Lewis, Bryan PhD, MPH; Swarup, Samarth PhD; Bisset, Keith PhD; Eubank, Stephen PhD; Marathe, Madhav PhD; Barrett, Chris PhD
Network Dynamics and Simulation Science Laboratory, Virginia Bioinformatics Institute, Virginia Tech, Blacksburg.
Correspondence: Bryan Lewis, PhD, MPH, Social Decisions and Informatics Laboratory, Virginia Tech Research Center, 900 North Glebe Rd, Arlington, VA 22203 ( email@example.com).
This study was supported by the Department of Defense through fund DTRA CNIMS Contract HDTRA1-11-D-0016-0001, Defense Threat Reduction Agency Comprehensive National Incident Management System Contract and DTRA grant HDTRA1-11-1-0016, Defense Threat Reduction Agency Validation grant; National Institutes of Health through NIH MIDAS grant 2U01GM070694-09, National Institute of General Medical Sciences—Models of Infectious Disease Agent Study grant; National Science Foundation through NSF PetaApps grant OCI-0904844, Accelerating Discovery in Science and Engineering Through PetaScale Simulations and Analysis grant and NSF NetSE grant CNS-1011769, and Network Science and Engineering grant.
Chris Barrett, Keith Bisset, Jiangzhou Chen, Stephen Eubank, Annette Feng, Kathy Laskowski, Bryan Lewis, Achla Marathe, Madhav Marathe, Bill Marmagas, Henning Mortveit, Paula Stretz, Samarth Swarup, Anil Vullikanti, Dawen Xie, Jose Jimenez, Junwhan Kim, Akshay Maloo, Nidhi Parikh, Guanhong Pei, Caitlin Rivers, Sudip Saha, and Balaaji Sunapanasubbiah contributed in building the simulation environment and the supporting data infrastructure. The authors also acknowledge the contributions of time, data, domain area expertise, and support from Todd Han, Dave Myers, and Mike Phillips (DTRA Reachback: Blast simulation and management); Mike Snow, Jian Lu, and Mandy Wilson (VBI CCF: High performance computing support); Ryan Quint, Yaman Evrenosoglu, Arun Phadke, and James Thorp (Electrical Engineering, Virginia Tech: expertise in power networks); Jeff Reed (Electrical Engineering, Virginia Tech: expertise in communication networks); Nishith Tripathi (Award Solutions: expert in communication networks); Dane Webster (School of Visual Arts, Virginia Tech: expertise in visualization and graphics); and Thomas Dickerson and Peter Sforza (CGIT, Virginia Tech: expertise in GIS).
The authors declare no conflicts of interest.
Supplemental digital content is available for this article. Direct URL citations appear in the printed text and are provided in the HTML and PDF versions of this article on the journal's Web site ( http://www.JPHMP.com).
Disasters whether natural or human-initiated remain a critical challenge for our societies. However, the increasing interconnectedness of our society, both through near-ubiquitous cellular communication and rise in social media use abetted by increased smartphone prevalence, presents a great opportunity to improve our response capabilities.1 Similarly, improvements in computer simulation technologies provide opportunities to enhance our ability to prepare. We present a novel simulation approach that facilitates the dynamic evaluation of policies and interventions in the days following disasters.
Much is understood about the keys to enhancing community resilience in the face of disasters.2–5 Much of this is based on previous experience and careful consideration of hypothetical situations,6–9 yet gaps remain in our understanding as disasters present challenges for scientific study.10–12 Data collection and experimentation are hardly concerns during the immediate response stage, and their unpredictable nature eludes controlled experimentation. However, we still need to collect data both to inform our decisions and to evaluate how to turn our scientific knowledge into action.13 Simulation frameworks provide a natural environment for this endeavor.
A primary challenge faced by simulations attempting to represent an event as complex as a major disaster is accommodating the actions of the people affected by the disaster and the cascading of effects of these actions. These cascades, a common characteristic of complex systems, result from the interdependencies between the entities in the system. We address these challenges by using an agent-based approach where individuals in the population are guided by central motivation and choose actions based on that motivation, their local environment, and their assessment of their situation. In addition, we use a relational database to coordinate multiple distributed computational modules that accommodates the inclusion of large numbers of interdependencies, thus allowing the effects of different actions to cascade through the simulated system.
To demonstrate the capabilities of this simulation environment, we conducted a study to evaluate the potential benefits of partially restoring cellular communications following a large-scale disaster, specifically an improvised nuclear detonation of a 10-kiloton device in downtown District of Columbia on a weekday morning (for our purposes 16th and K Street at 11:15 am on May 15, 2006). The prospect of an unanticipated nuclear detonation in a large metropolitan area has been considered and analyzed extensively.14–17 This event is overwhelmingly physical and destroys a large section of District of Columbia, causes widespread damage to buildings and roads, produces a plume of radioactive fallout that eventually stretches across rural Maryland, disrupts power to a broad area, injures 97 000 individuals, and kills 279 000 individuals. Because of space limitations, we provide an overview of the main features and capabilities of this approach and briefly describe the outcomes of the study.
