Antimicrobial minimum inhibitory concentrations can be imputed from phenotypic data using a random forest approach

Antimicrobial resistance (AMR) is a public health threat requiring monitoring across multiple sectors because AMR genes and pathogens can pass between humans, animals, and the environment. Idiosyncrasies in AMR data, including missing data and changes in testing protocols, make characterizing AMR trends over time and sectors challenging. Therefore, this study applied machine learning methods to impute missing minimum inhibitory concentrations.

Models were built using cattle-associated Escherichia coli from the National Antimicrobial Resistance Monitoring System. Random forest models were designed to predict the minimum inhibitory concentration of a given E coli isolate for 10 antimicrobials. Predictors included isolate metadata and the minimum inhibitory concentrations of other antimicrobials. Model performance was evaluated on held-out test data and 2 external datasets (E coli isolated from chickens and humans).

Overall, the accuracy within 1 minimum inhibitory concentration category was over 80% for all 10 antimicrobials and over 90% for 5 antimicrobials on test data. Six of the models performed as well on both external datasets as on test data, whereas the remaining 4 had similar accuracy on the human dataset but lower on the chicken data.

These results indicate that the models can predict minimum inhibitory concentration values at a level of accuracy that would be helpful for imputation in resistance datasets.

The imputation of missing minimum inhibitory concentrations would allow for better evaluation of AMR trends over time, helping inform stewardship policies. These models may also help streamline surveillance and clinical susceptibility testing because they suggest which antimicrobials need to be laboratory-tested and which can be extrapolated by modeling.