African Swine Fever: Computational Models for Early Detection and Spread Prediction in Wild Boar Populations

Introduction

African swine fever (ASF) is a highly contagious, hemorrhagic viral disease of domestic pigs and wild boar caused by the African swine fever virus (ASFV), a large, enveloped DNA virus of the family Asfarviridae. Since the re-introduction of ASFV genotype II into Georgia in 2007, the virus has spread across Eastern Europe, the Baltic region, and into Western Europe, with significant incursions into Italy, Germany, and other member states [1, 2]. The epidemiology of ASF in Europe is characterized by a sylvatic cycle involving wild boar (Sus scrofa) populations, which act as a reservoir and source of spillover to domestic pig holdings [3, 1]. The absence of an effective vaccine or treatment necessitates that control strategies rely on early detection, rapid culling, carcass removal, and biosecurity measures [4]. Computational models have become indispensable tools for understanding transmission dynamics, predicting spatial spread, and optimizing surveillance and intervention strategies in wild boar populations. This review examines the current state of computational approaches for early detection and spread prediction of ASF in wild boar, focusing on artificial intelligence (AI) based detection tools, spatial statistical models, and mechanistic transmission frameworks.

Early Detection Models

Early detection of ASF in wild boar is challenging due to the cryptic nature of the disease in free-ranging populations and the often delayed appearance of clinical signs. Several computational approaches have been developed to enhance detection sensitivity and reduce time to identification.

AI-Based Detection from Camera and Sensor Data

The use of surveillance cameras and health monitoring collars represents a non-invasive approach to early detection. Ryu and Tai [5] developed a high-performance AI-based 3D depth camera system for object detection and tracking in pig pens. The system employed two AI models, a fast model and an accurate model, to generate 3D bounding boxes, identification numbers, and distance measurements for individual animals. The accurate detection model demonstrated superior performance for 3D object tracking, suggesting potential for application in detecting behavioral changes associated with early ASF infection. The study emphasized the need for custom datasets trained on farm-specific conditions to achieve optimal performance.

Layton et al. [6] developed a non-AI preliminary algorithm using data from health monitoring collars that collected pulse rate, respiratory rate, and heart rate variability in research pigs. Following experimental challenge with highly pathogenic ASFV, the collar monitors detected decreased mean pulse rate and increased variability in pulse rate and heart rate variability, changes not captured by single daily point-in-time measurements. The algorithm detected disease in 100% of infected pigs and predicted disease onset in 67% of cases, with abnormal readings occurring prior to the onset of clinical disease. This approach provides a framework for developing early warning systems based on continuous physiological monitoring.

AI-Enhanced Diagnostic Assay Interpretation

Lateral flow assays (LFAs) are commonly used for field diagnosis of ASF, but their interpretation can be subjective, particularly in low-resource settings. Bakshi et al. [7] developed a deep learning assisted, smartphone-based AI diagnostic tool using You Only Look Once (YOLO) models for image classification of LFA test strips. The models achieved an average accuracy of 86.3%, precision of 96.3%, recall of 79%, and F1 score of 0.87 across multiple dataset splits. The tool was integrated into a prototype JavaScript web application for ASF reporting and visualization, with positive predictions displayed on a geographic map. This approach enhances the sensitivity and objectivity of LFA reading, facilitating rapid and accurate field diagnosis.

Mortality-Based Surveillance Algorithms

Monitoring mortality patterns is a practical approach for early detection in both domestic and wild populations. Faverjon et al. [8] compared the effectiveness of mortality thresholds at different epidemiological unit levels (pen, room, barn) for early detection of ASF in large commercial pig farms. Using a within-barn spread model that incorporated non-homogeneous probabilities of transmission within pens, between pens, and between rooms, the study found that room- or pen-based mortality thresholds provided detection within 8 days of disease introduction. Barn-level thresholds achieved similar detection performance but required testing a larger number of pigs. The study demonstrated that baseline mortality rate and pen size had limited impact on pen-level mortality thresholds, supporting the generalizability of this approach.

Environmental DNA Surveillance

Environmental DNA (eDNA) sampling offers a non-invasive method for detecting ASFV in the environment. Varzandi et al. [9] designed and evaluated an eDNA sampling method for highly turbid water and soil samples, using qPCR to detect both ASFV and wild boar DNA. The method was validated using samples from an ASF-free area spiked with synthetic ASFV DNA template. This approach has potential for early detection of ASFV in areas of recent viral introduction or ongoing outbreaks, complementing traditional carcass-based surveillance.

