BioWar: Simulating Disease Outbreaks using Social Networks
Download 123.03 Kb.
|
- Navigate this page:
- Representation of Disease course
- How symptoms relate to agent behavior
- Current Simulation Experiments and Preliminary Results
- High Alert vs. No-alert conditions
- Table 1. Simulation Variable Values
- Spatial epidemiological data
- Quick Medical Reference (QMR) data
- Looking up Health Information Online (LHIO) data
- Emergency Room Statistics (ERS)
BioWar: Simulating Disease Outbreaks using Social NetworksAbstractThe reality of life is embedded in social networks. At present, most epidemiological models do not consider the heterogeneity of social networks when predicting disease outbreaks. One of the challenges in modeling natural and man-made epidemics is understanding how social networks affect disease propagation and how the consequences of disease changes social networks. In particular, early detection of disease outbreaks or detection of bioterrorism events requires early detection, planning, and a keen understanding of how diseases propagate through social networks and alter agents behavior. It is critical to be able to systematically reason about the rate and location of spread of diseases, early presentation of diseases, potential of media and inoculation campaigns as counter measures, and the relative value of various possible early warning diseases (such as access to web-MDs). There is simply not enough actual data on weaponized attacks to begin to train personnel or to learn from best practices how to detect such attacks. What is needed is a cost effective, ethical system for reasoning about such events, for planning responses, and for gaining intelligence on what such attacks are likely to look like in early stages. Estimating the impacts of large scale biological attacks and the efficacy of containment policies is necessary from an intelligence and planning perspective. In this paper, we describe a simulation system called BioWar which uses cognitively realistic agents embedded in social, knowledge and work networks to describe how people interacting in these networks acquire disease, manifest symptoms, seek information and treatment, and recover from illness. Using a model of diseases and symptoms, agents who come in contact with infectious agents through their social and work networks become ill. These illnesses alter their behavior, changing both the propagation of the disease, and the manifestation of the disease on the population. BioWar is a city level multi-agent network model of weaponized biological attacks. In building BioWar advanced artificial intelligence techniques, such as autonomous agents and machine learning, are combined with state of the art tools from the social sciences, i.e., social network methods, current spatial epidemiology models, and empirical data drawn from diverse sources. This combination enables greater accuracy and flexibility in modeling the behavior of people and organizations. Presently, we have completed a number of simulations which examine the effect of contagious and non-contagious illnesses in high-alert (agents have knowledge of a potential disease outbreak) or low alert states. Agents who believe they may be ill and have knowledge of a potential outbreak are more likely to seek care than those who do not. We have compared results of low alert states to known influenza epidemics and to data containing emergency room visits, pharmacy purchases and absenteeism. Although the peak incidence of the simulated outbreak is larger than the peak incidence seen in the population data, the simulation results are temporally similar to those seen in the population data. Further work to increase the fidelity of both the simulation and the population data is on-going. It is hoped that this simulation framework will allow us to ask "what-if" questions regarding appropriate response and early detection strategies for both natural and man-made disease events. This simulation framework will also allow the more precise, timely, and localized modeling of naturally occurring disease “noise” versus the bioattack “signal”. This is especially crucial to enable early detection within the first critical 48 hours. Secondary data sources such as the sales indicators of drugs, orange juice, absenteeism would help in this respect. As the simulation framework gets refined in greater detail, it would become feasible to use it as a bioattack detector – in addition to as a simulator -- customized to a particular city/area. IntroductionThroughout history, the spread of disease has been linked to social and environmental factors. When people live in close proximity to each other, the risk of disease spread—both infectious and non-infectious—increases. Public health measure and aggressive vaccination policies have reduced the risk of disease spread, and in some cases, have eradicated diseases altogether. Smallpox for example, has been eradicated from the natural population through an aggressive and systematic vaccination program. However, public health officials remain vigilant to control and prevent disease outbreaks. Recent attention has turned to bio-terrorism and the risk that terrorists might use weapons of mass destruction to infect vulnerable populations. In these scenarios, the social networks in which we live can become the vehicle through which disease is spread. How best to detection and respond to a man-made disease outbreak is not always clear. The goal of our research is to develop tools to systematically think about, model and analyze how diseases would spread through socially connected groups. In trying to prepare for attacks, policy makers need to be able to think through the consequences of their decisions in various situations. Consider, for example, trying to decide if all US citizens should be vaccinated for smallpox. Speculations abound as to the potential devastation that smallpox could wreak. Medical experts, scientists, and policy makers need a way of thinking through the morass of complex interconnections to understand whether different inoculation or containment