessing BSI risk in ICU patients with pneumonia; however, limited external precision and the weaker performance of the admission-only sensitivity analysis indicate that further refinement is required before clinical implementation, especially for early decision-making.