Towards virtual epidemiology: an agent-based approach to the modeling of H5N1 propagation and persistence in

North-Vietnam

Edouard Amouroux1,2 , Stéphanie Desvaux3,4, Alexis Drogoul1,2

1IRD UR079 GEODES, 32 Avenue Henri Varagnat, Bondy - France 2Equipe MSI, IFI, 42 Ta Quang Buu, Ha Noi - Viet Nam

3 CIRAD, Campus International de Baillarguet F-34398 Montpellier - France 4 PRISE Consortium in Vietnam c/° NIVR, 86 Truong Chinh, Ha Noi - Viet Nam

{edouard.amouroux, alexis.drogoul}@ird.fr, stephanie.desvaux@cirad.fr

Abstract. In this paper we claim that a combination of an agent-based model and a SIG-based environmental model can act as a “virtual laboratory” for epidemiology. Following the needs expressed by epidemiologists studying micro-scale dynamics of avian influenza in Vietnam, and after a review of the epidemiological models proposed so far, we present our model, built on top of the GAMA platform, and explain how it can be adapted to the epidemiologists’ requirements. One notable contribution of this work is to treat the environment, together with the social structure and the animals’ behaviors, as a first-class citizen in the model, allowing epidemiologists to consider heterogeneous micro and macro factors in their exploration of the causes of the epidemic.

Keywords: Multi-Agent Systems, Agent-Based Models, epidemiological models, environmental models, GAMA platform.

1 Introduction

Over the past few years, avian influenza spread from Asia to Europe and some parts of Africa. Following this proliferation, a certain downturn in occurrence has been observed thanks to many measures (improved hygiene vaccination programs, etc). Although total eradication remains elusive and this disease is still a major threats both the economy and the public health. In this context, the challenge for current epidemiology is to eradicate the virus, which means understanding the factors that may impact this spread. In particular, its local propagation and the re-emerging mechanisms (epidemic outbreak in apparently virus free areas) are yet to be fully understood. Current hypotheses point towards: the presence of wild birds, traditional farming practices and trading activities — though none of these have been validated as of yet. Some recent studies [1] suggest that human activities, particularly within the agricultural system, in strong interaction with the dynamics of the environment, play a key role in the spread of the disease at local levels (within a given district). To have a better understanding of these mechanisms, one of the possible ways for

epidemiologists is to reason about simulated models of reality and to test hypotheses about the environment, the social structure, the behaviors of the birds, etc. However, existing modeling techniques are not really appropriate for such an exploratory use of simulation. The aim of this paper is to present an approach based on Agent-Based Modeling (ABM) and detail how it can be used to help epidemiologists answer their questions. After presenting the context and requirements of this work, we will review existing epidemiological models, explain why they do not fit the requirements and why we have chosen to develop our own tools on top of a multi-agent platform. We will then present our model, explain its implementation and advantages over present art and, finally, discuss possible methods of validating its adequacy.

2 The Context of Avian Influenza in Vietnam

Avian influenza (HPAI) epidemics occurred for the first time in mid-2003 then recurrently every year. These epidemics have a wide impact on human health and on the economy, especially in Vietnam (106 infected people, 52 deaths and 9 millions household farms infected) . The situation is similar in both North and South Vietnam yet the underlying variables of their respective poultry production industries are different: climactic and environmental conditions, variety of circulating virus strain [3] and poultry production organization.

2.1 Two Epidemiological Questions to Address

As of today, many basic epidemiologic questions (about avian influenza) still need to be answered. At a macro level, propagation is not precisely understood but the general tendencies are well. However, at the micro scale (e.g. village or commune level), the explanations of propagation mechanisms are hazy at best. Because understanding what is happening at these micro-levels is probably the key to control the epidemic, many investigations are focusing on them, and especially on these specific questions: (1) Propagation mechanisms: how does the virus propagate locally (from birds to

birds, humans, environments, farms, markets, etc. and back)? (2) Persistence mechanisms: how can the virus re-emerge in a previously uninfected

area, sometimes months after the end of the previous wave of the epidemic?

