key: cord-0291360-ynheirqv
authors: Picault, Sébastien; Vergne, Timothée; Mancini, Matthieu; Bareille, Servane; Ezanno, Pauline
title: The African swine fever modelling challenge: objectives, model description and synthetic data generation
date: 2021-12-21
journal: bioRxiv
DOI: 10.1101/2021.12.20.473417
sha: f24e8ae673ab1d05abd4723655afcb26b9b8c154
doc_id: 291360
cord_uid: ynheirqv

African swine fever (ASF) is an emerging disease currently spreading at the interface between wild boar and pig farms in Europe and Asia. Current disease control regulations, which involve massive culling with significant economic and animal welfare costs, need to be improved. Modelling enables relevant control measures to be explored, but conducting the exercise during an epidemic is extremely difficult. Modelling challenges enhance modellers’ ability to provide timely advice to policy makers, improve their readiness when facing emerging threats, and promote international collaborations. The ASF-Challenge, which ran between August 2020 and January 2021, was the first modelling challenge in animal health. In this paper, we describe the objectives and rules of the challenge. We then demonstrate the mechanistic multi-host model that was used to mimic as accurately as possible an ASF-like epidemic, provide a detailed explanation of the surveillance and intervention strategies that generated the synthetic data, and describe the different management strategies that were assessed by the competing modelling teams. We then outline the different technical steps of the challenge as well as its environment. Finally, we synthesize the lessons we learnt along the way to guide future modelling challenges in animal health. Highlights The 1st modelling challenge in animal health mimics ASF spread with synthetic data A mechanistic spatially explicit stochastic model was developed to represent ASF spread and control Synthetic data concerned date and location of detected cases in pigs and wild boar The wildlife-livestock interface is crucial in infectious animal disease spread

To raise livestock in a manner that is sustainable and respects animal welfare, animal health 40 must be managed, including infectious diseases which spread extensively between animal outputs and reflect on the lessons learnt. These are presented in the last article from this special 131 issue (Ezanno et al. submitted). 132 The data that were provided to the challenge participants were generated by a detailed agent-133 based model that was fed population data (spatial distribution of the host populations, 134 movements of live pigs, etc.) and parametrized with various key parameters defining 135 transmission processes as well as intervention strategies. This will be described in detail in 136 subsequent paragraphs. To be useful for the ASF-Challenge, we aimed to generate ASF-like were simulated based on land use data. 158 To simulate wild boar distribution, we obtained the hunting bags at the department level from 159 the Office Français de la Biodiversité, and assumed that a hunting season would reduce the 160 wild boar population by half, so that the total wild boar population size was twice the hunting 161 bag size at the level of the whole island. The location of individual wild boars was then 162 randomly simulated for each of the 500,366 wild boar, assuming that 18%, 80% and 2% of wild 163 boar would be in agricultural, forest and urban areas, respectively. Their geographical 164 coordinates were assumed to represent the centre of their home range. The spatial distribution 165 of individual wild boar was then summarized as department-level hunting bags, assuming again 166 that hunting bag sizes represented only half the size of the wild boar population in each 167 department.

To simulate a regionalization and a spatially heterogeneous distribution of the 4,775 pig farms, 169 we assumed that 1) 33% of farms would be in the Auvergne-Rhone-Alpes region while the 170 remaining 67% would be in the Occitanie region, and 2) 85%, 10% and 5% of pig farms would 171 be in agricultural, forest and urban areas, respectively. Based on these specificities, we 172 simulated geographical coordinates for the 4,775 pig farms that were registered in the two 173 French regions. Different farm characteristics, including whether they were commercial or 174 backyard, breeder, finisher or breeder-finisher, whether the pigs had access to an outdoor area, farm capacity and the regrouping of different farms into the same pig company were also 176 simulated successively. All of this information was used to generate a biosecurity score which 177 was used to vary a farm's susceptibility to, and transmissibility of, ASF (see below). Live-pig 178 movement data between farms also were simulated over time, assuming that breeder and 179 breeder-finisher farms could send pigs to finisher or breeder-finisher farms, that a breeder-180 finisher farm was more likely to fatten its own pigs than to send them to another farm, that 181 farms belonging to a pig company were more likely to send pigs to farms from the same 182 company than to farms not belonging to the same company, and that the total number of pigs 183 in a farm at a given time could not exceed the size of the farm. To produce these synthetic data, we needed a very detailed model that was expected to be more 186 complex than any model used by the ASF-Challenge teams. Hence, we designed a stochastic All epidemiological units (pig herds and individual wild boar) were subject to the same 208 infectious process, with the following states for animals in the units: susceptible (S), exposed 209 (E) where animals started being infectious but were still asymptomatic, fully infectious and 210 symptomatic (I). All infected animals eventually died, producing an infectious carcass (C). We 211 also assumed that wild boar were subject to natural mortality, giving either a healthy (D) or 212 infectious (C) carcass depending on their health state at death. The durations in states E, I, C, 213 D were distributed exponentially. In pig herds, carcasses were removed the next day, whereas 214 infectious (C) or healthy (D) wild boar carcasses could remain in the environment for several 215 weeks or months until they naturally decomposed or were removed when found by a passer-by.

Within pig herds, we assumed a frequency-dependent force of infection, exposed individuals 217 contributing to half the level of infectious animals or infectious carcasses. We also assumed a (Table S1 ). We (Table S2) . (Table S3 ).

The model incorporated all current European regulatory measures and triggered them as soon 252 as a primary case was detected in the simulation. All animals in detected pig farms were culled, 253 and the delay between detection and culling was fixed. A protection zone and a surveillance 254 zone were established (Table S6) , both subject to a trade ban and increased vigilance (increased 255 detection probabilities and biosecurity scores). Farms which had exchanged animals with a 256 detected farm during the previous three weeks (called "traced farms" in what follows) also were 257 subjected to the same restrictions and vigilance as in the protection zone, with the same 258 duration. Culled farms could be repopulated after a fixed period of time (Table S6) .

