A One Health scoping review of genomic and machine learning approaches for predicting antimicrobial resistance in low- and middle-income countries

Antimicrobial resistance (AMR) is a major global health threat, disproportionately affecting low- and middle-income countries (LMICs), where limited laboratory infrastructure constrains phenotypic antimicrobial susceptibility testing. Whole-genome sequencing (WGS)-enabled machine learning (ML) and artificial intelligence (AI) offer rapid approaches for AMR prediction, but their applicability across heterogeneous populations and One Health reservoirs remains uncertain. To map genomic ML/AI approaches for bacterial AMR prediction, focusing on genomic features, model architectures, validation, performance, and LMIC and One Health applicability. A scoping review of 27 studies published between 2020 and 2026 was conducted, including 22 with LMIC-relevant datasets. Evidence was narratively synthesized according to genomic features, modelling strategies, validation approaches, geographic coverage, and ecological context. Studies were categorized using predefined LMIC relevance and One Health criteria. Most studies used WGS-derived features, particularly k-mers and resistance-gene profiles. Tree-based ensembles and regularized regression models predominated, while neural networks were less frequently applied but generally reported high performance. Reported performance ranged from 78%–98% (accuracy or AUC; not directly comparable). External validation using independent LMIC-derived or cross-regional datasets was reported in approximately 30% of studies, whereas cross-country LMIC validation was limited. Six of 22 LMIC-relevant studies (27.3%) incorporated multi-reservoir One Health datasets. Performance varied with dataset composition, genomic feature representation, and model architecture. Explainable AI remained limited. Genomic ML/AI shows strong potential for AMR prediction but remains constrained by limited LMIC representation, external validation, and One Health integration. Expanding geographically diverse WGS datasets, strengthening independent cross-context and cross-country validation, harmonizing evaluation frameworks, and integrating explainable AI are priorities for equitable LMIC deployment.