2.1.1 Local Propagation Mechanisms Both mechanisms can be explained by the presence of wild birds [6] and by trading activities. Other recent hypotheses concern the role of the agro-system. On one hand, many recent policies focus on the role of traditional farms, as they have no bio-security system in place, though these farms act at a very local scale. On the other hand, scientists focus more on semi-industrial farms where studies are being conducted. At this level, the agro-sector has better bio-security, however not secure enough as semi-industrials have a much wider network of connections, and thus a wider span of influence. The fully industrialized farms is to be considered as well but as they very high bi-osecurity systems, they are not really much of a concern for the

experts. Whatever the considered production sector is, the whole agro-system is nowadays considered and monitored extensively by epidemiologists (the farm and its suppliers and customers) in order to understand how the virus propagates locally.

2.1.2 Re-emergence Mechanisms Re-emergence is another concern. The prevailing hypothesis is that wild birds act as reservoirs though there is, at the moment, no validation of it [6]. Another hypothesis concerns the environment itself as a reservoir. In-vitro experiments are currently being carried out on the persistence of H5N1 within the environment and especially in the water [7]. It has been demonstrated that avian influenza viruses are likely to remain infective several months while kept in places like ponds [7].

2.2 Addressing Epidemiological Questions through Simulated Experiments

Towards simulated experiments Local-scale propagation, persistence and re-emergence processes are not simple to understand. More and more epidemiologists now focus on studying and understanding the interactions between domestic birds, environment, wild-birds and human activity. Unfortunately, the data that has been produced to date is insufficient in term of completeness and reliability and it will probably remain the same in the near future. To have more reliable data, in-vitro studies are being conducted, but the conditions are too far from reality to be easily transposed to field reality. Since neither field nor in vitro studies can be completely satisfactory, experts turn towards the use of models and simulation experiments to validate their hypotheses, and explore new ones.

2.2.2 Requirements for ”Virtual Epidemiology” Experiments As we will see in the next section, however, existing modeling techniques are not really suited for such “in silico” experiments. Indeed, the ideal tool would need to answer the numerous requirements expressed by epidemiologists, since testing hypotheses about re-emerging and propagation mechanisms involves the possibility to reason about heterogeneous elements at different time and spatial scales. The most important are listed below: (1) The environment should be represented extensively together with its own

dynamics. This representation should also enable the use of various and heterogeneous descriptions (geographical, social, ecological, etc).

(2) Epidemiologists need the ability to work both at the population level and directly on models of individuals, depending on the hypotheses they need to validate.

(3) Interactions at the population or individual levels, but also with the environment, should be modeled explicitly, since a change in them may have an impact on the dynamics of the epidemic.

(4) These data are usually expressed using heterogeneous formalisms that the expert needs to be able to reuse in the modeling of the system.

(5) The simulation platform should be able to recreate an environment in which “virtual experiments” could be run (exploration of parameters, etc.).

3 Epidemiological Models: a Review

In this section, we review the existing epidemiologic models in the light of these requirements. Although the literature on the subject is quite important, we will see that most of the offers do not fulfill them completely yet.

3.1 Epidemiological questions and models

How to… Model Reference

Represent and study epidemics at a global level? SIR Bayesian Networks HMM

Yorke 79 Abbas 04 Durand 99

Account for the heterogeneity of the population? Micro-simulation Artzrouni 01

Represent contact patterns? Cellular automata Social networks

Turner 01 Ghani 07

Represent different individual behaviors? Individual based Wilensky 98

Study the role of the environment? Agent based Agent based + GIS

Muller 04 Badariotti 05

Table.1 : Summary of the relationships between questions and models

How to represent and study epidemics at a global level? The primary question asked to epidemiology by society is whether or not an epidemic may spread all over a whole (human or animal) population or fade out quickly? SIR compartments models [10], based on differential equations, are a way to determine it. Here individuals are classified according to their status regarding the infection (Suspicious, Infected, Recovered). To do so epidemiologists evaluate, with field data and previous epidemics data, the global transition rates from one compartment to another. Then they are able to evaluate if the epidemics is starting or fading.

