Application of machine learning to identify risk factors for bloodstream infections caused by different antimicrobial resistance phenotypes in chemotherapy patients.

PURPOSE: Bloodstream infections (BSIs) remain a major cause of morbidity and mortality in patients undergoing chemotherapy. The rising incidence of multidrug-resistant organisms (MDROs) further challenges the adequacy of empirical antimicrobial therapy. Therefore, this study aimed to identify features associated with different microbiological profiles of BSIs in chemotherapy patients using machine learning models. METHODS: We retrospectively analyzed 1435 confirmed BSI episodes among 1220 patients treated at a tertiary cancer center from 2017 to 2022. A multi-label classification approach was used to model BSIs associated with carbapenem-resistant Gram-negative bacilli (CR-GNB), methicillin-resistant staphylococci, third-generation cephalosporin-resistant Enterobacterales (3GCRE), fungal pathogens, and non-MDRO/non-fungal. Classifier chain models based on logistic regression, random forests, neural networks, and gradient boosting were trained and evaluated using cross-validation with an independent test set. Model outputs were interpreted using SHapley Additive exPlanations (SHAP) to estimate feature contributions. RESULTS: Most BSIs occurred during the first chemotherapy cycle and were due to non-MDROs or fungal agents (1193 episodes; 83.1%), CR-GNB was identified in 109 (7.6%). Model performance varied across pathogen groups, with the highest discrimination observed for CR-GNB (AUC-ROC 83.25%; 95% CI 73.91%-91.64%). Key features included hospitalization days, hematologic malignancy, prior MDRO infection, monocytopenia, and palliative chemotherapy. Chemotherapy class contributed differently across pathogen groups, with alkylating agents contributing to candidemia, antitumor antibiotics to CR-GNB, and antimetabolites to 3GCRE. Prior CR-GNB infection was the strongest predictor for CR-GNB BSI, while palliative chemotherapy correlated with non-MDRO/non-fungal BSI. CONCLUSION: Multi-label machine learning approaches may help disentangle the complex etiologies of BSIs in cancer patients by identifying associations between host factors, prior MDRO infection, and chemotherapy class within confirmed BSI episodes, informing future hypothesis generation.