Infected wild boar carcasses found in the environment were removed immediately and led to 260 an active search for other infected carcasses (Table S6) . During this active search, each wild 261 boar carcass (either infected or disease-free) could be found with an increased probability 262 compared to carcasses outside the search zone. New infected carcasses triggered themselves 263 new active search operations (assuming no logistic constraints). After the detection of the 264 primary case, it was also assumed that part of hunted wild boar were tested (Table S6) . 265 In addition to regulatory measures, alternative interventions were triggered in the simulation 266 (Table S7) . First, as the forest near the primary case could be considered a main threat for virus 267 diffusion, 300 km of fences were installed 60 days after the detection of the primary case. The hunting pressure was increased within the fenced area as well as in a buffer area outside the 269 fences, aiming to remove 90% of wild boar by the end of the hunting period (instead of 50% in 270 other areas). In all areas with increased hunting, all hunted animals were tested, the active search 271 for wild boar carcasses was suspended, and the probability of finding wild boar carcasses by 272 passive surveillance was much higher than before. The increased hunting effort in the buffer 273 area occurred during two months. Second, 90 days after the detection of the primary case and 274 until the end of the simulation, when an infected wild boar was found (either through hunting 275 or as a carcass), all animals from nearby farms (Table S7) The virus was initially introduced shortly before the beginning of the hunting season (Table   280 S4), through an exposed wild boar located near a forest close to the centre of the hypothetical The maximal geographical extent of the epidemics was also highly variable among repetitions, 329 both in the wild boar and pig herd populations (Fig. 3) . The seven repetitions satisfying all 330 selection criteria showed a low spatial spread in the wild boar population (Fig. 3B) , very close 331 to the surface of the fenced area. In contrast, they showed a much more variable spatial spread 332 in pig herds (Fig. 3A) , in relation with commercial movements of infected animals before 333 source herds were detected. These long-distance spreading events in pig herds did not impact 334 the spatial spread of the disease in wild boar, highlighting a low exposure of wild boar to 335 infectious pig herds. Despite a very large maximal geographical extent of the epidemics, the cases were highly 343 aggregated in the selected repetition, both for pig herds and wild boar (Fig. 4) . The local 344 probability of infection was very high within the fenced area and on its western limit (where 345 the primary case was introduced). It was high south of the fenced area, and low but not nil north 346 and east of that area. It was nil everywhere else, except where a few pig herds has been infected 347 in some of the stochastic model repetitions. 355 Before the ASF-Challenge started, the ASF-Challenge teams were provided with the available 356 population data, i.e., the hunting bag size per department, as well as the farm database that 357 comprised the location and characteristics of the pig farms (except for their biosecurity score).

To allow the teams to become familiar with the data format and start developing their analytical Overall, the synthetic data provided to the ASF-Challenge teams relied on a few detected cases 375 in pig herds, spread over the challenge period, and a massive observed epidemic in wild boar.

The number of detected cases highly correlated with the number of actual cases (Fig. 2) , but Movie S1) for a dynamic view of the synthetic epidemic. 

With each of the three data releases, narrative situation reports were provided to the teams.

These reports aimed to contextualize the development of the epidemic during the period that Days (after first detection)

is included in the supplementary information (SI7).

At the end of phase 1, i.e., 50 days after the first detection, participants were asked to 1) predict 405 the number and location of wild boar cases and outbreaks in farms that should be expected "looked like" an ASF epidemic by considering the right populations (domestic pigs and wild 473 boar) and appropriate transmission processes within and between the two populations. We

acknowledge that the calibrated model that was used to generate the synthetic data is only one 475 of a virtually infinite set of models that could have generated ASF-like epidemic trajectories. 476 We note that in our simulated environment, the virus was detected much more frequently in ). Yet while it is important that the contextual situation represents a credible scenario 480 inspired by real situations, there is no need to aim for a "perfectly calibrated" epidemic that 481 likely does not exist.

To make the modelling challenge interesting and useful, the model that generated the synthetic 483 data necessarily needed to be more detailed than the different models that were to be used to 

Epidemic disease in England

Mechanistic modelling 598 of African swine fever: A systematic review

Wrong but useful -what Covid-19 epidemiologic models can 601 and cannot tell us

An open challenge to advance probabilistic forecasting for dengue 615 epidemics

Management of wild boar populations in the European Union before and during the 619 ASF crisis

Models of foot-and-mouth disease

Alternative control strategies against ASF in wild boar populations

Six challenges in modelling for public health 627 policy

A systematic review 629 of studies on forecasting the dynamics of influenza outbreaks

Respiratory Viruses

A collaborative multiyear, multimodel assessment of 644 seasonal influenza forecasting in the United States

A proposal for standardized evaluation of 647

African Swine Fever Virus Isolate, Georgia

Globalization 653 and infectious diseases: a review of the linkages. World Health Organization Tech

African swine fever: Disease dynamics in wild boar 657 experimentally infected with ASFV isolates belonging to genotype I and II

African swine fever 661 in wild boar in Europe -A Review

Epidemic and intervention modelling -a 663 scientific rationale for policy decisions? Lessons from the 2009 influenza pandemic

Pig empire under infectious threat: risk of African swine fever 667 introduction into the People's Republic of China

How climate change and wildlife 670 management affect population structure in wild boars

The RAPIDD ebola forecasting challenge: Synthesis 674 and lessons learnt

The future of influenza forecasts

Diseases at the Wildlife -Livestock 678 Interface: Research and Perspectives in a Changing World