An agent-based model for collaborative learning to combat antimicrobial resistance: proof of concept based on broiler production in Senegal

Antimicrobial resistance (AMR) is a substantial global One Health problem. This paper reports on initial, proof-of-concept development of an agent-based model (ABM) as part of wider modelling efforts to support collaborations between groups interested in policy development for animal health and food systems. The model simulates AMR in poultry production in Senegal. It simultaneously addresses current policy issues, builds on existing modelling in the domain and describes AMR in the broiler chicken production cycle as seen by producers and veterinarians. This enables implementation and assessment of producer antimicrobial use and infection prevention and control strategies in terms of immediate economic incentives, potentially helping to advance conversations by addressing national policy priorities. Our model is presented as a flexible tool with promise for extension as part of AMR policy development in Senegal and West Africa, using participatory approaches. This work indicates that ABM can potentially play a useful role in fostering counter-AMR initiatives driven by food system actor behaviour in lower- and middle-income countries more generally.


Introduction
AMR is a major threat to global human health, with the latest evidence showing that the highest mortality burden of AMR is in West Africa [1].Action on AMR is accordingly a growing priority for global health systems, under a One Health approach designed to address food safety as well as zoonoses, with an emphasis on the development of knowledge and evidence through surveillance and research [2,3].
Antimicrobial use (AMU) in poultry production is vitally relevant to the counter-AMR agenda in this context because of this food system's importance to poverty alleviation and economic development efforts, including from a food security perspective [4].Unfortunately, too little is known about apparent causal pathways to AMR human health burdens.But there is another more immediate and potentially compelling aspect to the AMR threat from within animal health: the risk that treatment failures will drive production losses and food insecurity [5].This is the main motivation for efforts to develop relevant socio-technical interventions, for example through research on African farmers' knowledge, attitudes and practices (KAP) [6,7].KAP studies point both to the need and to the means to combine "top-down" regulatory interventions with "bottom-up" stakeholder-led local adaptation (plus infrastructural change to provide an "enabling environment" in resource-limited contexts), promoting investment in preventative practices relative to antimicrobial treatment to reduce reliance on antimicrobials.
Senegal is committed to action on AMR as a policy priority under the leadership of the High Council for Global Health Security (HCNSSM).The country stands out for its dynamic efforts to understand AMR in animal health.Antimicrobial use and consumption (AMU/C) aspects of livestock production have been a focus of attention for the Ministry of Livestock and other stakeholders, including the Emergency Centre for Transboundary Animal Diseases (ECTAD) supported by the United Kingdom's Fleming Fund [8].There is a vigorous animal health AMR research community in Senegal, with a variety of studies considering the poultry sector in particularfor example [9][10][11][12].Various reports have highlighted the increasing use of antimicrobials in the sector, with the resistance of bacteria such as Echerichia coli and Salmonella to commonly used antibiotics [13].Recent indications of high threat of drug resistance in Senegal relative to other African countries in the human health context [14], underline "calls for urgent policy intervention" [15].
Still various factors complicate the case for relevant action on AMR, notably obscure effects, including AMU economic incentives in terms of productivity and competitiveness, and the need for market-based solutions, especially given limits to regulatory supervision in animal health.Although economic studies point in this direction, for example for broiler producers in France [16], both animal health and LMIC are under-represented in economic evaluation of AMR-related interventions [17].Still, new initiatives like the Food & Agriculture Organization of the United Nations (FAO)'s Reduce the Need for Antimicrobials in agrifood systems initiative (RENOFARM) recognize economic objectives as the key to reducing AMU by aiming to improve countries' agrifood systems transformation through the provision of comprehensive support in the implementation of good production practices [18].In Senegal, the profile of the poultry sector demonstrates the urgency of the issues and policy development potential.Peri-urban areas of major cities Dakar and Thies exhibit intense livestock-rearing activity, especially chickens, with high animal density in close cohabitation with human settings.The great majority of poultry production is semi-intensive with average rotations of between 300 and 1000 subjects.Public policy plays a vital and active role, underpinning growth with protective measures since 2005 and regulating prices [19,20].
The implication of complex adaptive systems including both human behavioural interventions and disease dynamics make this a potentially fertile subject for ex ante ABM [21,22], as an input to ongoing AMR policy discussions in Senegal and beyond.ABM contrasts with more traditional modelling approaches, for example using sets of interrelated differential equations, in that it offers a natural way to describe systems characterized by many levels of interactions, and better capture emergent phenomena [23].The autonomy of individual agents in ABM allows them to be used as structures for simulation of systems with multiple "loci of control".Strikingly, ABM has been used extensively to simulate systems in several domains recognized to be constitutive of AMR as a policy problem, including infectious disease itself at both microbiological and population levels, animal herding/flocking (relevant to livestock), economic systems and behavioural policy.AMR ABM efforts can therefore draw together existing conceptual structures.
On the other hand, there are also considerable challenges for model development (including ABM) in the context of AMR in LMIC, especially the lack of relevant data on AMU/C in agriculture and food systems in Africa [24].Modelling under these circumstances has been conceptualised more as a means of empowering actors lacking social and political information than as a way of making sense of existing data flows [25].
This paper reports on model development to support collaboration between researchers, policymakers, veterinary professionals and producer businesses in efforts to combat antimicrobial resistance (AMR).Its immediate objective is to contribute to current policy dialogue over AMR in poultry production in Senegal and West Africa.Beyond that this research aims to advance agent-based modelling (ABM) as tool in the field, especially in lower-and middle-income countries (LMIC).

