Abstract
A recent review of the literature on surveillance systems revealed an enormous number of research-related articles, a host of websites, and a relatively small (but rapidly increasing) number of actual surveillance systems, especially for the early detection of a bioterrorist attack [BMS04]. Modern bioterrorism surveillance systems such as those deployed in New York City, western Pennsylvania, Atlanta, and Washington, DC, routinely collect data from multiple sources, both traditional and nontraditional, with the dual goal of the rapid detection of localized bioterrorist attacks and related infectious diseases. There is an intuitive notion underlying such detection systems, namely, that detecting an outbreak early enough would enable public health and medical systems to react in a timely fashion and thus save many lives. Demonstrating the real efficacy of such systems, however, remains a challenge that has yet to be met, and several authors and analysts have questioned their value (e.g., see Reingold [Rei03] and Stoto et al. (2004) [SSM04, SFJ06]). This article explores the potential and initial evidence adduced in support of such systems and describes some of what seems to be emerging as relevant statistical methodology to be employed in them. Public health and medical data sources include mortality rates, lab results, emergency room (ER) visits, school absences, veterinary reports, and 911 calls. Such data are directly related to the treatment and diagnosis that would follow a bioterrorist attack. They might not, however, detect the outbreak sufficiently fast. Several recent national efforts have been focused on monitoring "earlier" data sources for the detection of bioterrorist attacks or other outbreaks, such as over-the-counter (OTC) medication sales, nurse hotlines, or even searches on medical websites (e.g., WebMD). This assumes that people who are not aware of the outbreak and are feeling sick, would gen110 Galit Shmueli and Stephen E. Fienberg erally seek self-treatment before approaching the medical system and that an outbreak signature will manifest itself earlier in such data. According to Wagner et al. [WRT03], preliminary studies suggest that sales of OTC health care products can be used for the early detection of outbreaks, but research progress has been slow due to the difficulty that investigators have in acquiring suitable data to test this hypothesis for sizable outbreaks. Some data of this sort are already being collected (e.g., pharmacy and grocery sales). Other potential nontraditional data sources that are currently not collected (e.g., browsing in medical websites, automatic body sensor devices) could contain even earlier signatures of an outbreak.3 To achieve rapid detection there are several requirements that a surveillance system must satisfy: frequent data collection, fast data transfer (electronic reporting), real-time analysis of incoming data, and immediate reporting. Since the goal is to detect a large, localized bioterrorist attack, the collected information must be local, but sufficiently large to contain a detectable signal. Of course, the different sources must carry an early signal of the attack. There are, however, trade-offs between these features; although we require frequent data for rapid detection, too frequent data might be too noisy to the degree that the signal is too weak for detection. A typical solution for too frequent data is temporal aggregation. Two examples where aggregation is used for biosurveillance are aggregating OTC medication sales from hourly to daily counts [GSC02] and aggregating daily hospital visits into multiday counts [RPM03]. A similar trade-off occurs between the level of localization of the data and their amount. If the data are too localized, there might be insufficient data for detection, whereas spatial aggregation might dampen the signal. Another important set of considerations that limit the frequency and locality of collected data relate to confidentiality and data disclosure issues (concerns over ownership, agreements with retailers, personal and organizational privacy, etc.). Finding a level of aggregation that contains a strong enough signal, that is readily available for collection without confronting legal obstacles, and yet is sufficiently rapid and localized for rapid detection, is clearly a challenge. We describe some of the confidentiality and privacy issues briefly here. There are many additional challenges associated with the phases of data collection, storage, and transfer. These include standardization, quality control, confidentiality, etc. [FS05]. In this paper we focus on the statistical challenges associated with the data monitoring phase, and in particular, data in the form of multiple time series. We start by describing data sources that are 3 While our focus in this article is on passive data collection systems for syndromic surveillance, there are other active approaches that have been suggested (e.g., screening of blood donors [KPF03]), as well as more technological fixes, such as biosensors [Sul03] and "Zebra" chips for clinical medical diagnostic recording, data analysis, and transmission [Cas04]. Monitoring Multiple Data Streams for Biosurveillance 111 collected by some major surveillance systems and their characteristics. We then examine various traditional monitoring tools and approaches that have been in use in statistics in general, and in biosurveillance in particular. We discuss their assumptions and evaluate their strengths and weaknesses in the context of biosurveillance. The evaluation criteria are based on the requirements of an automated, nearly real-time surveillance system that performs online (or prospective) monitoring of incoming data. These are clearly different than for retrospective analysis [SB03] and include computational complexity, ease of interpretation, roll-forward features, and flexibility for different types of data and outbreaks. Currently, the most advanced surveillance systems routinely collect data from multiple sources on multiple data streams. Most of the actual statistical monitoring, however, is typically done at the univariate time series level, using a wide array of statistical prediction methodologies. Ideally, multivariate methods should be used so that the data can be treated in an integrated way, accounting for the relationships between the data sources. We describe the traditional statistical methods for multivariate monitoring and their shortcomings in the context of biosurveillance. Finally, we describe monitoring methods, in both the univariate and multivariate sections, that have evolved in other fields and appear potentially useful for biosurveillance of traditional and nontraditional temporal data.We describe the methods and describe their strengths and weaknesses for modern biosurveillance. © 2006 Springer Science+Business Media, LLC.