Peer Review
This scoping review is a synthesis of the evidence based on the existing systematic and scoping reviews on mathematical modelling and predictive analytics of antimicrobial resistance (AMR) dynamics. It seeks to trace the use of mechanistic transmission models and machine-learning (ML) methods to understand the emergence, spread and prediction of AMR in pathogens, hosts and settings. A total of 10 studies were included: six systematic reviews of mechanistic or transmission models, three systematic reviews of ML-based resistance prediction and one narrative review, providing a broader conceptual context on AMR modelling. Mechanistic examinations reveal that deterministic compartmental model structures in human medical centres or community contexts prevail with a paucity of One Health consolidation, limited external validation and under-representation of WHO priority Gram-negative organisms and low and middle-income countries. ML model reviews state that patient-level resistance prediction has promising discriminative performance (summary area under the receiver operating characteristic curve (AUC) of approximately 0.78–0.82 across ML-based prediction systematic reviews) based on heterogeneous data pipelines, however, the studies have mainly been applied in a retrospective, single-centre design, and are likely to be susceptible to bias and lack prospective impact assessment. In both fields, major gaps are associated with validation, transparency, reproducibility, coverage of pathogens and settings and correlation with policy-relevant economic and decision measures. Development of hybrid mechanistic-ML frameworks and explicit One Health-oriented modelling strategies, aligned with WHO GLASS and national AMR action plans, are required to strengthen AMR policy and clinical decision-making.
Antimicrobial resistance (AMR) is one of the most pressing threats to global public health. It occurs when pathogens acquire resistance to antimicrobial agents, rendering the effective treatment of common infections increasingly compromised.
Over the past decade, a substantial body of literature has examined AMR dynamics through mechanistic modelling.
In parallel, data-driven predictive approaches have emerged. These are time-series analyses and machine learning (ML) models that have been trained on surveillance or clinical data. ML algorithms (supervised and unsupervised) have been effectively used to forecast resistance patterns and inform therapeutic decision-making.
This review-of-reviews synthesises evidence from published systematic and scoping reviews on AMR modelling and prediction (2018–2025 for formal database searches; supplemented by a preprint repository hand-search to January 2026). It aimed to: (i) characterise model types and underlying assumptions; (ii) identify common parameters and data sources; (iii) describe prediction methodologies and reported performance metrics; (iv) quantify evidence overlap using corrected covered area analysis and (v) delineate persistent research gaps and future priorities, including practical guidance for decision-makers.
The scoping review was developed to map and synthesise existing literature on mathematical modelling and predictive analytics in analysing AMR dynamics. The methodological framework used in the review was that suggested by Arksey and O'Malley and expanded by Levac et al., and the review was presented in line with the Preferred Reporting Items of Systematic Reviews and Meta-Analyses Extension to Scoping Reviews (PRISMA-ScR). The search strategy, including full Boolean search strings and date limits, was conducted in the major electronic databases, such as PubMed, Scopus and Web of Science, to locate the literature on the topic published within the range of 2018–2025 for formal database searches, supplemented by the January 2026 preprint repository hand-search described above. Search terms combined MeSH headings and free-text keywords for antimicrobial resistance with modelling methods (mathematical modelling, transmission models, predictive analytics, machine learning, forecasting). Formal database searches were completed on January 15, 2025 (PubMed), January 16, 2025 (Scopus) and January 17, 2025 (Web of Science). A supplementary hand-search of preprint repositories (arXiv, medRxiv) was conducted on January 20, 2026, identifying one additional preprint (Schardong et al., 2026),
Titles and abstracts were screened independently by two reviewers (co-authors), followed by full-text evaluation of potentially eligible articles by the same reviewers. Disagreements were resolved by discussion, with adjudication by a third reviewer where consensus could not be reached. A calibration exercise on a random 10% sample of titles and abstracts was performed prior to full screening to ensure consistent application of eligibility criteria. The essential data extracted from included reviews were the type of models, the pathogens studied, the data sources, the scale of modelling, the validation techniques and the reported results. Given that two of the included studies are a preprint (Schardong et al., 2026)
The review protocol was not prospectively registered in PROSPERO or OSF; at the time of initiation, neither platform formally accommodated scoping reviews of reviews. This is acknowledged as a limitation.
