For 35 reportable communicable diseases for which cluster detection could inform programmatic activities (1), we analyzed disease counts for patients of all ages combined. For amebiasis, cryptosporidiosis, and giardiasis, for which outbreaks among young children are of particular interest, additional analyses were restricted to disease counts among patients <5 years of age, for 38 total daily analyses.

In BCD’s application, the space–time permutation scan statistic detects disease clusters in space–time cylinders centered on every census tract centroid; the circular base represents space (maximum geographic cluster size of 50% of all reported cases), and the height represents time (maximum temporal window length of 30 days, for most diseases). For each cylinder, a likelihood ratio–based test statistic is calculated. The test statistic is considered elevated if the observed disease count during the time window in census tracts with centroids inside the cylinder’s circular base exceeds the expected number of cases, which is a function of 1) the case count in the circle during a baseline period (which accounts for any purely geographic variations in disease occurrence, diagnosis, and reporting) and 2) the total case count citywide during the time window (which accounts for citywide purely temporal patterns, such as seasonality or secular trends) (3). The cylinder with the maximum test statistic is the cluster least likely to be due to chance under the null hypothesis that the same process generated disease counts inside and outside the cylinder.

To create a simulated dataset, cases’ dates are randomly shuffled and assigned to the original census tracts. The maximum statistic for each simulated dataset is calculated in the same way as for the observed dataset. For each disease, this process is repeated daily 999 times. The maximum value for the observed dataset is ranked among the 999 trial maxima. A p value (range 0.001–1) is derived from this ranking; p = 0.001 represents the highest significance relative to the permutation trials. The Monte Carlo approach to deriving significance by using repeated trials, each permuting observed data attributes, is designed to control for multiple testing.

A recurrence interval (RI) is calculated as the reciprocal of the p value and represents the number of days of daily surveillance required for the expected number of clusters at least as unusual as the observed cluster to be equal to 1 by chance (11). We defined a signal as any cluster with an RI >100 days; that is, during any 100-day daily analysis period, the expected number of clusters at least as unlikely as the current cluster is 1.

We developed a SAS program (SAS Institute, Inc., Cary, NC, USA) to generate case and parameter files (Table 1), read in a coordinate file of census tract centroids, invoke SaTScan in batch mode, read analysis results back into SAS for further processing, and output files to secured folders. For any signals, the program also generated emails notifying BCD leadership and staff responsible for follow-up (Technical Appendix).

Figure

Figure. Automated output from spatiotemporal analysis on July 17, 2015, indicating a cluster (dark gray) of 8 legionellosis cases over 8 days centered in the South Bronx, New York City, New York,...

This automated analysis detected the second largest US outbreak of community-acquired legionellosis (12), identifying a cluster of 8 cases centered in the South Bronx on Friday, July 17, 2015 (RI = 500 days) (Figure), before any human public health monitor noticed it. On Monday, July 20, an increase in cases was independently noticed by BCD staff members routinely investigating individual cases, and on July 21, an infection-control nurse working in the outbreak area called BCD to report an increase. The DOHMH and state and federal partners conducted an extensive epidemiologic, environmental, and laboratory investigation to identify and remediate the outbreak source, a cooling tower.

A shigellosis outbreak among the observant Jewish community in Brooklyn (13) began in late October 2014 and was detected with 9 cases on November 14, 2014 (RI = 333 days). BCD does not routinely investigate individual shigellosis reports, so automated analysis alone prompted early outbreak identification. Shigellosis outbreaks within this community occur cyclically and have been linked to daycare and preschool attendance (14). Starting in mid-November, BCD staff visited community schools, daycare centers, and health fairs to promote appropriate handwashing. The outbreak subsided by mid-March 2015. Other clusters prompting investigations included legionellosis (Queens, April–May 2015) and campylobacteriosis (Brooklyn, October 2014). During a 1-year period, 28 unique signals were observed across 15 diseases (Table 2), which staff perceived as a reasonable number for investigation.

Not all detected clusters were actionable. No public health response was conducted for an amebiasis cluster (Manhattan, April 2015; RI = 143 days) consisting of 6 men (34–49 years of age) diagnosed within a 12-day period and residing within a 0.35-mile radius because no case-patients were identified as food handlers or daycare workers. A public health response also was not conducted for a giardiasis cluster (Bronx, April 2015; RI = 1,000 days) that consisted of 6 household members who acquired the infection during international travel. Investigators were interested in being notified of and following such clusters over time, even if they ultimately were not actionable or verified as true outbreaks.

Several outbreaks in New York City, New York, were detected by daily automated spatiotemporal analyses. Early cluster detection facilitated prioritization of individual case investigations, outbreak recognition and investigation, provider and community outreach, and timely intervention to limit sickness and death. This method has proven particularly useful for identifying and monitoring outbreaks of shigellosis (6,8,9) and legionellosis and might be useful for monitoring additional diseases with outbreak potential, including pertussis, syphilis, and tuberculosis.

Key to the system’s success is a strong informatics infrastructure, especially electronic laboratory reporting and near real-time geocoding of surveillance data. Other facilitators include a powerful statistical disease surveillance methodology, knowledgeable epidemiologists to interpret signals, and adequate outbreak investigation resources.

These methods could be useful to other health departments receiving more reports than can be rapidly reviewed manually. State health departments could consider conducting similar analyses to detect clusters spanning multiple jurisdictions.