A three-state recursive sequential Bayesian algorithm for biosurveillance

Abstract

A serial signal detection algorithm is developed to monitor pre-diagnosis and medical diagnosis data pertaining to biosurveillance. The algorithm is three-state sequential, based on Bayesian thinking. It accounts for non-stationarity, irregularity and seasonality, and captures serial structural details of an epidemic curve. At stage n, a trichotomous variable governing the states of an epidemic is defined, and a prior distribution for time-indexed serial readings is set. The technicality consists of finding a posterior state probability based on the observed data history, using the posterior as a prior distribution for stage n+1 and sequentially monitoring surges in posterior state probabilities. A sensitivity analysis for validation is conducted and analytical formulas for the predictive distribution are supplied for error management purposes. The method is applied to syndromic surveillance data gathered in the United States (US) District of Columbia metropolitan area.