Automated biosurveillance: Searching for aberration detection algorithms with robust detection performance

Background: Use of robust and broadly applicable statistical alerting methods is essential for a public health syndromic surveillance system, such as CDC BioSense program.

Methods: Excess visit counts (at least two standard deviations above calculated baseline expectations) were artificially injected into gastrointestinal (GI) syndrome-related daily aggregate time series data from the BioSense Syndromic Surveillance system. Algorithm sensitivity was calculated as the ratio of days with injects for which the algorithm value exceeded a computed threshold. We compared these sensitivities among the three adaptive methods: a CUSUM (CS) chart and two Shewhart chart variations based on sums and ratios (C2a, C2b) adjusted for total visits. We also examined overall sensitivities before and after stratification by weekday versus weekend and holiday with different background ratios to total visits that are GI-related.

Results: At a daily background alert rate of 5% and 1%, the sensitivities ranged between 31%-79% and 10%-64%, respectively. Sensitivities were, in general, higher for CS and C2a and lower for C2b. However, the overall sensitivity for C2b increased significantly after stratification by weekday/weekend and holiday.

Conclusions: CS and C2a performed better than C2b for detection of short signals in GI-related visit series. These analyses demonstrated that sensitivity can be improved overall if day-of-week differential in background ratio to total visits is taken into account. Further analyses using time-series from additional syndromes and multi-day signal shapes to allow timeliness comparisons will be presented.