Material and methods
An ABM was developed in the NetLogo environment, simulating a broiler chicken production environment cycle typical of semi-intensive production characteristic of the Dakar and Thies regions of Senegal.Model code and a description conforming to the Overview, Design concepts, Details (ODD) standard that is a key component of the "TRACE" approach to model development for ABM [26,27] have been made available at (https://github.com/archiedrake/senegalabm.git).The simulation was configured to describe rearing of 500 birds in a standard facility over approximately 42 days, with multiple resistant and non-resistant dynamic bacterial strain populations (including Escherichia coli and non-typhoid Salmonella spp.) represented in addition to varying concentrations of multiple antimicrobial agents (designed to represent tetracycline and fluoroquinolones).Fig. 1 shows a comparison of the model observer interface with a practitioner's representation of a standard rearing facility.
Initial development aimed to elaborate a stable simulation, grounded in existing evidence, sufficient to be taken forward in participatory processes and to support initial observations about the added value of ABM in this application.Fig. 2 lays out simplified representations of model elements, with more detail available in the ODD documentation mentioned above.
Model design drew on existing mathematical modelling of AMR [28], in particular settling on Graesbøll et al.'s approach to modelling AMR in pig farming [29], which describes AMR dynamics as a function of competition between strains with varying susceptibility and fitness costs.Although this is necessarily a simplification of understandings of AMR dynamics, for example excluding horizontal transmission mechanisms [30], a competition model capable of supporting multiple bacterial strains and multiple antimicrobial agents with minimal modification offered considerable advantages in terms of representing real-world production environments in a flexible and extensible way.The clearest advantage was being able to move beyond exclusive focus on a single "bug-drug" combination, which is a clear limitation in many mathematical models in the field, with no limit to the number of bacterial strains or antimicrobial agents which our model could incorporate in principle.The competitive model also enabled relative simplicity in key modelling dynamics: bacterial strains defined exclusively in terms of growth curves according to antimicrobial concentrations; and a core equilibrium relationship ultimately governing the evolution of strains relative to each other.
Mathematical models were further deployed to add disease mechanics to the model, emphasising usefulness of representing individual agent immune responses to pathogenic bacteria and maximal simplicity.These considerations favoured mobilisation of a simplified version of Mochan et al.'s intra-host model for pneumonia in mice [31], with the model representing the complicated pathogenesis and epidemiology of colibacillosis and salmonellosis [32,33] in radically simplified terms.
The model was configured to produce a text-based output from each simulation run which summarised some of the outcomes relevant to stakeholders in economic, human health and environmental practices.An example output text is provided in Appendix A.
The team referred to practitioner experience within the author group, plus field data to improve representation of current realities in the chosen region of Senegal.For example, the "enclosure cleaning" originally included as a basic infection prevention and control (IPC) intervention which the model observer may deploy was replaced with an "add litter" intervention in response to the comment that producers rarely or never undertake cleaning during a production cycle.In addition the team analysed fresh AMUSE survey data from Senegal for model calibration purposes [34,35].
Model development culminated in testing and calibration, generating an initial parameter space (Appendix B) and the design of a set of experimental intervention strategies for evaluation over repeated simulation runs (Appendix C).

