Improved prediction of antimicrobial resistance in Klebsiella pneumoniae using machine learning

Abstract

Klebsiella pneumoniae is an important cause of healthcare-associated infections, with high levels of antimicrobial resistance (AMR) to critical antibiotics such as carbapenems and third-generation cephalosporins (3GCs). Accurate antimicrobial susceptibility detection is essential for guiding appropriate treatment. In this study, we evaluated the efficacy of machine learning (ML) models for predicting AMR phenotypes in K. pneumoniae particularly for antibiotics for which rule-based approaches fail. We analyzed a dataset of 5,907 K. pneumoniae genomes from public databases and a genomic surveillance project in Spanish hospitals. ML models were trained to predict AMR phenotypes using genomic features, and their performance was compared to ResFinder, which implements a conventional rule-based approach. Models were evaluated based on predictive accuracy across antibiotics. Additionally, we conducted a detailed analysis of the genomic features associated with AMR identified by ML to identify new putative AMR determinants. ResFinder exhibited low prediction accuracy for amikacin, fosfomycin, and piperacillin/tazobactam, whereas ML models significantly improved the prediction accuracy for these antibiotics. Moreover, we provide insights into why rule-based methods failed in these cases, specifically related to the genes acc(6)-Ib-cr, fosA, and blaOXA-1, respectively. Finally, we found possible genetic factors related to resistance for each antibiotic. Our findings underscore the value of ML models in AMR prediction based on genome information for K. pneumoniae, especially in challenging cases where traditional methods have low success rates. Continued evaluation and refinement of ML approaches are essential for applying these methods to enhance AMR detection in clinical and public health contexts. ImportanceTo combat antimicrobial resistance (AMR), the rapid and accurate identification of resistance phenotypes is essential for guiding appropriate therapy. In this study, we demonstrate the significant potential of machine learning (ML) to improve AMR prediction in Klebsiella pneumoniae using genomic data. Our findings reveal that gold standard rule-based methods for predicting AMR from genomic data underperform for antibiotics such as amikacin, fosfomycin, and piperacillin/tazobactam. In this study, we identified the genomic determinants that mislead resistance predictions in rule-based methods providing insights that can refine existing rule-based approaches. Moreover we used ML models that improved the prediction accuracy for these antibiotics and used these models to uncover potential new AMR-associated genes that contribute to a deeper understanding of resistance mechanisms. While these findings are specific to K. pneumoniae, the ML approach is broadly applicable to other pathogens facing similar challenges, enabling improved AMR prediction without reliance on prior knowledge.