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

Abstract

The World Health Organization considers carbapenem-resistant Enterobacteriaceae (CRE) an urgent public health threat due to global prevalence, limited treatment options, and high mortality rates. Carbapenem-resistant Klebsiella pneumoniae (CRKP), the most common CRE species, is especially prevalent in long-term acute care hospitals (LTACHs) in the United States. Among patients colonized with CRKP, only a subset develop clinical infection. We used CRKP whole-genome sequences and associated clinical metadata from unique patient samples from 20 LTACHs (n=366) to (1) identify both clinical and microbial features associated with infection compared to colonization, and (2) evaluate the contribution of microbial features in infection prediction. Modifiable clinical features associated with infection could inform clinical practice, while microbial features could help predict clinical infection. We performed L2 regularized logistic regression 100 times for each feature set. Clinical predictors of infection include having a central line or gastrostomy tube. Genomic predictors of infection include iron scavenging genes (known to be associated with invasive infections in healthy hosts) and the number of antibiotic resistance genes. Furthermore, we found that clinical and genomic features (separate or combined) have similar predictive power (AUROC IQRs: clinical=0.56-0.64, genomic=0.54-0.63, combined=0.59-0.67). This suggests that genomic features may be associated with certain clinical features. These results provide insight into potential clinical and microbial drivers of CRKP infection in LTACH patients and provide a starting point for investigating the biological basis of infection, identifying patients at high risk of infection, and devising targeted strategies to prevent infection.