Results
Our work shows that it is possible to mobilise computer-based ABM simulation of food production systems in LMIC in ways that illustrate the relevance of AMR to the practical challenges and priorities of producers and veterinarians.
The first result is production of a stable and robust model with a high degree of flexibility in terms of potential for integrating cross-disciplinary evidence for impact on practice and policy.Local sensitivity analysis using a one-at-a-time approach [36] with Netlogo's embedded Behav-iorSpace tool suggested a viable initial parameter space, as indicated in Appendix B, capable of supporting demonstrations for stakeholders illustrating the relevance of AMR factors to their day-to-day priorities including incidence of mortality and morbidity.
The second result is calibration of the model environment to bestavailable field data.The initial parameter space was explored to suggest a set of values obtaining good-enough fit for baseline purposes.
Although non-use of antimicrobial treatments in the study context is reportedly rare, with small sample sizes generating wide margins of error in estimates of applicable mortality rates, this represents the soundest empirical basis for calibrating the model.We used the BehaviorSearch tool to fit for target mean mortality with no interventions, followed by further BehaviorSpace testing to settle on a reasonable basis.Fig. 3 gives an example visualisation output from this process.
The third result is a set of in silico experimental observations obtained from simulation runs over the set of AMU strategies identified as being most relevant to the model setting (see Appendix C).See Table 1.As expected, this shows the advantage to broiler producers in terms of mortality reduction of using antimicrobial treatments (strategies 2-4) over purely hygiene-based (strategies 5 & 6) or no-intervention approaches (strategy 1).The prophylactic approach (strategy 4) was much less effective at reducing mortality than expected, probably because of the low antimicrobial concentrations used for experimental purposes.Meanwhile the "quick stop" strategy (strategy 2) was highly effective at reducing mortality, but at considerable externality cost in terms of nonrespect of withdrawal periods prior to slaughter and associated risks to human health through meat residues [37].The fourth result is a set of initial observations about AMR policy in the broiler production sector in Senegal based on experience of model development and behaviour.The most significant observation is the value of addressing human behaviour, especially in market systems context given animal health challenges.Our investigation originally sought to engage with underlying resistance dynamics in the study context, for example supplying explanations of how treatments can contribute to the general problem of growing resistance (Fig. 4).This proved significantly less interesting to domain specialists than flock health surveillance for the purposes of maximising productive efficiency (Fig. 5).The utility of modelling to support policy reflections in this context therefore depends on demonstrating the relevance of AMR and AMR-related initiatives to production goals, rather than vice versa.For modelling purposes, each of these dimensions remained as core emergent dynamics (ODD documentation mentioned above for more detail) and so are represented here to promote understanding.In terms of observed practical significance, productive efficiency was an apparent priority because of direct association with livelihoods.