Formal quality appraisal of included reviews using AMSTAR-2
A total of 10 studies were included in this scoping review. Among them, nine were systematic reviews examining mathematical modelling and predictive analytics approaches for antimicrobial resistance, while one study was a narrative review included to provide additional conceptual context and support the interpretation of modelling frameworks (
A scoping review of dynamic models of antibacterial use and resistance conducted by Ramsay et al.
Based on these previous syntheses, Schardong et al.
In a systematic review and meta-analysis on machine-learning models to predict AMR, Tang et al. identified 25 studies until the end of 2021 that utilised ML or risk scores to predict resistance (ML often ESBLs, MRSA or carbapenem resistance) and reported a pooled area under the ROC curve (AUC) of 0.82 in models that used machine-learning to predict resistance (with most models generally having higher specificity but similar sensitivity to traditional risk scores).
Among mechanistic modelling reviews, the majority of AMR models are population-level compartmental models and are given as deterministic ordinary differential equations and typically represent hospital or community systems with homogeneous mixing and few compartments that represent the susceptible and resistant colonisation or infection states.
According to Niewiadomska et al., in the period between 2006 and 2016, over 70% of AMR transmission models involved MRSA, tuberculosis, HIV, influenza and malaria, with very few models addressing the WHO priority bacteria like carbapenem-resistant
Similar gaps were identified by Brinch et al., who list
The specialised review of the HGT models by Leclerc et al. shows that the published literature has centred primarily on conjugation in Escherichia coli in vitro, and did not typically take into account actual exposure to antibiotics or multiple and independent AMR genes. Only 43 modelling studies of HGT followed more than one AMR gene independently, and very few followed transformation or transduction, even though all of these have been shown experimentally to have significant roles in resistance evolution in some species.
The effect of hypothetical interventions on the dynamics of AMR is often studied using dynamic models. According to Ramsay et al., 73 out of 81 included studies investigated at least one intervention, with most studies investigating changes in the overall consumption of antibiotics, switching of classes or altering the management of antibiotics, such as cycling, mixing or combination therapy, as well as infection-control measures, including improved hygiene or isolation.
Poor validation of models and the lack of full reporting of assumptions and data sources are consistent findings across mechanistic modelling reviews. Birkegård et al. applied the TRACE framework to find that none of the 38 AMR models they reviewed met any of the good-practice criteria of problem formulation, model description, data evaluation, internal and external (out-of-sample) validation, sensitivity analysis and output corroboration, and that only a portion of them did any formal sensitivity analysis and very few of them carried out any external (out-of-sample) validation with empirical data.
Deterministic compartmental models with fixed point-estimate parameters predominate and formal sensitivity analyses are inconsistently applied. Ramsay et al. reported that slightly more than half of the dynamic models in their scoping review did any sensitivity analysis,
Most systematic reviews of antimicrobial resistance (AMR) prediction models are patient-level, that is, they predict the likelihood of a certain resistance phenotype (or antimicrobial susceptibility test (AST) result) based on patient-level variables, as opposed to population-level prediction of resistance prevalence. Tang et al. compared the prediction of resistance of individual patients or isolates to selected agents, including methicillin, carbapenems or ESBL-producing status, based on clinical, microbiological and in some cases genomic data using ML algorithms with conventional risk-score methods.
Ardila et al. concentrated on the practical healthcare environment and the WHO critical and high-priority pathogens, and reviewed 21 cohort studies, which applied large hospital AST datasets to predict resistance at the point of care using a ML model.
