Markov Switching Models for Outbreak Detection
Infectious disease outbreak detection is one of the main objectives of syndromic surveillance systems. Accurate and timely detection can provide valuable information for public health officials to react to major public health threats. However, disease out
- PDF / 1,360,100 Bytes
- 34 Pages / 439.37 x 666.142 pts Page_size
- 37 Downloads / 211 Views
CHAPTER OVERVIEW Infectious disease outbreak detection is one of the main objectives of syndromic surveillance systems. Accurate and timely detection can provide valuable information for public health officials to react to major public health threats. However, disease outbreaks are often not directly observable. Moreover, additional noise caused by routine behavioral patterns and special events further complicates the task of identifying abnormal patterns caused by infectious disease outbreaks. We consider the problem of identifying outbreak patterns in a syndrome count time series using the Markov switching models. The outbreak states are treated as hidden (unobservable) state variables. Gibbs sampler then is used to estimate both the parameters and hidden state variables. We cover both the theoretical foundation of the estimation methods and the technical details of estimating the Markov switching models. A case study is presented in the last section. Keywords: Markov switching models; Infectious disease informatics; Markov chain Monte Carlo; Gibbs sampler; Bayesian inference
*
Department of Information Management, National Taiwan University, Taipei, Taiwan. [email protected]
D. Zeng et al. (eds.), Infectious Disease Informatics and Biosurveillance, Integrated Series in Information Systems 27, DOI 10.1007/978-1-4419-6892-0_6, © Springer Science+Business Media, LLC 2011
111
Chapter 6
112
1.
INTRODUCTION
Recent efforts in building syndromic surveillance systems try to increase the timeliness of the data collection process by incorporating novel data sources such as emergency department (ED) chief complaints (CCs) and over-the-counter (OTC) health product sales. Studies show that these data sources do contain valuable information reflecting current public health status (Espino and Wagner, 2001; Ivanov et al., 2002; Chapman et al., 2005a, b). However, they usually carry substantial noise that may interfere with the detection of infectious disease outbreaks. To overcome the problem, researchers have been working on developing statistical methods that can extract disease outbreak signals from the realtime data provided by syndromic surveillance systems. Typically, the data are classified and aggregated to generate univariate or multivariate time series at daily frequency. A univariate time series may be the daily counts of patients with a particular syndrome (for example, the gastrointestinal syndrome) from an ED. A multivariate time series may be the daily number of patients with a particular syndrome from multiple EDs. If geographic information such as the ZIP code is available, the multivariate time series may be the daily counts of patients with a particular syndrome from the ZIP code areas near an ED. A popular time series outbreak detection method in current literature is a two-step procedure (Reis and Mandl, 2003; Reis et al., 2003). At the first step, a baseline model describing the “normal pattern” is estimated using the training data (usually a historical time series without outbreaks). The baseline m
Data Loading...