Discussion
Model development for this study was undertaken in line with best practice in mathematical modelling of AMR [28].Our model deploys an advanced mechanistic approach (nested agent-based), incorporating stochastic elements to represent uncertainty and variability, for example in the growth rates and resistance profiles of bacterial strains, and including some initial sensitivity analysis with internal and external validation.
Care was taken to accommodate multiple bacterial strains and multiple antimicrobial agents, moving beyond the typical limitation to single organism and/or treatment perspective and address various "drug-bug" combinations.Although the model still only speaks to bacterial disease despite producers also facing viral, fungal and environmental issues, the multi-level approach enabled the initial model to address two combinations recognised as priority issues for AMR in relation to poultry rearing in Senegal and the West Africa region.Very high AMR to tetracycline have been reported in both Escherichia coli and Salmonella spp. in Senegal [13], as well as in Mali [38] and Togo [39].Although still considerable, rates of AMR to fluoroquinolones are reported to be lower in these studies; but this is arguably of even greater concern given their important role in human health [13].Given the lack of published data on the resistance profiles of resistant strains, it remains to be seen whether the level of detail included at strain level might be externally validated in due course, for example using surveillance data.
To our knowledge, existing similar work on AMR in broiler production focuses on flock colonisation by relevant bacteria without including pathogenic mechanics.This limitation in scope has enabled excellent work with more detail and external validation than possible for our model to date, notably in Becker et al.'s study [40], which was  also based on Graesbøll and generated insights on breeds, litter renewal and stocking densities.Our gross simplification of the relationship with infection, immunity, disease and control across multiple complex categories of poultry health conditions is difficult or impossible to justify against the modelling standards applied in such existing work.However, we believe that it is a reasonable starting point for the purposes of a model aiming to demonstrate the relevance of AMR ABM to producers and policymakers in Senegal.There is in any case a general lack of in vivo models of relevant disease on which to call here [41].
The clearest example of intervention assessment from our model development to date relates to the way in which research and policy assess current AMU by poultry producers.Experimental observations from the current model are generated in the context of AMUSE field survey data on the current prevalence of AMU strategies (Fig. 6).These observations can help focus discussions on potential intervention priorities, for example deciding amongst the following three channels for policy development.Firstly, food safety regulatory action on withdrawal period compliance in meat supply chains.Secondly, producer information campaigns on the cost-ineffectiveness of prophylactic strategies.Thirdly, peer-based promotion of good practices amongst producers and/ or animal health professionals.
More directly, these observations can support conceptual refinements to future rounds of data collection, for example aiming to understand specific molecules and concentrations used in prophylaxis.
Our discussion now turns from the current model status to consider prospects for future model development.Experience on modelling for decision support and policy indicates that the development process, particularly emphasising collaboration and stakeholder participation, acts a critical determinant of eventual value [42,43].In One Health contexts like AMR, pragmatic adaptation to address partner priorities appears to be a key determinant for sustaining initiatives over time [44].We envisage future model development to continue to emphasise participatory focus on stakeholder priorities, especially low-cost local disease surveillance for rearing programmes.This suggests some potential for wider model contributions to animal health and/or food systems policy, although we confine our detailed reflections here to AMR policy in line with study objectives.
Having established proof of concept, the next step is to engage with groups in AMR initiatives in Senegal to test assumptions about what ABM addsespecially alongside other modelling approachesto establish a foundation for further development.It is expected that our work to date can add reflections on ABM utility in this context alongside modelling work being undertaken under the Selecting efficient farm-level antimicrobial stewardship interventions from a One Health perspective (SEFASI) project [45].The SEFASI project is developing a quantitative System   Dynamic model (SDM) to test intervention scenarios and ex-ante impact assessment of the cost-effectiveness of the interventions at farm level and beyond using a One Health approach [46].Initial causal loop diagram development indicates a high degree of complementary between SEFASI SDM work and our work, especially offering different perspectives on the issues of farm productivity, profit and investment in good management practices [47].
One apparent strength of ABM in this context is its capacity to simulate the perspective of the people whose behaviour is understood to determine AMR-related outcomes.Most modelling involves more deterrent levels of abstraction, for example demanding system-wide overviews or concentrating on causal relationships that are assumed to be important.The observer interface of our ABM, by contrast, is designed to simulate a broiler flock rearing cycle in a way that is recognisable to producers and can help other groups develop a more detailed appreciation of their perspective.The simulation framework is flexible enough to accommodate the very wide range of dimensions relevant to assessing AMR-related interventions.Questions about how relevant dimensions have been or might be incorporated can at least help stimulate discussions about how interventions can be assessed and quantified [48].
Ideally, there is significant potential for further collaborative development of the model using participatory approaches, and perhaps even moving towards more elaborate interactive simulation or serious gaming [49].
More generally, there is an opportunity to build on mapping techniques for representing these system-level factors in terms of their composite actions and actors.This may provide a foundation to support AMR policy practice and programming through coherent integration of wider factors into future iterations of our ABM.In particular, conceptual frameworks for engaging with actor behaviour consistently across social systems is likely to prove helpful as a next step for assessing modelling priorities oriented towards evaluable intervention design [50,51].It seems unrealistic to expect that information system improvements, for example the availability of AMR and AMU surveillance data in Senegal, will transform quickly or profoundly enough to overcome the severe limitations of ABM in terms of external validity.Similarly, there is a risk that modelling for modelling's sake will exacerbate the existing issues of model complexity and resource-intensiveness.Instead it would be useful to use existing data and evidence, in participatory policy conversations, to focus modelling, interventions and evaluation on viable behaviour changes [52].
Our results are presented with acknowledgement of limitations.This paper presents results from proof-of-concept model development research.Its contribution is therefore conceptual, with no novel empirical or analytical insights claimed.We show the relevance and utility of ABM in this context as a reference point for future work rather than making a substantive contribution to ongoing debates at this stage.

