While several authors have advocated wavelets for biosurveillance, there are few published wavelet method evaluations using real syndromic data. Goldenberg et al. performed an analysis using wavelet predictions as a way of detecting a simulated anthrax outbreak. The commercial RODS application uses averaged wavelet levels to normalize for longterm trends and negative singularities. In line with the implementation in and in contrast to, we introduce two preconditioning steps to account for the strong day-of-week effect and holidays, and then use all levels of the wavelets to predict or alarm.
Objective
Syndromic data are created by processes that operate on different time scales (daily, weekly, or even yearly) and can include events of different durations from a 1-2 day outbreak of foodborne illness to a more gradual, protracted flu season. The duration of an outbreak caused by a new pathogenic strain or a bioterrorist attack is indeterminate. Wavelets are well suited for detecting signals of uncertain duration because they decompose data at multiple time and frequency scales. This study evaluates the use of several wavelet-based algorithms for both time series forecasting and anomaly detection using real-world syndromic data from multiple data sources and geographic locations.