Antimicrobial resistance modelling is decades behind climate science, and the gap is costing lives

International research collaboration. Published via Shanghai Jiao Tong University Journal Center.

Antimicrobial resistance kills more than a million people per year and threatens to unravel modern medicine's dependence on effective antibiotics. Yet the mathematical models that should inform AMR policy are, by the standards of other global threats, remarkably underdeveloped. A new review by an international research team lays out exactly how far behind the field stands and what it would take to close the gap.

89% of models look at humans alone

The team, drawn from the University of Edinburgh, the London School of Hygiene and Tropical Medicine, North Carolina State University, and the International Centre for Antimicrobial Resistance Solutions, analyzed 273 population-level AMR models published to date. The results are sobering.

Of those 273 models, 89% considered only humans. Seven percent included animals. Two percent included plants. Not a single model integrated all three sectors, despite the scientific consensus that AMR is a One Health problem spanning human medicine, veterinary practice, agriculture, and the environment.

Only 9% of models included any economic cost-benefit analysis. Forty percent included no sensitivity or uncertainty analysis. And none, zero out of 273, met the TRACE modelling guidelines established in 2010 for transparency and reproducibility in mathematical modelling.

The "wicked problem" that fails to mobilize

The researchers frame AMR as a "wicked problem" in the policy science sense: its impacts are cumulative and largely invisible, it involves diverse microorganisms with different resistance profiles across different drug classes and contexts, and the costs and benefits of intervention are distributed unevenly. Unlike an acute pandemic, AMR does not generate the public urgency that drives political action.

A fundamental obstacle is that the relationship between antimicrobial use (AMU) and resistance remains poorly understood across many specific bug-drug combinations. Global AMU estimates rely on multiple layers of inference, often built on unavailable or biased baseline data. The degree to which livestock and aquaculture antimicrobial use drives resistance in human infections remains actively contested.

What climate science got right

The review draws a pointed comparison to climate change modelling, which has developed standardized abatement cost curves, the "social cost of carbon" metric for economic decision-making, and the Intergovernmental Panel on Climate Change (IPCC) as a coordinating body that synthesizes evidence across models and disciplines.

AMR has none of these. There is no equivalent of the social cost of carbon for antimicrobial resistance. There is no integrated modelling architecture for framing national or international cost-benefit decisions. And while the UN Quadripartite Group on AMR is negotiating the creation of an Independent Panel on Evidence for Action against Antimicrobial Resistance (IPEA), it does not yet exist.

"AMR policy is enacted without economic efficiency evidence," the researchers note. Even for well-studied antimicrobial use interventions, the relative cost-effectiveness across different settings lacks clarity.

The validation bottleneck

Mathematical models progress through a hierarchy: from theoretical models, to fitted models with internal validity, to models externally validated against independent datasets, to multi-model comparison exercises. Most AMR models remain stuck at the lower rungs. External validation is hampered by limited independent data, and multi-model comparisons, which proved valuable during COVID-19 modelling, are currently not feasible for AMR because the models are too heterogeneous and the comparable data too scarce.

Where the field needs to go

The review calls for transdisciplinary collaboration to build integrated modelling architectures that span human, animal, and environmental sectors. It acknowledges that data harmonization across phenotypic, genomic, and metagenomic measurement methods remains a major challenge. Human infection samples are overrepresented in surveillance programs while environmental data is scarce.

The practical recommendations include digital One Health frameworks to maximize surveillance efficiency, mandatory data and code transparency in scientific publishing, and genuine cross-sector model development rather than the siloed approaches that currently dominate.

Whether these recommendations gain traction depends on political will, and the review's concluding argument is that political will itself depends on better models. Without convincing economic evidence, AMR will continue to be treated as a slow-moving crisis that someone else should solve. The modelling gap is not just a scientific problem. It is a policy problem that perpetuates inaction on a threat that is already here.