|
Abstract:
|
We present a nonparametric scan statistic (NPSS) for multivariate cluster detection. Given multiple streams of spatial time series data, we detect spatial regions where recent counts for some subset of streams are higher than expected. We first compute "empirical p-values" (EPV) for each stream and location, comparing current counts to the distribution of past counts. We scan over regions consisting of some subset of streams for some spatial area, and find regions with lower than expected EPV. We evaluated NPSS on disease outbreaks injected into hospital and pharmacy data from western PA. NPSS outperforms other approaches on datasets where parametric models are incorrect, and improves detection power by integrating information from multiple streams. Other advantages include characterization of affected streams, low false alarm rates, and ability to incorporate nontraditional data types.
|
