Approaches to syndromic surveillance when data consist of small regional counts

Statistical systems designed for syndromic surveillance often must be able to monitor data received simultaneously from multiple regions. Such data might be of limited size, which would eliminate the possibility of using more common surveillance methods that assume data from a normal distribution.

The objectives of this study were to design and illustrate a multiregional surveillance system based on data inputs consisting of small regional counts, where frequencies are typically on the order of

Cumulative sum (CUSUM) methods designed for cumulating the sum of the deviations between observed and expected Poisson-distributed data were modified to account for changing expectations over time, including weekly and monthly effects. Data on lower respiratory tract infections during 1996-1999 at multiple Boston clinics among residents from 287 census tracts were used to illustrate the approach.

When each region was monitored, 19% of the census tracts signaled a departure during 1999 from the base period (1996-1998) rates. When local statistics were used to monitor tracts and neighborhoods consisting of surrounding tracts, 60% of tracts experienced departures during 1999 from the base period. These results imply that the increases in lower respiratory tract infection that occurred during 1999 were geographically pervasive.

Poisson CUSUM methods are useful for monitoring small regional counts over time. The methods can be generalized to account for time-varying expectations in the counts.