515 – Time Series and Benchmarking

A Comparison of Regression Models with Adaptive Control-chart Algorithms for Aberration Detection in Biosurveillance Systems

at CDC

Biosurveillance systems require robust anomaly detection methods. For detection-method performance comparisons, we injected multi-day lognormal distributed signals into gastrointestinal (GI) syndrome-related time series of aggregated daily counts from the Centers for Disease Control and Prevention's BioSense syndromic surveillance system. CDC is part of the U.S. Department of Health and Human Services. We included a sample of facilities with data reported every day and with median daily syndromic counts = 3 over the entire study period from Jan. 2010 through May 2011. We compared alerting algorithms on the basis of a Poisson regression model (including covariates for day of week, total visits, and seasonality) with two adaptations of the cumulative sum (CuSUM) chart and three adaptations of the Shewhart chart (with variations on whether/how to adjust for total visits). We assessed sensitivity and timeliness of these methods for detection of injected multi-day signal events. Sensitivity was defined as the ratio of number of events detected before the event peak to the total number of injected signals. Alerting timeliness was calculated as the number of days from the start of injection to the first algorithm alert not later than the peak day of the injected signal. At a daily background alert rate of 1%, the sensitivities and timeliness measured as delay (in days) before signal detection ranged from 26%?66% and 2.6?3.3 days, respectively. We also examined sensitivity and timeliness with and without stratification by weekday versus weekend/ holiday. For time series with mean daily counts = 10, the Poisson regression-based method achieved higher sensitivity and slightly shorter mean detection delays than the chart-based methods for detection of multi-day signals in GI-related visit series.