The systematic reviews all find common methodological shortcomings in studies of ML-based AMR prediction. Common limitations include retrospective single-centre designs, non-standardised preprocessing, selective incorporation of features, small or biased datasets, scant external (out-of-sample) validation, and reporting centred on AUC as opposed to calibration or clinical-utility measures. Tang et al. and Ardila et al. observed that few ML models have been prospectively assessed or implemented in clinical decision-support systems, and thus their practical implications on prescribing behaviour and patient outcomes are unknown.
Across both mechanistic and ML modelling traditions, certain pathogens, settings and geographic regions remain markedly under-represented. High-burden LMIC settings—particularly sub-Saharan Africa and parts of South Asia—are rarely the focus of modelling studies, despite carrying the greatest global AMR burden. Long-term care facilities, informal healthcare settings and the hospital–community–agriculture–environment interface are similarly neglected, despite their centrality to a credible One Health approach. Both modelling traditions show limited coverage of WHO critical-priority Gram-negative pathogens outside acute hospital contexts, and community-acquired resistant infections remain substantially under-modelled.
The present review offers three translational advances beyond existing individual reviews. First, the CCA-based overlap analysis provides an empirical basis for judging whether the two modelling traditions draw on sufficiently distinct evidence bases to support independent synthesis conclusions. Second, the decision-maker guidance table (
| Review (Year) | Domain | Studies/Models | Key Metric | Value | External Validation | Sensitivity Analysis | Principal Limitation |
|---|---|---|---|---|---|---|---|
| Ramsay et al. (2018) | Mechanistic | 81 dynamic models | Studies with sensitivity analysis | ~55% | <10% | 55% | All aggregate compartmental; no individual-based models |
| Birkegard et al. (2018) | Mechanistic | 38 models | TRACE good-practice criteria met | 0/38 (0%) | <10% | <30% | No model met any TRACE criterion |
| Niewiadomska et al. (2019) | Mechanistic | 273 models | Calibrated to data / externally validated | 43% / 14% | 14% | ~50% | >70% of models on MRSA/TB/HIV; only 2% One Health scope |
| Leclerc et al. (2019) | Mechanistic (HGT) | 43 HGT models | Models including ≥2 AMR genes | <5% | Minimal | Not reported | Focused on conjugation in E. coli in vitro; limited in vivo context |
| Brinch et al. (2025) | Mechanistic | 170 models (39 assessed) | External validation reported | 22% (39/170) | 22% | ~33% | Zero code-verification compliance (TRACE criterion 5) |
| Schardong et al. (2026)† | Mechanistic | 36 models mapped | PK/PD integration | Growing minority | Not reported | Not reported | †Preprint (supplementary Jan 2026 search); not peer-reviewed; analysed separately |
| Tang et al. (2022) | ML prediction | 25 studies | Pooled AUC (95% CI) | 0.82 (0.78–0.86) | <20% | Not reported | Retrospective; single-centre designs predominant |
| Ardila et al. (2025) | ML prediction | 21 cohort studies; n>688,000 | GBTs/RF vs. LR superiority | Consistent across pathogens | <15% | Not reported | No prospective validation; non-standardised feature preprocessing |
| Lv & Wang (2024) | ML prediction | Multiple studies (n not reported) | Summary AUC (range) | 0.78 (~0.65–0.90) | <20% | Not reported | Moderate–high RoB; publication bias detected |
Abbreviations: AUC = area under the receiver operating characteristic curve; CI = confidence interval; GBTs = gradient-boosted trees; HGT = horizontal gene transfer; LR = logistic regression; ML = machine learning; PK/PD = pharmacokinetics/pharmacodynamics; RF = random forest; RoB = risk of bias; TRACE = documentation standard for model credibility. Values are approximate proportions or summary statistics extracted from review-level data; heterogeneity across primary studies is substantial and estimates should be interpreted as indicative. †Preprint identified via supplementary January 2026 preprint repository search. This table provides indicative guidance; optimal approach depends on local data, capacity, and policy context.
