Detection Algorithm Method

The detection algorithm method determines the date that a statistical detection algorithm first detects an anomaly in surveillance data. The evaluator compares this date with a gold standard reference date. The evaluator may establish the gold standard reference date by convening a panel of experts to review the outbreak, or he may determine the date by running a detection algorithm on a second time series, one with face validity for the outbreak (Figure 21.3).

If the evaluator obtains the reference date from a second time series, he may elect to use the same detection algorithm on both time series, or different algorithms. It is important that his choice not bias the results, so that in general, an evaluator will use the same detection algorithm unless the time series have different statistical properties (e.g., one has weekly periodic effects and the other is nonstationary) (See Chapter 14 for a discussion of periodic effects in surveillance data). The use of two different algorithms creates an experimental situation in which two variables have been manipulated, thus the eval-uator must use care in the selection of the algorithms so that the differences observed can be reasonably attributed to underlying differences in the data, not a bias that was introduced by his selection of algorithms.

The detection algorithm method was first introduced by Quenel et al. (1994) in a survey of novel types of surveillance

/ 1 2 3 4\ 5 6 week figure 21.3 Detection algorithm method.Two time series (A and B) show outbreak effect. The detection algorithm method finds the dates on which a detection algorithm (whose threshold is indicated by the dashed lines) first detects the increase in the surveillance data. An evaluator uses this difference as a measure of the relatively earliness of two types of surveillance data.

data for detecting influenza. It was next used by Tsui et al. (2001) for a study of chief complaints coded with the ICD-9-CM (International Classification of Diseases, 9th Revision, Clinical Modification). We discuss other examples of its use in the next chapter (Hogan et al., 2003, Campbell et al., 2004, Ivanov et al., 2003).

0 0

Post a comment