However, epidemiologists usually manipulate many other factors (age, sex, etc) that are known to impact this dynamics though SIR models are not able to take those into account. Bayesian network [11] and hidden Markov chains [12] are a way to do so. Here global factors (age, sex, etc) are used to adapt a general model of an epidemic to the actual situation. Unfortunately these models are not suited for our concern as they operate only at a global scale. How to take the heterogeneity of the population into account? In the previously presented models, the population is always considered homogeneous — which is obviously untrue in reality. Epidemiologists know that population heterogeneity can greatly influence the propagation of epidemics. To represent this heterogeneity, it is necessary to explicitly represent, in a model, individuals (or groups) instead of the global population: computational models have offered this possibility. Micro-simulation is the first type of model to have addressed these issues: individuals are represented by a vector of parameters, and the whole population by a matrix [individuals x parameters] on which global computations are made. For example, Artzrouni’s work on sleeping sickness moved from classical differential equation based models to micro-simulation ones [13] in order to represent systems

with small populations and be able to discriminate between situations with similar characteristics; a small difference in the number of infected Tsetse (sleeping sickness vector) may dramatically change the outcome of the global epidemic. This kind of models also allows evaluating the influence of the population’s heterogeneity on the dynamics of the epidemic. However, micro-simulation reasons on structurally identical individuals and processes only occur at the global level.

How to represent contact patterns? Micro-simulation represents the individuals themselves but not their relationships. These relationships can however be crucial to understand the dynamics of epidemics. Of particular interest for epidemiologists is the spatial organization of the individuals particularly the neighborhood organization [14]. Obviously, two individuals located at distant locations have a smaller probability of interaction than two neighbors. A way to model this hypothesis is to represent this topology using a grid where individuals can be located. As in the previous model, every cell of the grid contains a parameters vector that represents an individual. Their evolution (which occurs locally) is based on this vector but it is also influenced by the state of its “neighbors”. Turner et al. showed in [15] that taking local neighborhood topology into account can correctly treat cases where the population is not homogeneously distributed over space. However, interactions between individuals hardly follow such fixed contact patterns when there is a necessity to take more realistic populations into account.

In order to take into account more complex neighborhood, graph model have been proposed. Ghani [16] proposed a graph-based representation of the poultry production network in the U.K. and proposed a method to determine which nodes of the production network needed to be actively monitored when the epidemic started in some other node. Although network-based representations offer flexible ways to simulate the influence of topology, they are usually static and apply to homogeneous individuals. Changes in the interaction graph, like movements, for instance, cannot be easily taken into account. How to represent heterogeneity in behaviors? Taking relationships among individuals into account is a first step towards a more realistic representation of (human or animal) communities. Nevertheless, the previous models do not offer any help in representing the behavioral heterogeneity of the individuals. Yet, according to specialists, it can play a major role. For instance, vectors able of “long-distance translocations” [17] may transform a successful containment into a wide propagation and it is not possible to model this using previous models. Epidemiologists have then started to use ABMs in order to address this issue. In these models, it is possible to integrate all the information (global factors, heterogeneity of individuals, relationships) represented in the previous models as well as individualized behaviors. Each action of each agent can be designed according to its internal state and perceptions (neighbors and their parameters, global parameters, even environmental data, etc). An example is provided by [18], in which a SIR model is implemented in Netlogo. It simulates the transmission of a disease by individuals moving randomly on a grid. Though this model is simple, nothing prevents modelers from augmenting it, notably in terms of behaviors or environment, and using it to test hypotheses. More generally, ABMs allow epidemiologists to test

hypotheses related to the behaviors of the individuals that may be impossible to test on the field. These models are interesting, at least in our perspective, in that they are able to represent global parameters, individuals, their behaviors and their effect on the environment, their relationships, in combination with any of the environmental representation above. However, considering more detailed descriptions of the environment is currently beyond their scope and they may have difficulties in dealing with unstructured or dynamic ones.

