Date Published: June 26, 2007
Publisher: Public Library of Science
Author(s): Ben Y Reis, Isaac S Kohane, Kenneth D Mandl, Lauren Ancel Meyers
Abstract: BackgroundAdvanced disease-surveillance systems have been deployed worldwide to provide early detection of infectious disease outbreaks and bioterrorist attacks. New methods that improve the overall detection capabilities of these systems can have a broad practical impact. Furthermore, most current generation surveillance systems are vulnerable to dramatic and unpredictable shifts in the health-care data that they monitor. These shifts can occur during major public events, such as the Olympics, as a result of population surges and public closures. Shifts can also occur during epidemics and pandemics as a result of quarantines, the worried-well flooding emergency departments or, conversely, the public staying away from hospitals for fear of nosocomial infection. Most surveillance systems are not robust to such shifts in health-care utilization, either because they do not adjust baselines and alert-thresholds to new utilization levels, or because the utilization shifts themselves may trigger an alarm. As a result, public-health crises and major public events threaten to undermine health-surveillance systems at the very times they are needed most.Methods and FindingsTo address this challenge, we introduce a class of epidemiological network models that monitor the relationships among different health-care data streams instead of monitoring the data streams themselves. By extracting the extra information present in the relationships between the data streams, these models have the potential to improve the detection capabilities of a system. Furthermore, the models’ relational nature has the potential to increase a system’s robustness to unpredictable baseline shifts. We implemented these models and evaluated their effectiveness using historical emergency department data from five hospitals in a single metropolitan area, recorded over a period of 4.5 y by the Automated Epidemiological Geotemporal Integrated Surveillance real-time public health–surveillance system, developed by the Children’s Hospital Informatics Program at the Harvard-MIT Division of Health Sciences and Technology on behalf of the Massachusetts Department of Public Health. We performed experiments with semi-synthetic outbreaks of different magnitudes and simulated baseline shifts of different types and magnitudes. The results show that the network models provide better detection of localized outbreaks, and greater robustness to unpredictable shifts than a reference time-series modeling approach.ConclusionsThe integrated network models of epidemiological data streams and their interrelationships have the potential to improve current surveillance efforts, providing better localized outbreak detection under normal circumstances, as well as more robust performance in the face of shifts in health-care utilization during epidemics and major public events.
Partial Text: Understanding and monitoring large-scale disease patterns is critical for planning and directing public-health responses during pandemics [1–5]. In order to address the growing threats of global infectious disease pandemics such as influenza , severe acute respiratory syndrome (SARS) , and bioterrorism , advanced disease-surveillance systems have been deployed worldwide to monitor epidemiological data such as hospital visits [9,10], pharmaceutical orders , and laboratory tests . Improving the overall detection capabilities of these systems can have a wide practical impact. Furthermore, it would be beneficial to reduce the vulnerability of many of these systems to shifts in health-care utilization that can occur during public-health emergencies such as epidemics and pandemics [13–15] or during major public events .
In this paper, we describe an epidemiological network model that monitors the relationships between health-care utilization data streams for the purpose of detecting disease outbreaks. Results from simulation experiments show that these models deliver improved outbreak-detection performance under normal conditions compared with a standard reference time-series model. Furthermore, the network models are far more robust than the reference model to the unpredictable baseline shifts that may occur around epidemics or large public events. The results also show that epidemiological relationships are inherently valuable for surveillance: the activity at one hospital can be better understood by examining it in relation to the activity at other hospitals. In a previous paper , we showed the benefits of interpreting epidemiological data in its temporal context—namely, the epidemiological activity on surrounding days . In the present study, we show that it is also beneficial to examine epidemiological data in its network context—i.e., the activity of related epidemiological data streams.