| Policy/Clinical Question | Recommended Modelling Approach | Minimum Data Requirements | Resource Setting | Key References |
|---|---|---|---|---|
| Population-level resistance prevalence forecast (3–5 years) | Deterministic ODE model calibrated to longitudinal surveillance data | Resistance surveillance (GLASS/EARS-Net); antibiotic usage statistics | Any; GLASS data enable LMIC model calibration | Niewiadomska 2019; Brinch 2025 |
| Antibiotic cycling vs. mixing vs. combination therapy for ICU | Stochastic within-hospital transmission model with multiple intervention arms | Ward admissions/discharges; local resistance prevalence; antibiotic protocols | Hospital (any income setting) | Ramsay 2018; Birkegard 2018 |
| Individual patient risk of resistant organism carriage on admission | Supervised ML classifier (gradient-boosted trees or random forest preferred) | EHR: age, comorbidities, prior hospitalisation, antibiotic history, travel history | Hospital with EHR; HIC primary; LMIC where data infrastructure available | Tang 2022; Ardila 2025; Lv & Wang 2024 |
| AMR gene dissemination across the animal–human–environment interface | Agent-based or network ODE model incorporating HGT and inter-sector transmission | One Health surveillance: human, animal, and environmental resistome data | National/regional; multi-sector data sharing required | Leclerc 2019; Acharya 2023 |
| Cost-effectiveness of stewardship intervention vs. standard of care | Hybrid mechanistic–economic model (Markov/decision-tree linked to ODE) | Resistance prevalence; treatment costs; hospitalisation rates; DALYs | HIC; adaptable to LMIC with local parameter estimation | ;Brinch 2025 |
| AMR trend forecasting in LMIC with limited EHR or laboratory data | Mechanistic ODE with Bayesian inference OR ensemble ML on aggregated surveillance | WHO GLASS aggregates; proxy indicators (antibiotic sales volumes, livestock density) | LMIC; limited data infrastructure | Niewiadomska 2019; Schardong 2026a |
Abbreviations: DALYs = disability-adjusted life years; EHR = electronic health record; GBTs = gradient-boosted trees; GLASS = WHO Global Antimicrobial Resistance and Use Surveillance System; HGT = horizontal gene transfer; HIC = high-income countries; ICU = intensive care unit; LMIC = low- and middle-income countries; ML = machine learning; ODE = ordinary differential equations; RF = random forest.
aPreprint identified via supplementary January 2026 preprint repository search. This table provides indicative guidance; optimal approach depends on local data, capacity and policy context.
Most AMR models are still research prototypes with little translation into decision-support models. Key barriers include poor transparency, the absence of user-friendly interfaces, poor integration with health-economic outcomes (DALYs or cost-effectiveness) and stakeholder engagement.
Box 1 provides an illustrative case vignette demonstrating how WHO GLASS surveillance data streams can operationalise the mechanistic-ML pipeline advocated in this review into a directly actionable stewardship and procurement decision tool for an LMIC setting.
Illustrative vignette: GLASS-calibrated AMR transmission model for stewardship and procurement in a sub-Saharan African hospital setting. Context: A district hospital in a high-burden LMIC seeks evidence to guide carbapenem procurement and empirical prescribing for Klebsiella pneumoniae bloodstream infection. Step 1 (Data sourcing): WHO GLASS aggregated carbapenem-resistance proportions for Step 2 (Calibration): A deterministic two-compartment SIR model is fitted to GLASS resistance proportions via maximum-likelihood estimation of transmission coefficient βᴵ and clearance rate μ; sensitivity analysis explores ±20% parameter variation. Step 3 (Projections): The calibrated model projects carbapenem-resistance prevalence under three stewardship scenarios over 5 years: (a) no intervention; (b) 20% reduction in carbapenem consumption (AWaRe Watch category); (c) introduction of ceftazidime/avibactam as first-line therapy. Step 4 (Economic coupling): Projected prevalence estimates are linked to a cost-effectiveness module using local treatment cost data and WHO-CHOICE unit costs to estimate cost per DALY averted, directly informing national procurement decisions aligned with the national AMR action plan. This workflow illustrates the type of model-to-policy translation that the reviewed literature has consistently identified as critically absent.