strategies will be effective. Unfortunately many existing models are quite limited in that they only apply to a single disease, discount factors such as the urban geography which can influence disease spread, or discount how people use their social networks (who is a friend with whom) to pass information such as when to go to the doctor to be treated. In general, being able to estimating the impacts of large scale biological attacks and the efficacy of containment policies is necessary from an intelligence and planning perspective and requires reasoning about social response and disease processes as a complex social system. While conventional epidemiology has achieved significant successes in managing diseases and epidemics, the approach is inadequate in dealing with the high noise to signal ratio in case of bioattacks where the focus is early detection. Part of the reason for this is that conventional epidemiology does not inherently account for spatial, geographical, and social dimensions in its mathematical modeling of diseases. There is a recent progress in spatial epidemiology which takes into account the spatial and geographical dimensions directly [Lawson 2001][Gimblett 2002], however it does not consider the social dimension nor the intricate interplay between different dimensions. The interplay would affect the shape of signal fed into detection systems. Moreover, the organizational, media, government/institutional action dimensions also play significant roles in the event of bioattacks. To aid the analyst and decision maker we are developing BioWar. In BioWar we are combining state-of-the-art computational models of social networks, communication media, disease models, demographically accurate agent models, wind dispersion models, and a diagnostic error model into a single integrated model of the impact of an attack on a city. Unlike traditional models that look at hypothetical cities, in BioWar the analyst can examine real cities using census, school track, and other publicly available information. Moreover, rather than just providing information on the number of infections, BioWar models the agents as they go about their lives – both the healthy and the infected. This enables the analyst to observe the repercussions of various attacks and containment policies on factors such as absenteeism, medical web hits, medical phone calls, insurance claims, death rate, and over the counter pharmacy purchases. Analysts could use BioWar to ask and answer “what if questions” of the form “what would happen to this city if three people returned from vacation with smallpox?” Moreover, these questions could be answered from the person’s perspective, the perspective of the organization where he or she works, and the city’s perspective. Social NetworksThe world we live in is a complex socio-technical system. Although, social, organizational and policy analysts have long recognized that groups, organizations, institutions and the societies in which they are embedded are complex systems; it is only recently that we have had the tools for systematically thinking about, representing, modeling and analyzing these systems. These tools include multi-agent computer models and the body of statistical tools and measures that have arisen in social networks. Using these tools we can begin to think about these complex socio-technical systems in which people live, and in which diseases can propagate. Such an illustration is particularly salient in lieu of the tragic events of October 2001 when anthrax was distributed through the US postal system. Over the past few years, major advances have been made in the ability to model and analyze social networks and graphs of agent interactions. Techniques including the P* family of tools, and the graph level metric tools (such as that for clustering networks) can be used to locate patterns. Among the data that can be represented as graphs are interaction or communication networks, monetary networks, inter-organizational alliances, mental models, texts, web pages, who was present at what event, story lines. The combination of these techniques with machine learning is likely to be especially powerful for locating anomalies, assessing coherence, and locating underlying fundamental patterns. Multi-agent network models, if based on known information about general or specific characteristics of groups, can suggest general or specific guidance about how to affect or protect the underlying group, organization or society. Exactly what these models can address depends on the model. For example, food poisoning is often discovered through analysis of social links and shared experiences. More worrisome, infectious bioagents spread through the social networks of people who are infected. While non-contagious bioagents such as anthrax do not spread through social networks, it affects the behavior of social networks and more importantly social networks would facilitate piecing up together the puzzle of non-contagious attack (e.g., if anthrax is released at a stadium, sport-fan networks would show a malignant pattern). Since BioWar has imbedded within it agents that are linked through social, work, and knowledge networks, it allows us to recognize patterns with the network that may suggest a disease outbreak or bioterrorism event has occurred. BioWar is an effort to develop a scaleable and precise simulation tool to examine disease propagation and agent behavior in response to disease and illness. We believe it will serve to help researchers understand, predict, and analyze weaponized biological attacks at the city level and engage in “what-if” analyses to help inform decision-making in this complex socio-technical policy domain. For example, it can be used in a “what-if” mode to examine the impact of and response to various weaponized attacks for contagious and non-contagious diseases under high-alert and no-alert conditions. BioWarThe following sections describe the components of BioWar and the underlying rationale of the disease, network, and agent models. Graphically, these components are illustrated in Figure 1. 
