A multi-state spatio-temporal Markov model for categorized incidence of meningitis in sub-Saharan Africa

Meningococcal meningitis is a major public health problem in the African Belt. Despite the obvious seasonality of epidemics, the factors driving them are still poorly understood. Here, we provide a first attempt to predict epidemics at the spatio-temporal scale required for in-year response, using a purely empirical approach. District-level weekly incidence rates for Niger (1986-2007) were discretized into latent, alert and epidemic states according to pre-specified epidemiological thresholds. We modelled the probabilities of transition between states, accounting for seasonality and spatio-temporal dependence. One-week-ahead predictions for entering the epidemic state were generated with specificity and negative predictive value >99%, sensitivity and positive predictive value >72%. On the annual scale, we predict the first entry of a district into the epidemic state with sensitivity 65∙0%, positive predictive value 49∙0%, and an average time gained of 4∙6 weeks. These results could inform decisions on preparatory actions.