Machine learning-based prediction of bloodstream infection in rectal carbapenem-resistant Enterobacterales carriers: a multicenter retrospective cohort study.

BACKGROUND: Carbapenem-resistant Enterobacterales (CRE) carriers are at risk of developing subsequent CRE bloodstream infection (BSI). This study aimed to identify features contributing to BSI development in rectal CRE carriers and interpret their interactions using machine learning (ML) and explainable artificial intelligence (XAI), while implementing individualized BSI risk prediction. METHODS: We included patients with first-time positive rectal CRE surveillance cultures during hospitalization at three hospitals in Korea from January 2014 to December 2023. A total of 173 features-laboratory data including the carbapenem minimum inhibitory concentrations (MIC) of rectal CRE isolates and complete blood cell count, demographics, vital signs, comorbidities, medications, transfusions, and procedures -were collected for the 30-day and 14-day follow-up periods, of which 151 remained after removing redundant features. The entire dataset was randomly split into training (80%) and test (20%) sets, and XGBoost ML models were developed. Model performance was evaluated using the area under the receiver operator characteristic curve (AUROC) with 1, 000 bootstrap iterations. SHAP (SHapley Additive exPlanations) was employed as the XAI method. RESULTS: Among 639 included CRE carriers, 72 (11.3%) and 46 (7.2%) developed BSI within 30 and 14 days, respectively. The 30-day and 14-day models demonstrated excellent performance with AUROCs of 0.917 (95% CI: 0.889-0.943) and 0.854 (95% CI: 0.821-0.885), respectively. In the 30-day model, the most contributory feature was minimum inhibitory concentration (MIC) of imipenem or meropenem against rectal CRE, with BSI risk increasing when MIC ≥8 μg/mL. In the 14-day model, increased neutrophil count (≥7.50×10³/μL) was the most contributory feature. Several features showed nonlinear relationships with BSI risk and contrasting contributions depending on feature interactions. Individual 30-day and 14-day prediction models required at least 12 top-contributing features, achieving AUROCs of 0.873 (95% CI: 0.728-0.969) and 0.864 (95% CI: 0.727-0.959), respectively. CONCLUSIONS: The influence of contributing features varied by time periods and interactions. This study comprehensively interpreted complex factors using ML and XAI, achieving excellent predictive performance. Individualized BSI risk stratification can facilitate determining appropriate timing for preemptive therapy and inform antimicrobial stewardship decisions.