Figure 1. The design of BioWar. Input characteristics of the agents, diseases, geography, and communication technology are used by the agent model to change behavior. The raw output can be used to test early detection algorithms, and hypothetical what-if scenarios used to analyze policy impact. Disease ModelIn constructing our disease model, we used historical accounts of known anthrax releases, documents from the October, 2001 bioterrorism attacks, and known disease knowledge bases. The disease model considers risk factors (to identify subpopulations at more risk of developing a disease), symptoms, diagnostic tests, and disease progression to describe how a bioagent release would affect an individual, and focuses on how these individual characteristics affect behavior and the data that might be collected from non-routine sources. We have drawn on the experience of other medical expert systems developed to assist in diagnosis to ground our disease model in well-founded medical knowledge representations [Miller 1982]. 
Representation of Disease courseWe divide the disease symptoms into three groups, divided by time course: those symptoms and risk factors that predispose a person to development of a disease, those symptoms that might prompt a visit to the doctor or pharmacist, and those symptoms or tests that would help with diagnosis or may affect mortality and morbidity. Pre-existing risk factors affect who has a higher probability of getting a disease, and will use information about agent demographics during simulation to afflict agents with disease. For example, an elderly or immunosupressed patient may be at a higher risk of contracting anthrax if exposed. Once infected, the patient then begins to acquire symptoms as their disease progresses. This affects patient behavior in the simulation. For example, a patient with multiple, non-specific symptoms might be more likely to go to a physician or visit a pharmacy. When a patient has developed enough symptoms to warrant further investigation, a portion of the patients who are ill will seek medical care. It is at that point a patient may have additional tests performed, use more healthcare resources, and likely have their information recorded in databases that serve as a source for early detection. This progress is shown in Figure 2. Risk factorsAs we discovered in the anthrax releases in October 2001, certain demographic groups are more likely to be susceptible to particular diseases than other. These risk factors increase a person’s susceptibility to diseases either through host factors, or environmental factors to which that person is exposed. For example, individuals who have contact with animals (sheep shearers, for example), are more likely to contract cutaneous anthrax than other occupations. Risk factors are distributed to individuals in the population, according to demographic characteristics, and we can determine these risks based on age, sex, and disease prevalence. SymptomsSymptoms are important in our model in that they motivate behavior changes. People with symptoms will stay home from work, will visit their doctor or pharmacist, and will change their patterns of interacting with others. This change in interaction patterns leaves a “footprint” on the data streams used to detect bioterrorism events. It is this manifest pattern of symptoms that enables early detection of bioattacks. The symptoms have two different measures that influence which symptoms a person gets and how that changes their behavior. The frequency measure is a qualitative measure of how frequently people with a particular disease will manifest this symptom. The evoking strength is a qualitative measure of how frequently someone with a particular symptom will think they have a particular disease. For example, the symptom of cough may have a very high frequency among patients with anthrax—however cough would have a low evoking strength because people with a cough do not immediately think they have anthrax. FrequencyFrequency is a number between 1 and 5 which answers the question: “In patients with disease x, how often do you see symptom y?”