Poisson conditional autoregressive model of spatio-temporal data is proposed. Markov property and probabilistic characteristics of this model are presented. Algorithms for maximum likelihood estimation of the model parameters are constructed. Optimal forecasting statistic minimizing probability of forecast error is given. The “plug-in” principle based on ML-estimators is used for forecasting in the case of unknown parameters. The results of computer experiments on simulated and real medical data are presented.