A comprehensive scoping review examining artificial intelligence and genomic technologies for bacterial antimicrobial resistance surveillance has revealed that while predictive algorithms report high overall accuracy, their real-world reliability remains constrained by severe geographic imbalances and a lack of rigorous external validation. Published in Discover Public Health, the systematic evaluation identified and synthesized findings from 27 peer-reviewed investigations published between January 2020 and March 2026, establishing that published predictive performance measures across diverse bacterial pathogens ranged between 78% and 98% in terms of accuracy or area under the receiver operating characteristic curve. Crucially, the review documented that only approximately 30% of examined studies performed external validation using independent or cross-regional cohorts, and models evaluated on independent datasets frequently suffered substantial performance declines, underscoring that current computational architectures frequently capture dataset-specific artifacts rather than biologically robust and generalizable resistance determinants.
The critical urgency surrounding antimicrobial resistance prediction stems from an escalating global health crisis that disproportionately impacts low- and middle-income countries, where traditional diagnostic systems face substantial resource constraints. Antimicrobial resistance was directly responsible for an estimated 1.27 million human deaths and contributed to an additional 4.95 million fatalities worldwide in 2019 alone, with the heaviest clinical burdens concentrated in resource-constrained environments characterized by high infectious disease prevalence. Conventional phenotypic antimicrobial susceptibility testing is essential but slow and may delay detection of emerging or complex resistance mechanisms, limiting timely clinical and public health responses. Whole-genome sequencing and pan-genomic approaches enable high-resolution identification of resistance determinants, tracking of pathogen evolution, and detection of transmission clusters. Coupling these sequencing capabilities with machine learning and artificial intelligence algorithms provides an automated computational mechanism to rapidly infer phenotypic resistance directly from raw genomic sequence data, theoretically circumventing phenotypic incubation delays and informing clinical decision-making and public health surveillance.
To rigorously delineate the current state of computational resistance prediction in resource-limited environments, the authors structured their scoping review around established international methodological benchmarks, applying the Preferred Reporting Items for Systematic Reviews and Meta-Analyses extension for Scoping Reviews guidelines alongside the Joanna Briggs Institute conceptual framework. The research protocol adhered to an a priori classification framework specifically formulated to categorize studies according to their relevance to low- and middle-income countries as designated by the World Bank country income classification system, alongside predefined multi-reservoir One Health surveillance criteria. Systematic literature searches across major biomedical indexing platforms, including PubMed, Scopus, Web of Science, EMBASE, Google Scholar, and preprint repositories bioRxiv and medRxiv, initially retrieved 1,150 candidate records. Following duplicate removal and rigorous two-stage independent screening against structured eligibility criteria requiring original bacterial research utilizing whole-genome or pan-genomic sequencing, 74 full-text manuscripts were assessed in depth. Ultimately, 27 original investigations met all formal inclusion criteria, of which 22 studies were classified by the reviewers as directly or methodologically relevant to low- and middle-income country contexts.
A granular examination of the 22 studies identified as relevant to low- and middle-income territories exposed deep structural imbalances in data provenance, study leadership, and investigative infrastructure. The review categorized the relevant literature into four distinct operational tiers: six studies were classified as direct or specific low- and middle-income country studies derived exclusively from local datasets, three were designated as inclusive investigations incorporating multinational cohorts, two were characterized as mixed datasets combining isolates from both high-income and developing nations, and eleven were categorized as transferable studies conducted entirely within high-income settings but utilizing methodologies theoretically applicable to resource-constrained environments. Geographically, empirical evidence originating from low- and middle-income territories was restricted to a small cluster of nations, including Pakistan, Vietnam, Ghana, Uganda, Tanzania, Nigeria, and several multi-country African surveillance cohorts, alongside transferable high-income investigations conducted predominantly in China and the United States. Consequently, the authors emphasized that the existing literature remains overwhelmingly skewed toward models designed and calibrated in wealthy regions, with resource-constrained settings frequently serving merely as peripheral sources of biological validation data rather than active environments for end-to-end model development.
Across the investigated bacterial spectrum, the most frequently studied organisms in the synthesized research were Mycobacterium tuberculosis and Escherichia coli. Mycobacterium tuberculosis stood as the single most frequently investigated pathogen, serving as the focal organism in seven distinct studies, followed closely by Escherichia coli, which was evaluated in six dedicated research publications. In addition, Pseudomonas aeruginosa was assessed across three investigations, while eleven studies analyzed other Gram-negative organisms.
