Measuring Outbreak-Detection Performance By Using Controlled Feature Set Simulations

Publisher Website
Google Scholar
Abstract
The outbreak-detection performance of a syndromic surveillance system can be measured in terms of its ability to detect signal (i.e., disease outbreak) against background noise (i.e., normally varying baseline disease in the region). Such benchmarking requires training and the use of validation data sets. Because only a limited number of persons have been infected with agents of biologic terrorism, data are generally unavailable, and simulation is necessary. An approach for evaluation of outbreak-detection algorithms was developed that uses semisynthetic data sets to provide real background (which effectively becomes the noise in the signal-to-noise problem) with artificially injected signal. The injected signal is defined by a controlled feature set of variable parameters, including size, shape, and duration. This report defines a flexible approach to evaluating public health surveillance systems for early detection of outbreaks and provides examples of its use. The stages of outbreak detection are described, followed by the procedure for creating data sets for benchmarking performance. Approaches to setting parameters for simulated outbreaks by using controlled feature sets are detailed, and metrics for detection performance are proposed. Finally, a series of experiments using semisynthetic data sets with artificially introduced outbreaks defined with controlled feature sets is reviewed. These experiments indicate the flexibility of controlled feature set simulation for evaluating outbreak-detection sensitivity and specificity, optimizing attributes of detection algorithms (e.g., temporal windows), choosing approaches to syndrome groupings, and determining best strategies for integrating data from multiple sources. The use of semisynthetic data sets containing authentic baseline and simulated outbreaks defined by a controlled feature set provides a valuable means for benchmarking the detection performance of syndromic surveillance systems.

This publication has 0 references indexed in Scilit: