Moving from predicting hospital deaths by antibiotic-resistant bloodstream bacteremia toward actionable risk reduction using machine learning on electronic health records.

ABSTRACT

Drug-resistant bacterial infections are a global health concern with high mortality and limited treatment options. Several clinical risk-severity scores are available, e.g. qPitt, but their predictive performance is moderate. Here, we leveraged machine learning and electronic health records (EHRs) to improve prediction of mortality due to bloodstream infection with Klebsiella pneumoniae. We tested the qPitt score and new EHR variables (either expert-chosen or the full set of diagnostic codes), fitting LASSO, boosted logistic regression (BLR), support vector machines, decision trees, and random forests. The qPitt score showed moderate discriminative ability (AUROC=0.63), whilst machine learning models significantly improved its performance (best AUROC by BLR 0.80 for expert-chosen and 0.88 for full code set). Similar results were obtained in critically ill patients, and when excluding potential non-causal variables to evaluate an actionable model. In conclusion, current risk scores for bacteremia mortality can be improved and, with opportune causal modelling, considered for deployment in clinical decision-making.