How to study the impact of the Environment on the outcomes of an epidemic?

According to epidemiologists, the environment as a whole is often a key “actor” of an epidemic. At best, the previous models used to represent it simply (a grid) or only integrate its “global” impact. However, most environments are not homogenous, so speaking of global impacts can be inaccurate. Muller [19] proposed a first accounting of the heterogeneity of the environment by allowing the modeler to represent different “sites” in a simulation of a sleeping sickness epidemic. He then mimicked closely the situations observed, especially in the case of a low-level transmission where traditional approaches were simply not effective. To enhance the quality and realism of its model he then proposed to use a Geographical Information System (GIS) in place of what he called “agentified locations” (but no papers have been published on this). Such combinations of tools (ABM + GIS) allow epidemiologists to represent realistically the systems they are studying by having both the actors and the “background” modeled. It also enables them to “play” with parameters at multiple scales, which makes it promising for running virtual experiments on the inter-dynamics of individuals, populations and environments.

Badariotti proposed an implementation of this idea in [20] with a grid environment containing data extracted from a GIS. The author worked with epidemiologists studying the plague in Madagascar who needed to have a detailed representation of the environment. Indeed, epidemiologists known in details every part of the pathogen complex (where only the entities related to the epidemics are considered) but not their arrangement as a pathogen system (where their context is taken into account as well) [20] thus allowing them to explain this recurrent outbreaks situation. This situation is similar to our own though epidemiologists expressed the need to have field data directly as environment in place of GIS extracted data. Actually, in our case, the epidemiologists expressed the need to work with field data thus we will represent the environment using directly these GIS data.

3.2 Synthesis

Considering the requirements expressed by the epidemiologists in this research (individuals heterogeneity, global parameters, dynamic contact patterns, extensive representation of the environment with the same properties, and ability to conduct virtual experiments with the maximum of flexibility), the natural choice for us was to follow an agent-based approach, coupled with a detailed and flexible representation of the environment based on a combination of grids and GIS. In addition to this, we decided to make the environment (and its components) a first-class citizen of the model rather than just a topological surface, provided with its own attributes and

behaviors. We will use this formalism in section 5 to represent the environment, after we present, next section, the model of avian influenza we have built with epidemiologists.

4 Conceptual Model of Avian Influenza Propagation

In this section, we introduce the context of this research, followed by the representation of the environment and actors. All the choices made in accordance with epidemiologists and field specialists will be explained. As the model is likely to evolve in the future because of the addition of new data, we will explain how we have made provisions for such evolutions.

4.1 Frame of the Epidemiologic Study

The study takes place in North Vietnam, where epidemiologists focus on “local” mechanisms, i.e. mechanisms that occur at a scale comprised between the village and the district levels (around 50 km2). The model is then geographically limited by the bounds of a province (a few hundreds km2). The main geographical entity is the village, which is considered by epidemiologists as a “coherent epidemiologic unit”. Their hypothesis, present in the model, is that communes and districts are not really relevant to consider when it comes to study the local causes of propagation. This assumption may however be easily revised in future occurrences of the model. As the environment may be a reservoir for the virus we will consider that every place may allow virus survival. In addition every entities of the model can be also infective.

Fig. 1(Adapted) Vietnamese Village Transect from [21]

The “village” environmental unit

A « traditional » Vietnamese village consists of an inner-village space with a main street, a few dozen to more than a hundred households with some poultry (traditional farms) and pets. This inner space is surrounded by rice fields, watered lands (which enter the village), other cultures and is protected from the flooding by a dike. In our model, we describe the inner-village and the different zones with the structure presented on Fig. 1 and the data provided by the Vietnamese census and epidemiologists’ longitudinal surveys in order to define instances of villages.

