Abstract
Machine learning (ML) models for antimicrobial resistance (AMR) are typically trained and judged against reductionist antimicrobial susceptibility testing (AST) labels, which can fail to capture the biological nuance of host microenvironments, host-pathogen interactions, and drug bioavailability, limiting how well they predict real clinical outcomes. In this invited letter, we examine the origins and reliability of the AMR labels used to train ML models, and argue that while current ASTs constrain what ML can learn, they still offer critical mechanistic insights into AMR that should inform how future models are built and evaluated.
Type
Publication
Trends in Microbiology