Spatial and Spatio-Temporal Models for Spread Prediction

Understanding the spatial dynamics of ASF in wild boar populations is critical for predicting spread and targeting control measures. A range of spatial statistical and mechanistic models have been applied to this problem.

Space-Time Clustering and Risk Factor Analysis

Allepuz et al. [10] analyzed the temporal and spatial distribution of ASF-positive wild boar carcasses from 2017 to 2021 across ten European countries. The space-time K-function revealed significant clustering within distances of 2 km and within 1 week. A Bayesian hierarchical spatial model identified landscape factors associated with higher probability of finding ASF-positive carcasses, including transition zones between woodland and shrub, green urban areas, and mixed forests. The presence of paths and higher wild boar abundance also increased detection odds. These findings inform targeted search strategies for early detection.

Smolko et al. [11] analyzed six years of national surveillance data from Slovakia (2019-2024), estimating temporal variation in the effective reproduction number (Rt) and modeling spatio-temporal prevalence. Passive surveillance (found dead) showed greater diagnostic sensitivity than active surveillance (hunted) for case detection (PCR: 46.5% vs. 0.48%). Rt peaked at 3.83 in March 2021 and periodically exceeded 1.0 through late 2024. Virological prevalence exhibited strong late-winter/early-spring seasonality and a persistent east-to-west gradient. Wild boar density decreased by 36.3% over the study period, attributed to disease-related mortality and intensified hunting.

Habitat Suitability and Landscape Connectivity

Species distribution models and landscape connectivity analysis have been used to predict areas at high risk for ASF outbreaks. Faustini et al. [12] mapped suitable habitats for wild boar and their potential dispersal corridors in Northern Italy. The distribution of ASF-positive wild boar along major corridors predicted by the model validated the approach, supporting its use for surveillance and carcass early detection.

Choi et al. [13] used the MaxEnt model and shortest-path betweenness centrality analysis to predict suitable areas and geographical paths for ASF outbreaks in wild boar in South Korea. High-risk zones were defined as areas with suitability values of 0.4 or higher and areas within 1.8 km of predicted paths. The analysis identified 37 pig farms in high-risk zones on the suitability map and 499 farms on the shortest-path map, enabling targeted biosecurity measures.

De Petris et al. [14] applied GIS-based spatial analysis to monitor ASF outbreaks in northwestern Italy, identifying central-western municipalities as higher risk due to dense wild boar populations and favorable environmental conditions. The study developed a spatial risk model using geostatistical techniques, demonstrating the effectiveness of geospatial modeling for identifying high-risk zones and supporting targeted surveillance.

Directional Spread Estimation

Gervasi et al. [15] developed a simple regression-based method to estimate the directional speed of ASF spread using limited data. Applied to the ASF outbreak in northwestern Italy (2021), the model estimated average spreading speeds ranging from 33 to 90 m/day in different directions. The estimated infected area was 2216 km2, approximately 80% larger than the area identified through field-collected carcasses alone. The model also estimated that the actual initial date of the outbreak was 145 days earlier than the first notification, highlighting the importance of inferential tools for early outbreak characterization.

Mechanistic Transmission Models

Mechanistic models simulate the underlying biological processes of transmission and can incorporate host ecology, landscape features, and intervention strategies.

Density-Dependent and Spatially Explicit Models

Hayes et al. [16] developed a spatially explicit detection-delay SIR mechanistic model of ASF transmission among wild boar in northern Italy, parameterized using approximate Bayesian computation. The model incorporated static and seasonal transmission rates and linear relationships between habitat susceptibility/infectivity and wild boar density. The best-fitting model used a seasonal transmission rate but did not support a wild boar density effect across the entire study period. However, further analysis suggested that density likely played a role during the second wave (October 2022 to September 2023), possibly due to differences in surveillance rates or changes in density distributions.

Han et al. [17] developed a novel ASF transmission model for wild boar in South Korea, estimating that roads and rivers reduced transmission rates to approximately 37% on average. Fence-lines provided only limited reduction, suggesting they should be considered temporary measures. The probability of transmission to adjacent habitats decreased considerably with increasing distance, supporting the slow spatial transmission speed observed in European studies.