The surrounding environment

The surroundings of the village consist in agricultural and “natural” lands (mountain, forest, etc.). As we do not focus on long distance propagation, we consider this distant space as homogeneous. Conversely, the inner-village and agricultural lands are represented in detail, especially the possibility that the latter may act as a reservoir for the virus. Though detailed data about environmental survivability of the virus are not yet available, epidemiologists consider that attributes like altitude, Ph, temperature, solar exposure, level of organic matter present in water, watered or not are needed to represent the virus survivability in the environment. These parameters define the ecological dynamics of the system. While the village’s ecological dynamic can be considered static all through the year, it is not the case for agricultural lands and especially wetlands (more dynamic regarding the propagation of viruse).

Fig. 2 Hierarchy of the different “types” of environment found in the model

4.2 Relevant Actors

Actors of the pathogen system act at different scales and are organized in 2 main structures: the village and the poultry production chain. The other levels of organization (like administrative levels) are neglected in the model.

Actors of the Traditional Village In the village important actors are organized around the farm, which contain at least a farmer and a poultry flock (chicken and/or duck). The type and dynamics of both are defined according to the production type and the production sector (see below). In addition, village markets are important in terms of contacts between poultries. Farm We can classify farms by their size and production techniques in four sectors according to [22]: • Traditional farming: mixed poultry, local scale interactions, no bio-security at all. • Semi-industrial & industrial: targeted production, district to province size scale

interactions, low/medium bio-security level (medium/high for industrial ones). • Fully integrated: targeted production, province size scale interactions, very high

bio-security level (thus, not considered in the model). The production type and sector determine most of the farm’s characteristics in term of herd (size, lots organization, vaccination coverage, species and breeds), the type of

premises (caged, fully confined, with an access to a pond, etc.), which impacts the bio-security level, etc. They also determine the dynamics of the farm and its acquaintances. Poultry According to experts, all the poultry within a flock are very similar in terms of behaviors (gregarious animals), and characteristics (homogeneous lots). So we can aggregate these individuals and represent only the poultry flock. This, plus homogeneous mixing occurring within the flock allow us to represent the virus transmission within the flock with a SIR model. A specific SIR will be developed for traditional farm’s flock (heterogeneous population). The flock behavior is also not very complex. Depending on the farm type and sector, the poultry can be confined, access an adjoining or distant secured zone (a pond, channel, etc), or be free ranged. Markets and transporters According to epidemiologists, markets can play a role of reservoir. A lot of contacts between living and processed poultries occur in them. Since farmers are not modeled explicitly, we still need to be able to represent the transportation process and we introduce a “traditional” transporter to do so. Humans and other animals At the village scale, many other actors may have been taken into account, but because of the specific focus of this epidemiologic study, most of them will not be included in this first model.

Humans, for instance, can play a role of mechanical vectors in both local and long-range propagations. However, the epidemiologists are not interested in the latter, and the effects of the former are considered as very low when compared to poultry flocks. Consequently, we neglect them in this first model. Wild birds, which are considered by most epidemiologic surveys as not playing a noteworthy role in local propagation, are not to be considered. Peridomestic birds, farm animals (pigs, buffaloes, etc.), pets (cats, dogs) and fighting cocks are also removed from this first model, as the data is insufficient on their role on avian influenza propagation.

4.2.2 Outside the Traditional Village In the avian influenza case, the village is a coherent epidemiological unit but is linked to poultry production networks, which have a much wider scale dynamic and a possible strong impact because they connect all the farms and may constitute a good propagation system [4]. Several networks co-exist, though the fully industrialized sector can be neglected (including the farm) as the bio-security level is very high. Conversely, industrial and semi-industrial sector production chains are tied together by farms sharing middlemen.

Mainly, we can consider ([4] and expert knowledge) four actors of this production chain: traders (highly variable in term of size), transporters, slaughterhouses and selling points. Traders can stock, trade among them and manage their transporters, who carry living or slaughtered poultry from a chain production node to another.

Slaughterhouses “just” process the poultry and thus have no relevant dynamic. The only point of sale of processed poultry (i.e. supermarket) can be neglected as it is considered as a disease end point (no infection possible). In contrary, local, district or province market need to be represented explicitly.

