Abstract
Antimicrobial resistance (AMR) is a dynamic evolutionary process in which bacteria acquire and disseminate resistance through mutation, recombination, and horizontal gene transfer. This review synthesizes machine learning (ML) and deep learning (DL) applications in AMR diagnostics, genomic and metagenomic surveillance, and antibiotic discovery, emphasizing prediction of resistance evolution rather than classification of its current state. We map genomic, spectral, imaging, metagenomic, and clinical data to predictive architectures, including tree ensembles, convolutional and recurrent networks, graph neural networks, transformer-based genomic encoders, and generative models. A 57-study systematic review of Klebsiella pneumoniae found that 74.3% of AUROC-primary studies reported AUROC ≥0.90, yet validation remained predominantly retrospective: internal cross-validation was used in 96.5% of studies, compared with <30% external validation, 15.8% temporal validation, and only one prospective evaluation. High risk of bias affected 78.9% of studies, while approximately half relied on automated susceptibility testing rather than reference-grade broth microdilution. Fewer than 10% of AI-microbiology studies originated from low- and middle-income countries. These findings expose the gap between cross-sectional classification and evolution-aware prediction. We propose a six-component roadmap integrating longitudinal genomic surveillance, phylogeny-aware representation, temporal and relational architectures, forecast-oriented validation, federate learning, and explainable AI with prospective laboratory validation to anticipate resistance and accelerate countermeasure discovery.