Multi-Host Models

Hayes et al. [3] constructed a spatially explicit stochastic mechanistic model of ASF transmission between domestic pigs and wild boar in Romania. The model estimated that 69.4% of domestic pig herd cases originated from other infected domestic pig herds, 20.4% from infected wild boar sources, and 8.4% from external sources. For wild boar, 31.9% of infections originated from domestic pig herds and 68.1% from neighboring infected wild boar populations. Habitats with forest coverage greater than 15% were 2.6 times more infectious and 5.3 times more susceptible than other habitats. Alternative control scenarios, including reactive culling of entire villages, improved epidemic outcomes.

Scenario Tree Models for Freedom from Disease

Christensen and Duizer [18] developed a scenario tree model (STM) to support claims of freedom from ASFV in commercial swine in Western Canada. The model used data from the Canada West Swine Health Intelligence Network (CWSHIN) and the CanSpotASF surveillance program. The probability that a herd veterinarian would report a suspicion of ASF to authorities (seVet) was identified as a critical parameter affecting model outcomes. When seVet was low (0.01), additional pathology examination and rule-out testing improved evidence of freedom. When seVet was high (0.7 or higher), additional testing provided no benefit.

Event-Based Surveillance Models

Boudoua et al. [19] evaluated the EpiDCA model, an unsupervised event-based surveillance system that integrates epidemiological and environmental data for early detection of disease outbreaks. Applied to ASF in Europe, the model achieved weighted F-scores between 0.64 and 0.85. Sensitivity analysis demonstrated model robustness, with incorporation of environmental data and finer spatial granularity significantly improving classification precision.

Integrated Framework for Computational Surveillance

The integration of multiple computational approaches can create a comprehensive surveillance and response framework. The following diagram illustrates a decision tree for deploying computational models in ASF management in wild boar populations.

graph TD
    A[ASF Surveillance in Wild Boar], > B{Detection Method}
    B, > C[Passive Surveillance]
    B, > D[Active Surveillance]
    B, > E[Environmental Surveillance]
    
    C, > F[Carcass Detection]
    F, > G[AI-Enhanced LFA Diagnosis]
    G, > H[Case Confirmation]
    
    D, > I[Camera Traps / Collars]
    I, > J[AI Behavior Analysis]
    J, > K[Early Warning Signal]
    
    E, > L[Soil/Water eDNA Sampling]
    L, > M[qPCR Detection]
    M, > N[Viral Presence Confirmed]
    
    H, > O{Spatial Analysis}
    K, > O
    N, > O
    
    O, > P[Space-Time Clustering]
    O, > Q[Habitat Suitability Modeling]
    O, > R[Landscape Connectivity Analysis]
    
    P, > S[Risk Zone Identification]
    Q, > S
    R, > S
    
    S, > T[Mechanistic Transmission Model]
    T, > U[Spread Prediction]
    T, > V[Intervention Optimization]
    
    U, > W[Targeted Carcass Search]
    V, > W
    W, > X[Control Implementation]
    X, > Y[Monitoring and Model Updating]
    Y, > O

Discussion

The computational models reviewed here represent a diverse toolkit for addressing the challenge of ASF in wild boar populations. AI-based detection methods, including camera systems, health monitoring collars, and smartphone-based LFA interpretation, offer the potential for earlier detection than traditional surveillance alone. However, these methods require validation under field conditions and may be limited by infrastructure requirements and cost.

Spatial statistical models, including Bayesian hierarchical models, MaxEnt, and landscape connectivity analysis, provide valuable insights into risk factors and high-risk areas. These models are particularly useful for targeting surveillance resources and designing search strategies for infected carcasses. The consistent finding that passive surveillance (found dead) has higher diagnostic sensitivity than active surveillance (hunted) underscores the importance of prioritizing carcass search efforts [11, 1].

Mechanistic transmission models, ranging from simple SIR frameworks to complex spatially explicit multi-host models, enable prediction of spread dynamics and evaluation of intervention strategies. The use of approximate Bayesian computation for parameter estimation allows these models to be fitted to observed epidemic data, improving their predictive accuracy [16, 3]. The identification of key parameters such as wild boar density, landscape features, and reporting probabilities provides actionable insights for control programs.

