A Nonparametric Scan Statistic for Multivariate Disease Surveillance

OBJECTIVE We present a new method for multivariate outbreak detection, the “nonparametric scan statistic” (NPSS). NPSS enables fast and accurate detection of emerging space-time clusters using multiple disparate data streams, including nontraditional data sources where standard parametric model assumptions are incorrect. BACKGROUND Expectation-based scan statistics [1] extend the traditional spatial and space-time scan statistics [2-3] by inferring expected counts for each location from past data and detecting regions where recent counts are higher than expected. While these methods have been shown to achieve high detection power across a variety of datasets and outbreak types [4], they make strong parametric model assumptions (e.g. Poisson or Gaussian counts), and performance degrades when these models are incorrect. Our solution, NPSS, is a new scan statistic approach which does not assume a parametric model, but instead combines empirical pvalues across multiple locations, days, and data streams to discover significant disease clusters. METHODS Given a time series of observed counts ci,m for each data stream Dm at each location si, we wish to detect spatial regions where the recent counts for some subset of streams are higher than expected. To do so, we first compute empirical p-values Pi,m for each stream and location for each recent day: these are defined as (Tbeat+ 1) / (T+1), where Tbeat is the number of past days (for that data stream and location) with higher residuals and T is the total number of past days. We then scan over the space-time regions (D, S, W), each consisting of some subset of streams D for some spatial area S for the last W days. We search for regions where the Pi,m are significantly lower than expected, corresponding to higher than expected counts.