4.3 Conclusion

We have a fairly complex system with both detailed environment and numerous actors to represent but the general structure has been defined. The environment consists geographically of three main areas: inner village, agricultural land, natural land, “organizationally” of the province (our whole system), district & commune (which has no real interest in our concern) and the village, and conceptually in static and dynamic environments. The actors can be organized in two ways. Considering the “action scale”, we have entities acting within the village such as: poultry and other animals, transporter, farm, etc. and entities acting at a larger scale: farm, trader, transporter, and slaughterhouse.

As all needed data are not yet available at the creation process of the model, it will be incremental. Although, we need to take care to be able to represent any entities that may be declared relevant from surveys being conducted at the moment

5 From Model to Simulation

One of the main purposes of our model is to identify and test hypotheses about local propagation mechanisms and the reservoir role of environment for the avian influenza disease in North Vietnam.

As explained in section 3, we decided to use an ABM for two main reasons: first, the ability of this type of model to represent the system in a detailed way and more important in a incremental manner. Secondly, the possibility of running controlled experiments to test numerous hypotheses within the simulation.

Our model derives its data from existent epidemiological sources. Firstly, this means that we have to integrate several data sources which imply very different formalisms: formal data and informal data that can be reliable or not. Secondly, all this data is not available at the beginning of the modeling process--this means that we must plan the use and possible interactions with other data with the help of epidemiologists knowledge (informal data).

As the latter point is well addressed by modern simulation platforms we will focus on the first one. We have to be able to integrate together the data that can be collected. To do so we used GAML’s metamodel, which is defined within the GAMA platform [23], that is generic enough to our application and allow detailed and dynamic environment representation. Indeed, this metamodel divide the agent into a decision architecture representing behaviors and internal state while a “body” takes care of localization and perception issues.

Finally, we will need to be able to conduct “virtual experiments” though this issue is still under consideration.

5.2 Implementing the Conceptual Model

We will present here how we implement the conceptual model and the choices made to represent it correctly using GAML’s [23] metamodel. In contrary to many ABMs we will treat the environment as a very important and detailed entity of the model.

5.2.1 Representing the Environment

5.2.1.1 Outline Data As we do not want explicitly select relevant locations as in Muller [19] we use directly the GIS because it can be uneasy and we may not select every needed locations. Indeed, places like ecotones (a transition area between two adjacent ecosystems) are always difficult to model (according to [19]), but with our approach, if a place plays a special role--we will see it emerge.

This GIS data is not detailed enough, just areas and no individualization of buildings or lots, thus we will use the transect (of Fig.1) in order to have a detail enough representation of the village.

5.2.1.2 Dynamic Data Some of the static provided data should be dynamic (i.e. vary seasonally), for example when we have only the mean value of the temperature.

To address this issue we decided to generate a discreet environment from the GIS, a grid where each cell is typed and parameterized according to the GIS data and added to the GIS. This idea is similar to the GIS layers decomposition idea but extending it as we can pile up any kind of environment like this as much as needed. Here the grid makes sense as it much more computer efficient than any continuous space (like GIS).

In order to be able to synchronize different type of environments, GAMA defines every environment as an organized aggregation of spatial entities/locators which can have an intrinsic dynamic defined like reactive agents. In addition this dynamic can also impact external elements (agents localized at that place or neighbor cells). Intrinsic dynamic will be use to represent virus survivability according to in vitro experiments currently conducted in Cambodia (Pasteur Institute). For example, a river’s cell may propagate the virus to its neighbors following a stream.

5.2.1 Representing the Actors

5.2.2.1 Poultry In this first model we focus mainly on the poultry flock. As expressed in section 4 we can consider the flock as a coherent epidemiological unit and thus have a aggregated agent representing it. Actually, this agent we represent the day to day dynamic of the poultry (going to the field/pond or stay confined) according to the farm type. This dynamic will be represented using a task-based decision architecture. This architecture is similar to the Mae’s one [24] but introduced motivations to tasks. This allows epidemiologists to finely tune the behavior of the agents.

