Utilization of machine learning to predict antibiotic resistant event outcomes in acute myeloid leukemia patients undergoing induction chemotherapy

Introduction Acute myeloid leukemia (AML) patients are highly susceptible to infection. Moreover, prophylactic and empirical antibiotic treatment during chemotherapy disrupts the gut microbiome, raising the risk for antibiotic-resistant (AR) opportunistic pathogens. There is limited data on risk factors for AR infections or colonization events in treated cancer patients, and no predictive models exist. This study aims to combine metagenomic and antibiotic administration data to develop a model predicting AR event outcomes. Methods Baseline stool microbiome, antibiotic administration, resistome, and clinical metadata from 95 patients were utilized to build a Random Forest model to predict AR infection and colonization events by serious AR threats. Additionally, sparse canonical correlation analysis assessed correlations between microbiome and resistome data, while Spearman correlation networks identified direct associations with AR event outcomes and secondary variables. Results AR-events were identified in 14 of the 95 included patients, with 8 developing AR infections and 9 identified as AR colonized. A Random Forest model predicted AR event outcomes (AUC = 0.73), identifying bacterial taxa and antibiotic resistance gene (ARG) classes as key variables of importance. Methanobrevibacter smithii, Clostridium leptum , and Bacteroides dorei were identified as key taxa associated with reduced risk of AR events, suggesting the potential roles of commensals in maintaining gut microbial resilience during chemotherapy. ARG classes, particularly those conferring resistance to lincosamides, macrolides, and streptogramins, were negatively associated with AR events. Conclusion These results underscore the value of integrating microbiome and resistome features to reveal potential protective mechanisms and improve risk prediction for AR outcomes in vulnerable patients.