Conclusions
This paper has presented findings from the proof-of-concept stage of research into ABM for AMR, using Senegal poultry production as a case study and aiming to inform collaborative policy development.We present some results from initial model development, with discussion focused on the potential future elaboration of this approach.
Conventional application of the terms "irrational use", "inappropriate use" or "misuse" of antimicrobial treatments tends to underestimate the extent to which day to day AMU in LMIC agriculture addresses immediate economic goals.Senegal is well-placed, in collaboration with international partners, to contribute conceptual frameworks which can help drive closer engagement with the issues involvedincluding through further development and validation of future versions of the model proposed here.Many poultry producers in Senegal feel that they have little choice but to make extensive use of antibiotics, faced with the perceived alternative of disease and death driving them out of business.
ABM may prove useful in this context by suggesting ways to think about the problem differently, for example the possibility that farmers may be wasting their time and money on strategies that don't work well or the likelihood that some farming practices may be making it more difficult to fight human diseases.In terms of national policy, our model might contribute to novel thinking which considers the problem as a potential stimulus for investment in improved farming practices rather than as a competing priority.
Although we present associated results and discuss related ideas in this paper, its main contribution is to introduce the model and make the model available to wider research and policy communities.It is hoped that this will promote debate and collaboration around AMR in Senegal, as well as ABM for AMR generally.independent evaluation of the Fleming Fund.This is not a material conflict since the Fleming Fund is only mentioned in passing as a notable AMR initiative in Senegal which is not controversial.
Michel Dione and Ardiouma Faye report an interest in the SEFASI project.They are, respectively, employed by and affiliated with the International Livestock Research Institute (ILRI) which is funded by SIDA to implement the SEFASI project (see Funding Statement).This is not a material conflict because no view of SEFASI is presented beyond descriptions of the work as supported by published material and the opinion that the model presented in this paper will be useful to advance SEFASI objectives.

Fig. 1 .
Fig. 1.Model observer interface (left) alongside representation of a standard semi-intensive broiler rearing facility (right).(Chicks are initially reared in circular starter enclosures (top), moving to occupy the entire rearing facility for later production phases (bottom) (Source: model observer interface/Dr.Njiemessa Nsangou)).

Fig. 2 .
Fig. 2. Simplified UML class and state diagrams showing key model elements.

Fig. 3 .
Fig. 3. Example density plots from exploratory calibration of model to base case (no AMU).(Survey ¼ estimated distribution of mortality rate from field data [34, 35].Fit 1 & fit 2 ¼ distribution of mortality rates using values shown in Appendix B (only varying seasonal impacts on bacterial growth rates in the model environment)).

Fig. 4 .
Fig. 4. Model populations of bacterial strains in an individual chicken over the course of simulation runs.(In the absence of antimicrobial treatment (left), susceptible and resistant strains co-exist in equilibrium in environmental context.But periods of antimicrobial treatment are disruptive (righttreatment over c.550-650 h), conferring competitive advantage on relatively resistant strains and so tending to promote AMR generally).

Fig. 5 .
Fig.5.Key model variables from a producer's perspective.(Apparent infections (left) may lead to the death of chickens (counted at centre), which impacts the total mass of birds (right) and therefore ultimately the value of the meat that will be sold).

Table 1
Summary mortality statistics and antimicrobial residue in meat due to nonrespect of withdrawal periods, over 200 simulation runs over different AMU strategies.