Machine-learning-based risk assessment tool to rule out empirical use of ESBL-targeted therapy in endemic areas

Background: Antimicrobial stewardship focuses on identifying patients who require extended-spectrum beta-lactamase (ESBL)-targeted therapy. ‘Rule-in’ tools have been researched extensively in areas of low endemicity; however, such tools are inadequate for areas with high prevalence of ESBL-producing pathogens, as almost all patients will be selected. Aim: To develop a machine-learning-based ‘rule-out’ tool suitable for areas with high levels of resistance. Methods: Gradient-boosted decision trees were used to train and validate a risk prediction model on data from 17,913 (45% ESBL) patients with Escherichia coli and Klebsiella pneumoniae in urine cultures. The predictive power of different sets of variables was evaluated using Shapley values to evaluate the contributions of variables. Findings: The model successfully identified patients with low risk of ESBL resistance in ESBL-endemic areas (area under receiver operating characteristic curve 0.72). When used to select the 30% of patients with the lowest predicted risk, the model yielded a negative predictive value ≥0.74. A simplified model with seven input features was found to perform nearly as well as the full model. This simplified model is freely accessible as a web application. Conclusions: This study found that a risk calculator for antibiotic resistance can be a viable ‘rule-out’ strategy to reduce the use of ESBL-targeted therapy in ESBL-endemic areas. The robust performance of a version of the model with limited features makes the clinical use of such a tool feasible. This tool provides an important alternative in an era with growing rates of ESBL-producing pathogens, where some experts have called for empirical use of carbapenems as first-line therapy for all patients in areas with high prevalence of ESBL-producing pathogens.