The simulation environment uses a data-centric, agent-based approach and is centered on a relational database to coordinate multiple distributed computational modules. The database captures the state of the world at different time points, and each of the modules updates this state for specific domains: Health, Behavior, Communication, Physical Effects, and Power. This database was initially populated by a series of data fusion and refinement steps specific to the population and the infrastructures represented.
For this study, we selected a population that would experience some effect from the blast. Specifically, we chose a contour line around at least 2.1 cal/cm2 of thermal radiation and at least 0.01 rad/h of radiation from fallout (at 1 hour following the blast). This includes 730 833 people in the District of Columbia and Prince George's County, Maryland (Figure 1, inset).
At the heart of the simulation, individuals are represented and possess key attributes that determine their motivations and responses to the changing environment in which they are immersed. Using demographic distributions and sample household information from the 2009 American Community Survey, we generate synthetic individuals and place them in households while maintaining the demographic distributions. An iterative proportional fitting algorithm18 is used to ensure that joint distributions are preserved down to the block-group level. Thus, the synthetic population has a detailed representation of household structure, age, gender, and income level that will produce the same block group distributions as the real population while maintaining the anonymity of the population.
Individuals in this synthetic population are assigned activity templates, which were derived from the National Household Travel Survey and the National Center for Education Statistics, based on demographics. These activity templates contained both the type (eg, home, work, school, shopping, other) and timing of the activities. This approach generates a dynamic “in silico” society where each individual moves between his or her activity locations throughout the simulated day in a very realistic approximation of the represented population.19 In addition, populations that are not represented in the 2009 American Community Survey, such as tourists and students living in dormitories, were added.
The behaviors and movements of the individuals are constrained by the physical world around them. A highly detailed representation of the physical world was constructed on the basis of locations. Locations are any geolocatable site that has an attribute that could be influenced by the event or would influence the movement or behavior of an individual. A listing of the data sources used and links when open source can be found in the Supplemental Digital Content (available at: http://links.lww.com/JPHMP/A27).
For the Power module to determine which locations maintain power after the event, the electric grid was added to the database and the impact from the blast estimated. In addition, the power load shift immediately following the blast was simulated separately and found a high probability that the power outage would cascade further, reducing the area with power.
For the Communication module to route cell phone calls, the location of all the cell towers were loaded into the database. The module handles each call in the order initiated by the agents and determines when the tower will be over capacity and starts to reject calls. Several towers are destroyed by the blast itself, and many more are without power. The power level of the backup batteries at each tower is also taken into account as is the “emergency broadcast” system (Commercial Mobile Alert System) messages. This module also determines the placement of the mobile “cells on wheels” that begin to arrive to restore communications in areas without power (in the branch of the study with partial communication restoration).
For the Physical Effects module to update the built infrastructure based on the effects of the event, data from a NucFast simulation of the nuclear blast were imported into the database. This included a temporally dynamic location of thermal radiation, blast overpressures, and their subsequent damage to the buildings, taking into account the type of building. In addition, the radiation both prompt and from fallout was also calculated over a 72-hour period and the module updates these dynamic effects.
In a disaster, several factors play into the actions and movements of individuals. Primarily, their actions are determined by the time from the event, their physical well-being, their knowledge of their current risk, and their ability to ascertain the safety of their loved ones. The simulation updates the state of individuals and is influenced by the physical environment, which is driven by the infrastructure modules (Figure 2).
The Behavior module coordinates the motivations and actions of an individual. The complex decision-making process of human behavior was simplified into a decision tree (Figure 3), which uses the individual agent's state (both physical and mental) to probabilistically determine a primary motivation. These motivations were then used, in turn, to direct the immediate actions of individuals. Household Reconstitution seeks to ensure the well-being of their household. To achieve this goal, they will make phone calls and travel toward the location of their household members. They will continue to do so until all members are together, or they know that all members are safely out of danger, at that point they (and whatever members they have gathered) will evacuate the area. Those following an Evacuation motivation seek to ensure safety by leaving the area as quickly and efficiently as possible. To achieve this goal if individuals are mobile, they will travel toward the closest major “evacuation route” until they are outside of the event area. When Shelter-Seeking individuals seek to limit exposure to harm from the event outside by finding shelter and remaining there. In this particular study, individuals would try to avoid the effects of the radioactive fallout by finding buildings with minimal damage and remain there for as long as they were healthy or in some cases until their patience ends. The Health Care-Seeking motivation seeks to obtain medical attention for an injury. Individuals will call 911 if they are unable to move, or if mobile, they will travel to the nearest hospital. If individuals begin to Panic, they become irrational and essentially begin following one of the other motivations without regard to other factors. The likelihood of panic reduces over time from the initial blast such that very few people are panicking after 5 hours. The Aid and Assist motivation seeks to rescue and aid small children and the injured. Individuals will seek to find others to help them reach at the nearest locations that can provide shelter, and upon finding them will take them to the nearest health care location, after which they will then evacuate.
