The high-fidelity detectability experiments (HiFIDE) technique extends the semisynthetic method by forming injections whose shapes and noise levels are derived from surveillance data collected during an actual outbreak (Wallstrom et al., 2004, Wallstrom et al., 2005). The technique also scales the height of the inject in a way that preserves the known relationship between the magnitude of the real outbreak and the strength of the signal in the surveillance data collected during the real outbreak. The scaling adjusts for differences in population size and in data completeness (defined as the proportion of data that is available in a region). This technique estimates the effect of an outbreak on surveillance data in one region and allows the evaluator to determine the detectability of that type of outbreak in a second region (where the two regions may differ in population size, population density, completeness of surveillance data, and sizes of outbreaks).
The mechanics of a HiFIDE analysis are similar to those of a semi-synthetic analysis. An evaluator creates multiple injects, varying in outbreak size and surveillance data completeness. Each inject is then combined with real surveillance data to form a time series. An evaluator runs detection algorithms on each injected time series, varying the detection threshold, and summarizes the detectability characteristics of the algorithms using AMOC curves and other generalized ROC curves that display the relationships between timeliness of detection, sensitivity, false alarm rate, outbreak size, and data completeness.
A software package, also called HiFIDE, is available from www.hifide.org that automates the analysis. The software can be freely downloaded for noncommercial use. Note that HiFIDE is limited to synthetic additions to univariate data. The problem of adding synthetic additions to multivariate data is an ongoing area of research. It is a complex problem because one would need to simulate correctly the interdepen-dencies between the variables. Buckeridge et al. (2004) are developing an approach to injecting simulated anthrax records into a set of relational tables.
Was this article helpful?