K-scan for anomaly detection in disease surveillance

Anomaly detection in disease surveillance has recently been the focus of active research. Often, this involves searching for clusters of unusually high incidence rates among current cases of disease incidence, against a background incidence rate which may be spatially varying due to underlying variation in population and environmental characteristics. We propose a method, called K-scan, to identify such anomalies, using components of the inhomogeneous K function. Specifically, we assign to each case i a value K_i which, when summed together over all case locations, yields the overall inhomogeneous K function for the point pattern of disease locations. Clusters are identified using a bottom-up approach. Points with high values of K_i are first identified. Neighboring points are then iteratively added to form potential clusters. The inhomogeneous K function restricted to the region covered by each potential cluster is used as the criterion for selecting the reported cluster. Significance is computed using a parametric bootstrap approach with the inhomogeneous Poisson process. The intensity function is estimated using observations of case locations at prior time points. This intensity function is also used in the computations of the K_i values and provides a way to account for the spatially varying incidence rates when searching for clusters. We study the performance and behavior of this K-scan method through a simulation study, and find that the method works well in identifying the correct cluster, is computationally fast and can yield irregularly shaped clusters. We also include a comparison with the spatial scan statistic as implemented in SaTScan. Finally, we apply the method to dead bird data from Contra Costa county in California to identify anomalies in dead bird sightings.