Machine learning models to identify patient and microbial genetic factors associated with carbapenem-resistant Klebsiella pneumoniae infection

Abstract Carbapenem-resistant Klebsiella pneumoniae (CRKP) is a critical-priority antibiotic resistance threat that has emerged over the past several decades, spread across the globe, and accumulated resistance to last-line antibiotic agents. While CRKP infections are associated with high mortality, only a small subset of patients acquiring CRKP colonization will develop clinical infection. Here, we sought to determine the relative importance of patient characteristics and CRKP genetic background in determining patient risk of infection. Machine learning models classifying colonization vs. infection were built using whole-genome sequences and clinical metadata from a comprehensive set of 331 CRKP isolates collected across 21 long-term acute care hospitals over the course of a year. Model performance was evaluated based on area under the receiver operating characteristics curve (AUROC) on held-out test data. We found that patient and genomic features were predictive of clinical CRKP infection to similar extents (AUROC IQRs: patient=0.59-0.68, genomic=0.55-0.61, combined=0.62-0.68). Patient predictors of infection included the presence of indwelling devices, kidney disease and length of stay. Genomic predictors of infection included presence of the ICEKp10 mobile genetic element carrying the yersiniabactin iron acquisition system, and disruption of an O-antigen biosynthetic gene in a sub-lineage of the epidemic ST258 clone. Altered O-antigen biosynthesis increased association with the respiratory tract, and subsequent ICEKp10 acquisition was associated with increased virulence. These results highlight the potential of integrated models including both patient and microbial features to provide a more holistic understanding of patient clinical trajectories. ImportanceMultidrug resistant organisms, such as carbapenem-resistant Klebsiella pneumoniae (CRKP), colonize alarmingly large fractions of patients in endemic regions, but only a subset of patients develop life-threatening infections. While patient characteristics influence risk for infection, the relative contribution of microbial genetic background to patient risk remains unclear. We used machine learning to determine whether patient and/or microbial characteristics can discriminate between CRKP colonization vs. infection across multiple healthcare facilities and found that both patient and microbial factors were predictive. Examination of informative microbial genetic features revealed features associated with respiratory colonization and higher rates of infection. The methods and findings presented here provide a foundation for future epidemiological, clinical, and biological studies to better understand bacterial infections and clinical outcomes.