Several limitations should be acknowledged. Many models rely on assumptions about wild boar ecology and behavior that may not hold across different regions and ecosystems. The quality and completeness of surveillance data, particularly for passive surveillance, can introduce bias. The detection sensitivity of carcass searches is imperfect, and models must account for this to avoid underestimating the extent of infection [20]. Additionally, the role of human-mediated spread, including through contaminated fomites and transport of infected animals, is difficult to quantify and may not be adequately captured by models focused on wild boar ecology.

Future Directions

Future research should focus on integrating multiple data streams, including remote sensing data, animal movement data, and genomic epidemiology, into unified modeling frameworks. The development of real-time or near-real-time surveillance systems that combine AI-based detection with spatial modeling and mechanistic prediction would enable rapid response to emerging outbreaks. The application of machine learning methods for parameter estimation and model selection in mechanistic models may improve predictive accuracy. Finally, cross-border collaboration and data sharing are essential for modeling ASF spread in wild boar populations that do not respect national boundaries.

References

[1] More, S., Miranda, M., Bicout, D., et al. (2018). African swine fever in wild boar. EFSA Journal. European Food Safety Authority. https://www.semanticscholar.org/paper/5df51d201b07712558a568bcc4e37811bfb1cff4

[2] Nurmoja, I., Schulz, K., Staubach, C., et al. (2017). Development of African swine fever epidemic among wild boar in Estonia - two different areas in the epidemiological focus. Scientific Reports. https://www.semanticscholar.org/paper/a1cad1c1fdd2971ac1983a244b65ee72245e9c79

[3] Hayes, B., Vergne, T., Rose, N., et al. (2024). A multi-host mechanistic model of African swine fever emergence and control in Romania. bioRxiv. https://www.semanticscholar.org/paper/f94cd8d56995a86a24d772ec926aaf621bc86f5c

[4] Barongo, M. B., Bishop, R., Fèvre, E., et al. (2016). A Mathematical Model that Simulates Control Options for African Swine Fever Virus (ASFV). PLoS ONE. https://www.semanticscholar.org/paper/56bd27ececdcfa73d9dadb7809813638845a416b

[5] Ryu, H. W., & Tai, J. (2021). Object detection and tracking using a high-performance artificial intelligence-based 3D depth camera: towards early detection of African swine fever. Journal of Veterinary Sciences. https://www.semanticscholar.org/paper/53c9587db3c50ebfc762918ff15365347ac1652a

[6] Layton, R., Beggs, D., Mansell, P., et al. (2026). A non-AI preliminary algorithm for the prediction and detection of highly pathogenic African swine fever in pigs using health monitoring collars. Animal Welfare. https://www.semanticscholar.org/paper/7743a626febd9acbb8b782611a81d7903c30e4cd

[7] Bakshi, A., Stetson, J., Caragea, D., et al. (2025). Toward a rapid, sensitive, user-friendly, field-deployable artificial intelligence tool for enhancing African swine fever diagnosis and reporting. American Journal of Veterinary Research. https://www.semanticscholar.org/paper/98c1ab5b7a8ed97878fa2646dbf698279f375c96

[8] Faverjon, C., Meyer, A., Howden, K. J., et al. (2020). Risk-based early detection system of African Swine Fever using mortality thresholds. Transboundary and Emerging Diseases. https://www.semanticscholar.org/paper/631690e376d728885d9d4ed569af6f1b04307fb8

[9] Varzandi, A. R., Zanet, S., Barroso Seano, P., et al. (2023). Detection of African swine fever virus and wild boar eDNA in soil and turbid water samples: towards environmental surveillance. Zeitschrift für Jagdwissenschaft. https://www.semanticscholar.org/paper/caab8e32c89f7748558c68eddb0a1878246e7159

[10] Allepuz, A., Hóvári, M., Masiulis, M., et al. (2022). Targeting the search of African swine fever‐infected wild boar carcasses: A tool for early detection. Transboundary and Emerging Diseases. https://www.semanticscholar.org/paper/4749da2c47f421c979f7b5752bc378a4bd7dbcbc

[11] Smolko, P., Bučko, J., Štefanec, M., et al. (2025). Spatio-Temporal Dynamics of African Swine Fever in Free-Ranging Wild Boar (Sus scrofa): Insights from Six Years of Surveillance and Control in Slovakia. Veterinary Sciences. https://www.semanticscholar.org/paper/f9ada93b6ca88e214c96ab2aacdb45d3002750ec