In addition poultry are able to excrete virus in the environment and to be infected from it while the internal infection mechanism is addressed by different SIR (depending on the poultry type) models under construction by epidemiologists.

5.2.2.2 Structure actors Poultry production networks node can be represent as non mobile agents that may be infected or not. Traders have a very simple dynamic as they manage poultry collection and distribution through transporters. These transporters are mobile agents that follow a collect plan then deliver living poultries to slaughterhouses then deliver slaughtered poultries to selling point. In particular, traditional farms transporter are limited to a farm and a few market within the vicinity (according to field experts). These selling points, excluding markets, are static element as they are a disease end point. Markets have their own model of virus survivability according to data under collection. Farms are more dynamic as they determine the poultries’ dynamic. Thus it will be represented by a non mobile agent with characteristics such as its type and sector which will define: bio-security level, day-to-day dynamic of the poultry, trade dynamic with other production nodes.

Finally, (para) veterinarians are mobile agents that check farms. They build a journey plan then they move from a farm to another. During these movements they made carry the disease from a farm to another.

5.2.2.3 Actors to be represent in following models Other actors such as humans, domestic animals (pets and breeding), peridomestic birds are to be considered in upcoming models. They may also carry the disease mechanically or by being infected. These will be represented using mobiles agents using characteristics and behaviors that are still under evaluation.

5.2 Validating such a model

Validation of models has always been a strong issue for modelers to address, especially in the case of IBMs and ABMs [24]. Although we focus on one main issue first: adequateness of the model to validate epidemiologists’ hypotheses. Many authors (i.e. [19] [20]) already justified the need of an explicit and detailed environment. In addition, we compared several theoretical models about the propagation of a disease that can be transmitted trough contacts and the environment within a homogeneous population. We found that if we use an heterogeneous and dynamic environment model, the expressivity of the model is larger. Indeed, even simple contacts or basic homogeneous environment models are able to represent slow and fast epidemics and endemics. Although environmental models are the only ones able to represent resurgence situations while dynamic environment allow more different disease dynamic (i.e. longer or shorter infected environment only period). Later on we intend to validate our mode as follow. Each representing unit during the modeling process by checking with field expert if their “no-interaction” behavior are

correct or not. Then we check if interactions among several individuals are correct. Finally, epidemiologists can check the model “realisticness” (i.e. in a qualitative way) by comparing its global dynamic to documented real epidemics situation.

5 Conclusion

We presented in this paper the evolution of epidemiological questions and the model used to addressed them. We moved from the field of mathematics, addressing global-scale questions, for computer-based models and especially IBM which are better fitted for small-scale problematic. Afterwards we presented a model to study small-scale epidemiological phenomenon in the context of avian influenza in North Vietnam: local propagation and re-emergence mechanisms. This model differs from previously proposed as it treats the environment as a first-class citizen of the system. Here the environment is not only a topological surface, it can contain heterogeneous data, have its own dynamic and can be multiple (GIS + grid in our application). Then we presented the implementation of this model using the GAML metamodel. Finally we presented the question of validating such ABM. To do so we studied the representation expressivity of the environment and presented some ways of globally validating our model.

References

1. Gilbert, M., Xiao, X., Pfeiffer, D.U., Epprecht, M., Boles, et al.: Mapping H5N1 highly pathogenic avian influenza risk in Southeast Asia, Proc Natl Acad Sci USA. 105, 4769-4774

2. Food and Agriculture Organisation Vietnam, http://www.fao.org.vn/ (28/06/08) 3. Dung Nguyen, T., Vinh Nguyen, T., Vijaykrishna, D., Webster, R.G., Guan, Y., Peiris,

J.S.M., Smith, G.J.D.: Multiple sublineages of influenza A virus (H5N1), Vietnam, 2005 2007, Emerging Infectious Diseases, 14, 4 (2008)

4. Agrifood consulting international: The Impact of Avian Influenza on Poultry Sector Restructuring and its Socio-economic Effects. (2006)