. For example, patients with the diagnosis of anthrax will have a fever frequency of 5 – nearly all patients with anthrax will have fevers at some point in the course of their disease. Evoking StrengthEvoking strength is a number between 0 and 5 which answers the question: “When you see symptom y, how often do you think the cause is disease x?” For example, fever symptoms are not specific to any one disease – in our disease profile of anthrax, fever is given a ES of 1. However, widened mediastinum is a more specific manifestation of anthrax – in patients who have a widened mediastinum, the diagnosis of anthrax should be considered. Evoking strength is similar to specificity. How symptoms relate to agent behaviorThe frequency and evoking strength associated with individual symptoms are used in different ways within BioWar. Although individuals will get symptoms based on the symptom frequency for the disease that they have, they will actually alter their behavior based on the evoking strength of that symptom for a serious illness. For example, if a person contracts an anthrax infection and displays fevers and chills, in the absence of a known attack, may consider their symptoms to be an influenza infection or simple cold and not significantly alter their behavior. However, if they began heaving shortness of breath, chest pains, or other symptoms more suggestive of a serious infection, they would likely stay home from work, seek medical attention or go to an emergency room. Diagnostic testsDiagnosis is simulated based on the apparent symptoms. It produces certain accuracy in the form of ROC curve similar to the real diagnosis. The result of diagnosis would determine whether a person is treated properly and recovered. To determine which disease a person has, the evoking strengths among potential diseases are compared and the highest one is chosen as the diagnosed disease. Current Simulation Experiments and Preliminary ResultsFigure 3 shows a potential view of the results of a BioWar simulation. Within BioWar, it will be possible to examine the impact of weaponized attacks for the city as a whole, particular geographic sub-locations, the entire population, or specific sentinel populations such as school children, military personnel or police for all diseases or for a specific disease. F  igure 3. An interface into BioWar. This illustrates the GIS and graphical capabilities that BioWar will have to display simulation information.
In addition, a comprehensive set of information will be tracked by location and time that will include over-the-counter drug purchases, web access, symptoms, accurate and inaccurate diagnoses, physician and emergency room visits, deaths and recoveries. Each type of information will be capable of being displayed separately or in aggregate for the population as a whole or by sentinel group or physical sector. (It is also possible to track a simulated health & life history of a particular kind of person.) For example, over the counter purchases are displayed in Figure 4. 
Figure 4. Illustrative BioWar Output for Over the Counter Drug Purchases Virtual Experiment
In the following section, we discuss how knowledge of the alert conditions—high alert versus no-alert conditions affect the behavior of agents and ultimately the outcomes and mortality of the populations. This is an example of how policy decisions and response strategies can have a very significant effect on the outcomes of a bioterrorism attack. For our initial experiments, we assume that in the states in which there was not a high state of alert, patients would not immediately seek medical care, and physicians and health care providers would not consider the illness to be a bioterrorism attack. In these settings, treatment is delayed and is less effective, and patients would be more likely to self-treat their symptoms. In conditions in which there is a state of high alert, both patients and healthcare provides are more likely to assume their illnesses represent a potential bioterrorism attack, and seek early intervention at a clinic or emergency room. Healthcare providers have a higher index of suspicion that patients may have been exposed, and treatment is more effective and more efficient. Figure 5 and Figure 6 show non-alert versus high-alert conditions for a non-contagious disease like anthrax. In both cases, the number of exposed and infected patients is similar. However, in the non-alert conditions, more patients seek over-the-counter medications and fewer are seen by healthcare providers. The result is a higher death rate with non-alert conditions than with alert conditions.