The Transportation module calculates routes from the current location to the chosen destination (eg, health care center, home) and updates the location of individuals as they progress to their goal on the basis of time traveled. Routes are based on the road network (also including prominent footpaths), new routes are dynamically generated when the route is blocked by rubble, and progress is slowed when travel is over damaged terrain. To initially place individuals, a normal day was simulated and each individual's location at 11:15 was determined, including locations of those in transit either by public transportation or by personal vehicle.
The Health module updates the health state of individuals, which, in turn, drives their behaviors, affects their mobility, and influences their response to treatment. For these purposes, a simple model that represents health on a continuum for injury triage (based on the SALT triage20) is used as the main health state. This continuum consists of states from 0 (death) to 7 (full health), with states 4 and lower corresponding to moderate injury or worse. Secondary effects of health (mobility, health care requirements, etc) are based on this state, as are the effects of additional injuries and treatments. Risk of injury is based on the location of the individual and thus captures damage to buildings, roadways, etc, and, in the case of this study, exposure to radiation.
Health care locations consist of the existing hospitals in the area and mobile health centers that may be deployed following the disaster. The capacity of the health care center is based on the number of employees at the center and the damage it has incurred. Individuals seeking care are evaluated on a first-come-first-serve basis; after evaluation, a triaged queue is created. Individuals who are moderately injured may give up if the wait is too long. If individuals survive until they are treated, their health is adjusted on the basis of the severity and type of their injury (not all treatments are successful).
Response policies and interventions
Disasters of the nature of this scenario have been extensively studied by many different parties16,17,21 to inform preparedness activities. Many response policies and proposed interventions have been developed and were included in the simulation environment for this study. Deployment of emergency health care centers, ranging from a single well-equipped Humvee to an entire federal medical station, was included in the environment and placed nearly 200 mobile centers in appropriate locations over the 72 hours simulated. Emergency broadcasts representing the use of the Commercial Mobile Alert System were included in the environment, and receipt of these broadcasts influenced individuals' behavior (Figure 3). Several independent groups provide “Safe and Well” registries to help disseminate news of an individual's well-being. We expanded on these policies and represented a centralized registry that is accessed and updated by any individual who was initially triaged at a health care center. Evacuation routes and centers have been identified and were represented in the simulation, which guided the destination choices of individuals seeking to evacuate.
The developed simulation environment represented the complexities of a disaster of this extent and the following 72 hours. The overall effects in terms of injuries (roughly 97 000) and deaths (roughly 279 000) were similar to those produced by other studies. In addition, a high-resolution estimate of both place and time of the surviving population was generated. This includes where individuals are seeking health care and their level of injury (Figure 1; see Supplemental Digital Content Movie, available at: http://links.lww.com/JPHMP/A28), the call capacities on each cell tower, traffic flow throughout the area, and many other details.
The demonstration study focused on evaluating the effectiveness of partially restoring the communications infrastructure. As shown in Figure 4, in the partially restored case, more individuals are seeking sheltering and fewer are panicking than in the case with no restoration. In addition, in the first few hours, more individuals are attempting to reconnect with household members due to enhanced situational awareness from emergency broadcasts. Then, as more of them are successful, they switch to health care seeking, evacuation, and aid and assist. These behavioral differences peak in the early stages following the blast (2.5 hours) yet remain significant for the first 6 hours. Individuals who are able to make contact with household members can make plans so they are more likely to shelter in place or evacuate rather than remain ignorant of the fate of their household members and risk potential injury trying to physically reach them by any means possible. Additional analyses (not shown) indicated that this could lead to several hundred additional survivors and decreases the time it takes survivors to leave the disaster area.
This simulation environment provides the capability to explore the complex interplay of policies and interventions over time following a disaster. The demonstration study illustrates the interdependence of human behavior on the communication infrastructure (which, in turn, impacts the use of the transportation network for evacuation, access to health care centers, etc). The study also provides an example for how this tool could be used by emergency planners to inform and evaluate their policies and real-world intervention capabilities. With contributions from a wider range of stakeholders, this framework can provide a path to an evidence-based approach to defining our disaster response policies.
The level of detailed data needed to represent a realistic scenario limits this study and simulation environment. While an extensive literature search was conducted and many outside experts were consulted, gathering this volume of data was time-intensive and still remains incomplete. Furthermore, the ability of simulation to represent complex cascades of interdependent events presents some unique challenges for analysis and study design. As a result, and many of the dynamics captured in these simulations have yet to be analyzed and because of space limitations many of the existing results are not reported.
