Learning Outbreak Regions in Bayesian Spatial Scan Statistics

Abstract

The problem of anomaly detection for biosurveillance is typically approached in an unsupervised setting, due to the small amount of labeled training data with positive examples of disease outbreaks. On the other hand, such model-based methods as the Bayesian scan statistic (BSS) naturally allow for adaptation to the supervised learning setting, provided that the models can be learned from a small number of training examples. We propose modeling the spatial characteristics of outbreaks from a small amount of training data using a generative model of outbreaks with latent center. We present the model and the EM-based learning of its parameters, and we compare its performance to the standard BSS method on simulated outbreaks injected into real-world Emergency Department visits data from Allegheny County, Pennsylvania.