Bayesian Inference for Graph-Based Markov Chain Model of African Swine Fever Transmission

African swine fever (ASF) remains a major challenge for veterinary surveillance because of its complex spatial spread across neighbouring regions. We present a Bayesian susceptible–infected–susceptible (SIS) model on a graph in which European NUTS2 regions are the nodes and transmission from a reported region j to a susceptible region i decays with geographic distance between their centroids, with strength pj→i∝dji−α1 and a small spontaneous-infection term. Transitions to a state with no reported ASF outbreaks are modelled using a logit-link function that combines regional case counts with continuing infection pressure from nearby infected regions. Bayesian inference is performed using stochastic variational inference (VI) and validated against Hamiltonian Monte Carlo using the No-U-Turn Sampler (NUTS). The model was evaluated on twelve hold-out months and compared with a persistence baseline. The proposed model achieved a lower overall log-loss than the persistence baseline (0.099 for VI and 0.095 for MCMC versus 0.127), with the largest gains on newly infected regions (information gain of +2.51 nats for VI and +2.65 nats for MCMC relative to persistence). VI and MCMC produced nearly identical predictive performance (average precision 0.895 versus 0.894), indicating that the mean-field variational approximation reliably captures the posterior predictive distribution for this model despite overdispersed marginal posteriors. For early warning of emerging outbreaks, flagging the ten highest-risk previously clear regions each month would have identified 36–38% of the 53 new outbreaks in the hold-out period, compared with none under the persistence baseline. The proposed framework provides an interpretable Bayesian model for regional ASF surveillance, demonstrating that distance-weighted graph-based epidemic models can improve early detection of emerging outbreaks.