A comparative analysis of bayesian network and ARIMA approaches to malaria outbreak prediction

Disease outbreaks are important to predict since they indicate hot spots of transmission with high risk of spread to neighboring regions and can thus guide the allocation of resources. While numeric prediction models can be easily used for outbreak prediction by setting thresholds, an alternative is to build a model that specifically classifies situations into outbreak or none. In this paper we compare Bayesian network models built for the outbreak classification problem with Bayesian network, ARIMA and ARIMAX models built for numeric prediction and used for outbreak prediction by thresholding. We show that in most cases the classification models outperform the other models. We then investigate the reasons underlying the differences in performance among the models in order to shed light on their strengths and weaknesses. The models are developed and evaluated using two years of malaria and environmental data from northern Thailand.