Figure 5. No-alert condition: people self-treat themselves by going to pharmacies, resulting in more deaths. Figure 6. High-alert condition: people go directly to ER, resulting in more people treated and recovered and less deaths compared to that of no-alert condition (though the number of deaths is still high due to the severity of the anthrax attack). All agents within the simulation are socially situated and tend to associate with similar or like-minded individuals. Agents that share characteristics will tend to associate with other agents who are like them. For infectious diseases, the social network takes on more significance—those people infected will tend to associate and infect other agents within their social network. We run virtual experiments for infectious/contagious bioattacks in addition to non-contagious bioattacks. In one of these experiments, we simulate the occurrences of bioattacks on the background of influenza and staph pneumonia. The population of Pittsburgh is approximated by a representative sample of 2000 people, with the corresponding demographics ratios (age, gender, and race). The length of the simulation is 90 days. The background diseases will have varying degree of base rate (the susceptible proportion), transmissivity, life span, and time of occurrence. There are a total of 10 strains of influenza and 10 strains of staph pneumonia. The staph pneumonia is assumed to be community acquired and mild; it has the duration of 7-10 days. The bioattacks will occur at day 15, during which bubonic plague bioattack – which is contagious -- is released, and days 20, 30, 40 during which cutaneous anthrax is released. All background diseases and bioattacks have associated symptoms, evoking strengths, and frequencies data as contained in the QMR database. QMR is a diagnostic decision-support system with a knowledge base of diseases, diagnoses, findings, disease associations and lab information. QMR contains information from the primary medical literature on almost 700 diseases and more than 5,000 symptoms, signs, and labs. More concisely, the following table shows the parameters of the simulation. Table 1. Simulation Variable Values
ResultsFigure 7 below shows the result denoting the general statistic of health, counting the number of people infected, died, cured, and immune. Additionally, it gives supplemental statistics useful for the early detection of bioattack: the number of people who got sick & go home, go to pharmacy, go to doctor, and go to ER. 
Figure 7. General Statistics of Health over 1 Contagious Bioattack and 3 Non-Contagious Bioattacks. As Figure 7 shows, after the Bubonic Plague happens on Day 15, the number of people who died rises. Then on Day 20, an Anthrax attack occurs, which further increases the number of people who died afterwards of bioattacks. However, when the second and the third Anthrax attacks occur, the number of people who died does not increase as much afterwards. This is due to an increased people’s awareness of bioattacks and stockpiling of antibiotics like Cipro. This could be seen in the graph by “cumulative death” line. Figure 8 below shows the contribution of Bubonic Plague. Figure 8. Bubonic Plague Contribution
Figure 9 below shows the contribution of Cutaneous Anthrax. 
Figure 9. Anthrax Contribution As Figure 9 shows, while after the first anthrax attack there is an increase in number of death, during the second and third anthrax attacks there is fewer number of death. Figure 10 below shows the manifested symptoms during the time. 
Figure 10. Symptom Manifestations The above symptoms are combined symptom manifestations of background disease strains of influenza and staph pneumonia, and of bubonic plague and anthrax bioattacks. Data SourcesThere are several categories of data needed for generating BioWar simulation. As described earlier, BioWar simulation is complex; it involves sociological, geographical, spatio-epidemiological, technological, demographical, financial, and psychological data. It also involves issues of government, authority, and the media. Categories of data:
All the above factors interplay with each other resulting in a cumulative contribution to the fidelity of BioWar simulations. Not all of the above have been implemented, but work is progressing to include more and better data in BioWar. General Social Survey (GSS) This GSS data set describes the general and comprehensive social characteristics such as work status (working full time, working part time, unemployed, in school), occupation (transportation worker, farmers, military people), work hours, medication taking habits, who to talk to in time of need, family networks (sibling, brothers/sisters, single mothers), working in a small or big firm, Internet use, commuting time, etc. GSS in the Year 1985 describes the social network. Thus GSS data set describes the general characteristics of a heterogeneous population. The source of GSS we use is GSS 1972-2000 Cumulative Datafile, published by the National Opinion Research Center (NORC), with Inter-University Consortium for Political and Social Research (ICPSR). GSS is used to parameterize agents in biowar simulations to reflect the people in the real world. For example, the working time in the simulation is modeled after the real world data from GSS, resulting in a much fewer people work at weekends and during the night. If a bioattack occurs at night, it has a different effect than if it occurs during the working hours of most people. Other GSS data, such as the Internet use, provides estimates to what people would be able to effectively use during/after a bioattack. Census 2000The Census 2000 data describes the demographic characteristics of a population, income level, area profiles, birth & death rates, migration rate, estimate of population growth, among others. The source of this census data is Census 2000 from US Census Bureau. This census data set is used to parameterize agents with appropriate demographics for a certain area as agents have the spatial dimension. It will also be used to provide the natural birth, death, and migration rates of agents. ArcGISArcGIS data set describes the characteristics of geographical area. It provides maps, charts, views, and graphical user interfaces for geographical information systems. The source of ArcGIS data set is ESRI, in addition to publicly available mapping data from USGS. Additionaly, census tract boundary files, school district boundary files, and ZCTA (ZIP Code Tabulation Area) from Census Bureau are used.