This scoping overview is subject to five principal limitations that should be considered when interpreting the findings. First, grey literature was not systematically searched; WHO and ECDC technical reports, unpublished national AMR action-plan modelling studies and conference abstracts may not have been captured, potentially introducing selection bias towards published peer-reviewed evidence. Second, all conclusions are derived from secondary syntheses, and their quality is inherently bounded by the methodological rigour of the included reviews. Third, formal quality appraisal (AMSTAR-2/ROBIS) of included reviews was not performed; while the corrected covered area (CCA) analysis confirmed only slight-to-moderate citation overlap (median CCA = 0.07, range 0.00–0.19), review quality was not systematically weighted in the synthesis. Fourth, the review protocol was not prospectively registered, which is acknowledged as a methodological limitation. Fifth, only 10 review-level syntheses met eligibility criteria, reflecting the nascent state of AMR scoping overview literature and limiting the comprehensiveness of the evidence map. Sixth, the temporal scope of the formal database searches (2018–2025) means that rapidly evolving methodologies—including large language model applications to AMR genomics, foundation-model-based resistance prediction, and federated learning approaches for multi-site resistance forecasting—published after the search date are not captured. These limitations are partially mitigated by dual-reviewer independent screening, a pre-screening calibration exercise on a 10% random sample, a formal sensitivity analysis confirming robustness of primary conclusions, and the quantitative CCA-based evidence overlap assessment.
This scoping review synthesised evidence from 10 review-level studies spanning mechanistic transmission models and ML-based resistance prediction for AMR dynamics. Mechanistic modelling remains dominated by deterministic compartmental frameworks with persistent deficits in external (out-of-sample) validation, One Health integration and LMIC coverage; intervention modelling is rarely linked to health-economic endpoints, limiting the direct policy utility of intervention projections. For ML-based prediction, the consistently high discriminative performance (summary AUC approximately 0.78–0.82) is encouraging; however, near-universal reliance on retrospective, single-centre designs means these estimates may not generalise across diverse healthcare settings, and the near-complete absence of prospective impact evaluations precludes any conclusions about clinical benefit. The principal implication is therefore not simply to develop more models, but to build models that are better validated, transparently reported and directly linked to decision-relevant outcomes, such as cost-effectiveness and patient outcomes. Future research priorities should include expanding pathogen and geographic coverage—particularly for WHO critical-priority Gram-negative organisms and LMIC settings that carry the greatest AMR burden—integrating health-economic endpoints such as DALYs and cost-effectiveness, and developing hybrid mechanistic-ML frameworks capable of reflecting the multi-scale, multi-sectoral complexity of AMR. Realising this potential will require not only methodological innovation, but also greater investment in data infrastructure, open-code practices and structured engagement between modellers, clinicians and policymakers. Aligning future modelling efforts with existing global frameworks—including the WHO Global Antimicrobial Resistance and Use Surveillance System (GLASS) and national AMR action plans—will be essential to ensure that model outputs are relevant to and usable by decision-makers at national and international levels. GLASS currently aggregates standardised AMR surveillance data from over 70 countries, providing an unprecedented platform for calibrating and validating population-level AMR models in LMIC settings where primary data infrastructure remains limited. Future modelling consortia should prioritise co-production with LMIC researchers and health ministries, leveraging GLASS data streams alongside the WHO Global Action Plan on AMR and the Tripartite Joint Plan of Action on AMR (2022–2026), to develop regionally calibrated, decision-ready models directly informing national stewardship programmes, antibiotic procurement policies and One Health action plans.
Additional material is published online only. To view please visit the journal online.
N/a
N/a
N/A
There is no conflict of interest
Sumit Sharma: Writing – original draft
Sumit Sharma
Unsolicited and externally peer-reviewed
This is a review paper, and we have already submitted the data in paper itself. But if data is required during the peer review process, we will make it available