The advantage of this environment, however, is that it provides a framework for collecting these disparate data sets. By collecting and defining both the resolution and relation to other data layers, the framework provides a natural environment for collaboration. Analysts from different agencies and academics from different institutions can contribute data and/or computations specific to their area of expertise, which is integrated into the simulation dynamics. This approach supports multiple stakeholders and experts in a way that enables innumerable policy studies.
1. Merchant RM, Elmer S, Lurie N. Integrating social media into emergency-preparedness efforts. N Engl J Med. 2011; 365:(4):289–291. doi:10.1056/NEJMp1103591.
2. Lasker RD, Macdonald N, Hébert PC. Fixing the fatal flaw in emergency planning. Can Med Assoc J. 2009; 181:(10):661
3. Lasker RD.. Redefining Readiness. New York, NY: The New York Academy of Medicine; 2004; .
4. Auf der Heide E. The importance of evidence-based disaster planning. Ann Emerg Med. 2006; 47:(1):34–49. doi:10.1016/j.annemergmed.2005.05.009.
5. Castleden M, McKee M, Murray V, Leonardi G. Resilience thinking in health protection. J Public Health. 2011; 33:(3):369–377. doi:10.1093/pubmed/fdr027.
6. Sherman MF, Peyrot M, Magda LA, Gershon RRM. Modeling pre-evacuation delay by evacuees in World Trade Center towers 1 and 2 on September 11, 2001: a revisit using regression analysis. Fire Saf J. 2011; 46:(7):414–424. doi:10.1016/j.firesaf.2011.07.001.
7. Drabek TE, Boggs KS. Families in disaster: reactions and relatives. J Marriage Fam. 1968; 30:443–451.
8. Liu S, Murray-Tuite P, Schweitzer L. Analysis of child pick-up during daily routines and for daytime no-notice evacuations. Transp Res Part A: Policy Pract. 2012; 46:(1):48–67.
9. Perry RW, Lindell MK. Understanding citizen response to disasters with implications for terrorism. J Contingencies Crisis Manag. 2003; 11:(2):49–60.
10. Barnett DJ, Thompson CB, Errett NA, et al. Determinants of emergency response willingness in the local public health workforce by jurisdictional and scenario patterns: a cross-sectional survey. BMC Public Health. 2012; 12:164
11. Smith SD, Smith R, Albanese J, et al. Disaster and exercise performance information collection tool: capturing observations in four minutes or less. J Bus Contin Emer Plan. 2012; 6:(2):151–163.
12. Yeager VA, Menachemi N, McCormick LC, Ginter PM. The nature of the public health emergency preparedness literature 2000-2008: a quantitative analysis. J Public Health Manag Pract. 2010; 16:(5):441–449. doi:10.1097/PHH.0b013e3181c33de4.
13. Lurie N. The need for science in the practice of public health. N Engl J Med. 2009; 361:(26):2571–2572. doi:10.1056/NEJMe0911050.
15. Coleman CN, Knebel AR, Lurie N. Preparing and planning for a catastrophic incident of a nuclear detonation [Foreword]. Disaster Med Public Health Preparedness. 2011; 5:(suppl 1):S11–S12. doi:10.1001/dmp.2011.19.
16. McNally R, Stopford B, Oancea V, et al. Blood and Tissue Gap Analysis for an Improvised Nuclear Device Event. McLean, VA: Science Applications International Corporation; 2010; :1–62.
17. Buddemeir B, Valentine J, Millage K, Brandt I. Key Response Planning Factors for the Aftermath of Nuclear Terrorism. Livermore, CA: Lawrence Livermore National Laboratory; 2011; :1–58. Technical Report No. LLNL-TR-512111/O1049.
18. Beckman RJ, Baggerly KA, McKay MD. Creating synthetic baseline populations. Transp Res Part A: Policy Pract. 1996; 30:(6):415–429. doi:10.1016/0965-8564(96)00004-3.
19. Eubank S, Guclu H, Kumar VSA, et al. Modelling disease outbreaks in realistic urban social networks. Nature. 2004; 429:(6988):180–184. doi:10.1038/nature02541.
20. SALT mass casualty triage: concept endorsed by the American College of Emergency Physicians, American College of Surgeons Committee on Trauma, American Trauma Society, National Association of EMS Physicians, National Disaster Life Support Education Consortium, and State and Territorial Injury Prevention Directors Association. Disaster Med Public Health Preparedness. 2008; 2:(4):245–246. doi:10.1097/DMP.0b013e31818d191e.
21. Guterbock T, Lambert J, Bebel R, Ellis J, Kermer D. Population Behaviors in Dirty Bomb Attack Scenarios: A Survey of the National Capital Region. Charlottesville, VA: Center for Survey Research; 2010; :1–64.