Cluster Detection Comparison in Syndromic Surveillance

OBJECTIVE To use the New York City Department of Health and Mental Hygiene’s (NYC DOHMH) emergency department (ED) syndromic surveillance data to evaluate FleXScan’s flexible scan statistic and compare it to results from the SaTScan circular scan. A secon d objective is to improve cluster detection in by improving geographic characteristics of the input fil es. BACKGROUND The NYC DOHMH collects data daily from 50 of 61 (82%) EDs in NYC representing 94% of all ED visits (avg daily visits ~10,000). The information collect d includes the date and time of visit, age, sex, home zip code and chief complaint of each patient. Observations are assigned to syndromes based on the chief complaint field and are analyzed using SaTScan to identify statistically significant clusters of synd romes at the zip code and hospital level [1]. SaTScan employs a circular spatial scan statistic and cluster s hat are not circular in nature may be more difficult to detect. FlexScan employs a flexible scan statistic using an adjacency matrix design [2][3]. METHODS Counts of syndrome visits were aggregated at the zi p code level for 2005. FleXScan’s flexible scan and SaTScan’s circular scan were analyzed by comparing the most likely cluster (primary cluster) identifie d; the secondary clusters identified; location and are a of identified cluster; P-value and relative risk. Both projected and unprojected coordinate systems were used to identify sensitivity in clusters to changes in m easurement and coordinate systems. Improving the FleXScan matrix file provided a method for capturin g area connectivity where bridges, tunnels, or subway lines existed between them. This was not possible t o do in SaTScan. ZIP code area centroids were weighted to reflect the underlying population distr ibution of the areas. Both FleXScan and SaTScan were run again using the reweighted centroids. RESULTS FleXScan and SaTScan both detected similar, overlapping areas in three of the time periods investigated. Non-circular clusters with a high relative r isk were detected by FleXScan’s flexible scan, but this was not detected by SaTScan (Figure 1). However, known clusters were detected at a more significant pvalue by SaTScan than FleXScan (p=0.002 vs. p=0.179). Weighting ZIP code centroids based on population and improving the connectivity matrix changed results; over a one week period p-values increased 50% of the time, decreased 36% of the time, and stayed the same 14% of the time when weighted centroids were employed. The differences were most prominent where unweighted centroids had not been representative of underlying populatio n distributions in the areas.