[12] Faustini, G., Soret, M.-G., Defossez, A., et al. (2025). Habitat suitability mapping and landscape connectivity analysis to predict African swine fever spread in wild boar populations: A focus on Northern Italy. PLoS ONE. https://www.semanticscholar.org/paper/59715814ea6ad721f002b614af766942ca31ea76

[13] Choi, J. H., Namgung, H., Lim, S., et al. (2023). Predicting Suitable Areas for African Swine Fever Outbreaks in Wild Boars in South Korea and Their Implications for Managing High-Risk Pig Farms. Animals. https://www.semanticscholar.org/paper/fda6ee739596199fc2a854389545434cea681005

[14] De Petris, S., Orusa, T., Viani, A., et al. (2025). Spatio-Temporal Pattern Analysis of African Swine Fever Spreading in Northwestern Italy, The Role of Habitat Interfaces. Animals. https://www.semanticscholar.org/paper/64b4e9dabcd2e2414350afb910af5e0f77c31fd0

[15] Gervasi, V., Sordilli, M., Loi, F., et al. (2023). Estimating the Directional Spread of Epidemics in Their Early Stages Using a Simple Regression Approach: A Study on African Swine Fever in Northern Italy. Pathogens. https://www.semanticscholar.org/paper/c144df98320f6c85f5591ff97bc227497349a335

[16] Hayes, B., Lim, J. S., Andraud, M., et al. (2024). Elucidating the influence of wild boar density on African swine fever spread in wild boar populations, Italy, 2022–2023. EFSA Supporting Publications. https://www.semanticscholar.org/paper/8727164851a9524d9d7d94846eff889c0baad1bd

[17] Han, J.-H., Yoo, D.-s., Pak, S., et al. (2021). Understanding the transmission of African swine fever in wild boars of South Korea: A simulation study for parameter estimation. Transboundary and Emerging Diseases. https://www.semanticscholar.org/paper/236c0e97a3315196000538199f40fe831af7482a

[18] Christensen, J., & Duizer, G. (2025). A scenario model to support freedom from African swine fever virus in Western Canada populated with data from Canada West Swine Health Intelligence Network and CanSpotASF. Preventive Veterinary Medicine. https://www.semanticscholar.org/paper/889dfc5cd8c014ebd266d41f267f8506a4935685

[19] Boudoua, B., Roche, M., Teisseire, M., et al. (2026). Spatio-temporal evaluation of EpiDCA: Case studies on avian influenza, African swine fever, and West Nile virus disease. Spatial and Spatio-temporal Epidemiology. https://www.semanticscholar.org/paper/685086693a70a306149b6b1c6b6bc178dbab985b

[20] Lim, J.-s., Vergne, T., Pak, S., et al. (2021). Modelling the Spatial Distribution of ASF-Positive Wild Boar Carcasses in South Korea Using 2019–2020 National Surveillance Data. Animals. https://www.semanticscholar.org/paper/deaacc10f112091f312c014cc944002e92fe6103

[21] Jarynowski, A., & Belik, V. (2019). Modeling the ASF (African Swine Fever) spread till summer 2017 and risk assessment for Poland. Journal. https://www.semanticscholar.org/paper/3a79de252bbb11e96c3076fa54d8379a4370f38b

[22] Pepin, K., Brown, V., Yang, A., et al. (2022). Optimizing response to an introduction of African swine fever in wild pigs. Transboundary and Emerging Diseases. https://www.semanticscholar.org/paper/b9a098f0ba96e54f73525c8e20679115d1d9c1c7

[23] Herrera-Ibatá, D. M., Martínez-López, B., Quijada, D., et al. (2017). Quantitative approach for the risk assessment of African swine fever and Classical swine fever introduction into the United States through legal imports of pigs and swine products. PLoS ONE. https://www.semanticscholar.org/paper/f372926dbcdd9d0f47126ead553187c256fd49cb

[24] Nigsch, A., Costard, S., Jones, B., et al. (2013). Stochastic spatio-temporal modelling of African swine fever spread in the European Union during the high risk period. Preventive Veterinary Medicine. https://www.semanticscholar.org/paper/38bdece40ba340e64a2e8435ba4400a0781d31d0