UC Santa Barbara

Traditional epidemic detection algorithms make decisions using only local information. We propose a novel approach that explicitly models spatial information fusion from several meta-populations. Our method also takes into account cost-benefit considerations regarding the announcement of epidemic. We utilize a compartmental stochastic model within a Bayesian detection framework, which leads to a dynamic optimization problem. The resulting adaptive, non-parametric detection strategy optimally balances detection delay vis-a-vis probability of false alarms. Our algorithm can also be used to optimize an existing detection strategy. Taking advantage of the underlying state-space structure, we represent the stopping rule in terms of a detection map, which visualizes the relationship between the multivariate system state and policy making. It also allows us to obtain an efficient simulation-based solution algorithm that is based on the Sequential Regression Monte Carlo (SRMC) approach of Gramacy and Ludkovski (SIFIN, 2015). We present two models for pseudo-posterior and illustrate our results on two synthetic examples. We also quantify the advantages of our adaptive detection relative to conventional threshold-based strategies.