BigFred DatasetsBigFred datasets describe the drug purchase behaviors of a population in a particular area. These datasets come from UPMC. We use the datasets to validate BioWar simulation. Spatial epidemiological dataSpatial epidemiology (Lawson 2001, Gimblett 2002) is the study of epidemiology in time and space. There is an increasing quantity of spatial epidemiological data. Spatial epidemiological data sets are used to parameterize, compare, and validate BioWar simulations. For background influenza disease, spatial epidemiological data for influenza will be used to parameterize agents having influenza. Quick Medical Reference (QMR) dataQMR data set describes the interrelationship between diseases, symptoms, diagnosis, and treatments, among others. QMR itself is diagnostic decision-support system with a knowledge base of diseases, diagnoses, findings, disease associations and lab information. With information from the primary medical literature on almost 700 diseases and more than 5,000 symptons, signs and labs. The source of this data set is FirstDataBank, Inc. The QMR data set is used to parameterize the disease models on the relationship between diseases and symptoms. Hospital dataHospital data describes the capacity of a hospital, the patient admission rate, and the number of nurses and physicians. This hospital data will be used to parameterize hospital abstract agents in BioWar simulations. The hospital data is useful for determining the capacity of care in case of a bioattack and for researching an improvement to the health system with respect to bioterrorism. Emergency room dataEmergency room data describes the critical care capacity of health systems. This emergency room data will be used to parameterize the critical care abstract agent in BioWar simulations. Looking up Health Information Online (LHIO) dataLHIO data describes the percentage of people looking up health information on the Internet based on their demographics (gender, race, age). This data set is useful to parameterize how agents behave online after enduring certain symptoms or after knowing a bioattack occurred. Emergency Room Statistics (ERS)ERS data describes emergency room visits based on demographics. This data set is used to parameterize how demographics affect agent’s decisions to go to Emergency Room. Doctor Visits (DV) DV data describes doctor visits based on demographics. This data set is used to parameterize how demographics affect agent’s decisions to go to a doctor. Purchase of Over-the-Counter Drugs (OTC) OTC data describes the over the counter drug purchasing rate based on demographics. This data set is used to parameterize how demographics affect agent’s decisions to purchase over the counter medicine. LimitationsThe current implementation is a work in progress. It has the following limitations:
ExtensionsIn addition to fixing the limitations mentioned in the previous section, we are extending the model to include:
REFERENCES[Gimblett 2002] Gimblett, H. Randy, “Integrating Geographic Information Systems and Agent-based Modeling Techniques”, Santa Fe Institute, Oxford University Press, 2002. [Lawson 2001] Lawson, Andrew B., “Statistical Methods in Spatial Epidemiology”, John Wiley & Sons Publisher, 2001. [Miller 1982] Miller RA, Pople HE, Myers JD, “Interist-I, An Experimental Computer-based Diagnostic Consultant for General Internal Medicine”, N Engl J Med 1982, 307:468-76. APPENDIX: PICTURES 
Directory: publications -> papers publications -> Swiss Federal Institute of Technology (eth) Zurich Computer Engineering and Networks Laboratory publications -> Quantitative skills publications -> Multi-core cpu and gpu implementation of Discrete Periodic Radon Transform and Its Inverse publications -> List of Publications Department of Mechanical Engineering ucek, jntu kakinada publications -> 1. 2 Authority 1 3 Planning Area 1 publications -> Sa michelson, 2011: Impact of Sea-Spray on the Atmospheric Surface Layer. Bound. Layer Meteor., 140 ( 3 ), 361-381, doi: 10. 1007/s10546-011-9617-1, issn: Jun-14, ids: 807TW, sep 2011 Bao, jw, cw fairall, sa michelson papers -> English: The Lightest Weight Programming Language of them all papers -> Causes of failure in it telecommunications networks papers -> Rare detections of North Pacific right whales in the Gulf of Alaska, with observations of their potential prey Download 123.03 Kb. Share with your friends:
|
