Spatial Scan Statistics on the GPGPU

Kulldorff’s spatial scan statistic and the software implementation (SaTScan) are widely used for the detection and evaluation of geographic clusters, particularly within the health care community. Unfortunately, the computational time of the scan statistic depends on a wide variety of variables, and, depending on the chosen parameter settings and operations, the computational time can be on the order of seconds to weeks. The greatest factors in computational time are the number of cases in the dataset and the number of time intervals over which these case are being aggregated. Fortunately, the scan statistic algorithm is highly parallelizable, where the runtime can be equally divided over the number of processors used. Given that a Graphics Processing Unit (GPU) is a high performance many-core processor, we can take advantage of the GPU’s speed and the parallelizability of the scan statistic algorithm to create a low cost means of efficiently reducing the runtime. In this work, we present an implementation of the spatial scan statistic for the GPU using the CUDA programming language. Our current results focus on purely spatial scan statistics (as opposed to spatiotemporal) with an underlying Bernoulli distribution model. We discuss the resultant speed increase, issues with porting such algorithms to the GPU, modifications to the algorithm for further speed increases, and future ideas for utilizing scan statistics and the GPU in a visual analytics environment.