I don’t think I need to convince anyone reading this that AMR is responsible for a substantial health burden! Unfortunately, “substantial” is not accurate enough to inform policy. Estimating the future burden of AMR is essential to understand the scope of the problem we have to solve, and implement the best solutions. Models are often used for this purpose, but we must keep in mind the important uncertainties of model estimates, and the sources of these uncertainties.

In a recent article, Hillock et al reviewed the feasibility and value of modelling to inform policy by predicting future rates of AMR. Cost-effectiveness is a key metric to decide which interventions should be implemented, balancing the costs with the benefits. Estimating costs is arguably the easier part. On the other hand, understanding the benefits of an intervention requires an understanding of the counterfactual: what will happen if we don’t implement that intervention? That’s where models can help.

The simplest models estimating future trends will take a past/present trend, and extend it over time. Sounds easy? Well, the problem with AMR is that there’s a substantial element of uncertainty: for example, the appearance of new resistance mechanisms in bacteria is highly unpredictable. Therefore, predictions at the human population level come with inevitable uncertainty. In fact, this uncertainty can even be seen in studies which “only” try to estimate current burden of AMR, not even future burden, like the GRAM study (see this Twitter thread by AMR Centre co-director Gwen Knight for some thoughts).

The magical solution to this problem: data. A complete, accurate picture to inform models reduces the need for assumptions, and suddenly the mist in the crystal ball fades away. However, as you might have guessed already, this is too good to be true, and the data is never perfect. Many models estimating AMR burden rely on surveillance data, such as the popular WHO GLASS (for some highlights revealed by this data, see this summary by AMR Centre member Sam Willcocks on the GLASS report released in summer 2020). While these datasets are incredibly useful, geographical gaps in surveillance are still frequent, and the values they present may be biased (e.g. mostly reporting isolates from more severely-ill patients). Hillock et al nicely discussed these gaps in their review, so I encourage you to go read this for more information.

In summary: predicting the future burden of AMR is essential, and models can help. However, if we don’t already know what’s currently happening, then we can’t accurately predict what will come later. To solve this, we must develop better surveillance systems right now. Members of the AMR Centre are working on these questions, using modelling to estimate the burden of latent drug-resistant tuberculosis for example, so watch this space for updates on this topic!