 |
The 130th Annual Meeting of APHA |
|
The 130th Annual Meeting of APHA |
Michael A. Stoto, PhD, Statistics group and RAND Health, RAND, 1200 South Hayes St., Arlington, VA 22202-5050, 703-413-1100 x5472, mstoto@rand.org
Syndromic surveillance systems are intended to give early warnings of bioterrorist attacks or other emerging health conditions. However, even with access to the requisite data and perfect organizational coordination and cooperation, the statistical challenges in detecting an incident are formidable, involving an intrinsic tradeoff among sensitivity, the false positive rate, and timeliness.
To illustrate and characterize this tradeoff, we analyzed four statistical detection algorithms using daily counts of patients with influenza-like illness from a hospital emergency department. We simulated two outbreak patterns: a “fast” pattern had 3, 6, and 9 cases over three contiguous days, and a “slow” pattern with the same total number of cases over 9 days.
Under these conditions outside of the flu season, the detection algorithms each had a roughly 50 percent chance of detecting a “fast” outbreak on the second day, increasing to nearly 100 percent on the third day. During the flu season detection thresholds must be set higher to avoid detecting the flu itself, and as a result sensitivity of the detection algorithms ranged from 3 to 19 percent on the third day.
The “slow” outbreak was more difficult to detect. Outside the flu season the three detection algorithms that integrate data from more than one day were roughly equivalent to each other but have only an approximately 20 percent chance of detection at day 6 and a 50 percent chance at day 9. In the flu season, none of the methods reaches a 10 percent detection rate by day 9.
Learning Objectives:
Presenting author's disclosure statement:
I do not have any significant financial interest/arrangement or affiliation with any organization/institution whose products or services are being discussed in this session.