The review conducted a detailed technical evaluation of the biological feature representations utilized to convert raw nucleotide sequences into numeric matrices suitable for algorithmic processing, identifying four dominant encoding paradigms across the literature. A central synthesis insight emerging from this comparative analysis is that genomic feature engineering exerts a far more decisive influence over final predictive model performance than the selection of complex machine learning architectures. Furthermore, multi-reservoir datasets demonstrated consistent, shared resistance-associated genomic signatures spanning human, veterinary, and environmental compartments, confirming the fundamental biological plausibility and cross-sector transferability of machine-learning-based resistance predictions.
In terms of computational model architectures, the analyzed studies evaluated a diverse spectrum of analytical frameworks ranging from classical statistical baselines to sophisticated deep neural networks, each exhibiting distinct performance characteristics. In contrast, tree-based ensemble models, particularly Random Forest and XGBoost, demonstrated the most consistent performance across heterogeneous and LMIC-relevant datasets, particularly in multi-reservoir contexts. More complex neural network architectures, encompassing convolutional neural networks and deep multilayer artificial neural networks, achieved upper-tier predictive metrics ranging from 86% to 94% on specialized k-mer and pan-genome representations. However, neural networks were less frequently applied in LMIC-relevant settings, and their advantage was inconsistent, with no consistent performance advantage of one model class demonstrated.
To overcome the notorious black-box limitations of advanced machine learning models and facilitate regulatory acceptance, a growing proportion of studies incorporated explainable artificial intelligence frameworks into their computational pipelines. Explainable AI approaches, particularly SHAP, were increasingly used to enhance interpretability by identifying key resistance-associated genomic determinants across model classes. However, explainable AI remains largely adjunctive rather than embedded within modelling pipelines, with limited independent biological or clinical validation.
A critical contribution of the scoping review was its systematic assessment of the One Health paradigm, which conceptualizes human health, animal populations, and shared environmental ecosystems as an interconnected biological continuum. Despite widespread conceptual consensus regarding the multi-reservoir dynamics of antimicrobial resistance dissemination, the empirical integration of One Health principles into machine learning research was found to be strikingly underdeveloped. Only nine of the total 27 studies, and merely six of the 22 low- and middle-income country-relevant papers (representing 27.3%), met the predefined One Health criterion of incorporating biological isolates from at least two distinct environmental, animal, or human reservoirs. Furthermore, most studies incorporated data from multiple reservoirs as pooled inputs rather than explicitly modelling interactions between human, animal, and environmental compartments, indicating that computational implementation of One Health principles remains limited. The review concluded that the currently available literature fails to demonstrate that multi-reservoir data integration provides a quantifiable boost to predictive accuracy, primarily because substantial heterogeneity in study populations and sampling strategies precludes direct quantitative comparisons.
Translating genomic artificial intelligence from experimental proof-of-concept models into functional, prospective public health surveillance networks requires overcoming substantial computational, infrastructural, and methodological barriers. The scoping review identified that the existing evidence base is overwhelmingly retrospective, with an almost total absence of prospective clinical implementation trials or cost-effectiveness studies conducted within resource-constrained operational settings. To bridge this implementation gap, the authors advocated for a fundamental transition toward data-centric systems design, prioritizing the generation of geographically expansive and representative whole-genome sequencing repositories across low- and middle-income regions. Furthermore, the review called for mandatory external validation across independent, cross-country datasets as a standard benchmark for publication, alongside standardized evaluation reporting encompassing sensitivity, specificity, precision, recall, F1-scores, and calibration metrics rather than isolated accuracy figures. Finally, decentralized privacy-preserving methodologies, such as federated learning frameworks, were highlighted as promising pathways to facilitate cross-border collaborative model training while safeguarding national biological data sovereignty, thereby paving the way for truly equitable global genomic surveillance.
Subject of Research: A One Health scoping review of genomic machine learning approaches for predicting antimicrobial resistance in low- and middle-income countries
Article Title: A One Health scoping review of genomic and machine learning approaches for predicting antimicrobial resistance in low- and middle-income countries
Article References: Kimera, Z. I., Mtebe, M., Kibwana, U., Masoud, S., Kamori, D., Mbugi, E., & Matee, M. I. (2026). A One Health scoping review of genomic and machine learning approaches for predicting antimicrobial resistance in low- and middle-income countries. Discover Public Health, 23(1), Article 1679. https://doi.org/10.1186/s12982-026-02998-5
Image Credits: AI Generated
DOI: 10.1186/s12982-026-02998-5
Keywords: antimicrobial resistance, genomic machine learning, whole-genome sequencing, scoping review, low- and middle-income countries, One Health surveillance, tree-based ensembles, neural networks, Mycobacterium tuberculosis, Escherichia coli, external validation, k-mer profiling, single nucleotide polymorphisms, explainable artificial intelligence, SHAP analysis, pan-genome matrices
News Source: Juliet Wilcox. (November 13, 2026). Genomic Machine Learning Shows High Accuracy but Limited Validation in Global Surveillance. Scienmag.



