Machine-Learning Model for Prediction of Cefepime Susceptibility in Escherichia coli from Whole-Genome Sequencing Data

The declining cost of performing bacterial whole-genome sequencing (WGS) coupled with the availability of large libraries of sequence data for well-characterized isolates have enabled the application of machine-learning (ML) methods to the development of nonlinear sequence-based predictive models. We tested the ML-based model developed by Next Gen Diagnostics for prediction of cefepime phenotypic susceptibility results in Escherichia coli. A cohort of 100 isolates of E. coli recovered from urine (n = 77) and blood (n = 23) cultures were used. The cefepime MIC was determined in triplicate by reference broth microdilution and classified as susceptible (MIC of ≤2 μg/mL) or not susceptible (MIC of ≥4 μg/mL) using the 2022 Clinical and Laboratory Standards Institute breakpoints. Five isolates generated both susceptible and not susceptible MIC results, yielding categorical agreement of 95% for the reference method to itself. Categorical agreement of ML to MIC interpretations was 97%, with 2 very major (false, susceptible) and 1 major (false, not susceptible) errors. One very major error occurred for an isolate with blaCTX-M-27 (MIC mode, ≥32 μg/mL) and one for an isolate with blaTEM-34 for which the MIC cefepime mode was 4 μg/mL. One major error was for an isolate with blaCTX-M-27 but with a MIC mode of 2 μg/mL. These preliminary data demonstrated performance of ML for a clinically important antimicrobial-species pair at a caliber similar to phenotypic methods, encouraging wider development of sequence-based susceptibility prediction and its validation and use in clinical practice.