The analysis of the geographical variations of a disease and their representation on a mapis an important step in epidemiology. The goal is to identify homogeneous regions in termsof disease risk and to gain better insights into the mechanisms underlying the spread of thedisease. We recast the disease mapping issue of automatically classifying geographical unitsinto risk classes as a clustering task using a discrete hidden Markov model and Poisson classdependent distributions. The designed hidden Markov prior is non standard and consists of avariation of the Potts model where the interaction parameter can depend on the risk classes.The model parameters are estimated using an EM algorithm and the mean field approximation. This provides a way to face the intractability of the standard EM in this spatial context,with a computationally efficient alternative to more intensive simulation based Monte CarloMarkov Chain (MCMC) procedures.We then focus on the issue of dealing with very low risk values and small numbers of observedcases and population sizes. We address the problem of finding good initial parameter values inthis context and develop a new initialization strategy appropriate for spatial Poisson mixturesin the case of not so well separated classes as encountered in animal disease risk analysis.We illustrate the performance of the proposed methodology on some animal epidemiologicaldatasets provided by INRA.