5. Dung Nguyen, T., Vinh Nguyen, T., Vijaykrishna, D., Webster, R.G., Guan, Y., Peiris, J.S.M., Smith, G.J.D.: Multiple sublineages of influenza A virus (H5N1), Vietnam, 2005 2007, Emerging Infectious Diseases, 14, 4 (2008)

6. Chen H., Li Y., Li Z., Shi J., Shinya K., Deng G. et al: Properties and dissemination of H5N1 Viruses isolated during an influenza outbreak in migratory waterfowl in Western China, J.of Viro. 80, 5976-5983 (2006)

7. Ito, T., Okazaki, K., Kawaoka, Y., Takada, A., Webster, R.G., Kida H.: Perpetuation of influenza A viruses in Alaskan waterfowl reservoirs. Arch. of Viro. 140(7):1163-72 (1995)

8. Stallknecht, D.E., Brown, J.D.: Wild birds and the epidemiology of avian influenza. Presentation at the FAO and OIE Conf on Avian Influenza, Rome, (2006)

9. René, M., Bicout, D.: Influenza aviaire : Modélisation du risque d’infection des oiseaux à partir d’étangs contaminés , Epidémiologie et santé animale, 51, 95 – 109 (2007)

10. McKendrick, A.G.: Applications of mathematics to medical problems, Proc. of the Edinburgh Mathematical Society, 44, 1-34 (1925)

11. Abbas, K., Mikler, A., Ramezani, A., Meneze, S.: Computational epidemiology: Bayesian disease surveillance, Proc. of ICBA'04 proceedings (2004)

12. Durand B., Mahul O.: An extended state-transition model for foot-and-mouth disease epidemics in France, Preventive Veterinary Medicine 47, 1-2(2000)

13. M. Artzrouni, et J.P. Gouteux, Population dynamics of sleeping sickness : A micro simulation, Simulation & Gaming, Vol. 32, No. 2, 215-227 (2001)

14. Brouwers ,L.: MicroPox: a Large-scale and Spatially Explicit Microsimulation Model for Smallpox Transmission, Proc. of the Intl. Conf. of Health Sciences Simulation (2005)

15. Turner, J., Begon, M., Bowers, R.: Modelling pathogen transmission: the interrelationship between local and global approaches. Proc Biol Sci.,270(1510): 105–112 (2003)

16. Ghani, A.: Models for an Outbreak of Avian Influenza in the UK Poultry Flock, Presentation at Journées de modélisation en épidémiologie animale du CIRAD (2007)

17. Smith, D.L., Waller, L.A., Russell, C.A., Childs, J.E., Real, L.A.: Assessing the role of long-distance translocation and spatial heterogeneity in the raccoon rabies epidemic in Connecticut, Prev. Vet. Med., 71, 3-4, 225-240 (2005)

18. Wilensky, U.: NetLogo Virus model (http://ccl.northwestern.edu/netlogo/models/Virus) (1998)

19. Muller, G., Grébaut, P., Gouteux, J.P.: An agent-based model of sleeping sickness: simulation trials of a forest focus in southern Cameroon, Éd. Sc. et Méd. Elsevier (2004)

20. Badariotti, D., Banos, A., Laperrière, V. : Vers une approche individu-centrée pour modéliser et simuler l’expression spatiale d’une maladie transmissible : la peste à Madagascar, cybergeo (http://www.cybergeo.eu/index9052.html) (2007)

21. Bénédicte, S.J.: Atlas Bac Hung Hai, Gret (personnal communication) (1999) 22. Desvaux S. Vu D-T: A general review and description of the poultry production in

Vietnam, Agricultural publishing house, 2008 23. Amouroux, E., Chu, T.Q., Boucher, A., Drogoul, A.: GAMA: an environment for

implementing and running spatially explicit multi-agent simulations, LNAI, Springer (2007) 24. Maes, P.: A bottom-up mechanism for behavior selection in an artificial creature. Proc. of

the first conf. “animals to animats”, 238 - 246 (1991) 25. Moss, S.: Alternative approaches to the empirical validation of agent-based models, JASS,

11, 1-5 (2008)