Comparing large-scale computational approaches to epidemic modeling: agent-based versus structured metapopulation models
Marco Ajelli, Bruno Gonçalves, Duygu Balcan, Vittoria Colizza, Hao Hu, Jose J Ramasco, Stefano Merler and Alessandro Vespignani
BMC Infectious Diseases 2010, 10:190.

In recent years, two major classes of methodologies emerged in the large-scale and spatial spreading simulation of influenza-like-illnesses (ILIs) and other emerging infectious diseases. The first one is the very accurate epidemic description with agent-based models, which keep track of each individual in the population in an extremely detailed way. The second scheme relies on metapopulation structured models that considers in a detailed way the long range mobility scheme at the inter-population level while using coarse-grained techniques at the intra-population level. It is clearly important to assess the level of agreement that the two different approaches can provide on the quantities accessible in both cases and the respective data needed and computational costs associated.

Snapshots of the epidemic evolution in GLEaM (top) and in the agent-based model (bottom) at three different timesteps of the simulation with R0=1.9. Maps report the average number of cases at the resolution scale of the Italian municipalities.

The paper by Ajelli and co-workers contains the first side-by-side comparison of the results obtained with an Agent-Based model and metapopulation approach offered GLEaM. The two models are carefully calibrated in order to simulate an epidemic described by the same natural history and key parameters.  The country used for the study is Italy, a large European country that provides the necessary geographical and population heterogeneity to assess the models performance in the case of highly structured populations. For the sake of clarity, the two models consider a hypothetical influenza pandemic event for which the same parameterization and initial conditions in the far east. Both models, despite the difference in the data integration and model structure, provide epidemic profiles with spatio-temporal patterns in very good agreement.

The good agreement of the two approaches reinforces the message that computational approaches are stable with respect to different data integration strategies and modeling assumption.The presented results hint to the possibility of combining the two methodologies in order to devise multiscale approaches that use the data parsimony of the metapopulation approaches at the global level and the high resolution of the agent-based model in specific locations of interest where detailed data are available.

The GLEaMviz Simulator is a software system with an intuitive and flexible GUI for the simulation of emerging infectious diseases spreading across the world, that we developed during the last 2 years. The software system levers on GLEaM, and its design maximizes flexibility in the definition of the disease compartmental model and in the configuration of the simulation scenario, allowing the user to set a variety of parameters, from compartment-specific features, to transition values, to environmental effects. The output of the simulation is then provided in terms of a dynamic map visualization and sets of charts to quantitatively describe the geotemporal evolution of the disease.

Download the GLEaMviz Simulator and explore your global epidemic simulations!

The software is based on a Client-server system. The Client can be installed on the user’s local machine, and it allows the user to setup the simulations that will be executed on the server, thus avoiding specific requirements for large computational capabilities on the user’s side. The Client consists of four principal components: 1) the main window with the Simulations History; 2) the Compartmental Model Builder; 3) the Simulation Wizard; and 4) the Visualization Windows. The main workflow and the role of the components are outlined in the diagram in the Figure below.

Overview of the workflow of the GLEaMviz Simulator.

The GLEaMviz Simulator Client uses the Adobe AIR 1.5 runtime and can thus be installed on recent versions of the following operating systems: Windows, Mac OS X,  Linux. The Public Edition of the client available from http://www.gleamviz.org/simulator/ is pre-configured to use the GLEaMviz server made available by gleamviz.org. There is thus no need to install the server in order to use this client. However, in order to avoid an overload on this server, a number of limitations are enforced in this setup. Research groups interested in an unlimited version of the GLEaMviz system are invited to contact us at info@gleamviz.org.

Check out all the software features at the GLEaMviz Simulator webpage!

For this purpose, we developed the GLEaMviz Simulator, which is a software that allows the simulation and visualization of the spread of epidemics at the worldwide scale. It allows the user to specify and remotely execute simulations, and retrieve and visualize the results. The Simulation Engine is based on GLEaM, and the Client is made of different components that enable a customizable compartmentalization to describe a specific infection dynamics, the set up of a specific simulation initialization (e.g. time and place of seed event in the world, presence or not of prior immunity, presence of seasonal effects, etc.), the remote execution of GLEaM simulations and the visualization of the results in forms of maps and plots for different geographic areas and levels of resolutions chosen by the user.

The Software architecture ensures great flexibility and maintainability as well as high speed in large-scale simulations. The Client is user-friendly and meant to be used to produce global simulations based on GLEaM with no specific knowledge on coding/modeling needed.
Have a look at this video presenting an overview of the GLEaMviz Simulator.

GLEaMviz Simulator Overview from GLEaMviz on Vimeo.

Developing the GLEaMviz simulator faces us with several computational challenges and requires to analyze the specific use of the simulator in the context of different decision making processes  (crisis rooms, public health planning, research etc.).  For this reason we are partnering with the JOINT RESEARCH CENTER (JRC) of the European Commission at Ispra, Italy. The Joint Research Centre is a research based policy support organisation and an integral part of the European Commission. The mission of the JRC is to provide customer-driven scientific and technical support for the conception, development, implementation and monitoring of EU policies. As a service of the European Commission, the JRC functions as a reference centre of science and technology for the Union. The objective of the collaboration is to share tools, including the GLEaM paltform to improve models for crisis management and decision making of emerging public health threats. We are proud to announce that the first version of the GLEaMviz Simulator has been released to JRC where usability tests have been started.

What is keeping us busy: we are now fine tuning the Software to include additional details and improve its usability.

We plan to post soon the first public release of the Client of the GLEaMviz Simulator. More info will soon be available on this website. Stay tuned!

Epidemic Planet - explore how H1N1 influenza travels around the world and how intervention measures may help

Epidemic Planet  is the visualization application developed in the context of the GLEaMviz project, which enables its users to interactively compare and learn about the effect of a number of intervention scenarios on a pandemic, simulated using GLEaM. Its first public appearance was at the 3 months long INFECTIOUS Art & Science exhibition at the Science Gallery, Trinity College, Dublin, Ireland.

In collaboration with the British Society for Immunology, Epidemic Planet lands at the Edinburgh International Science Festival. In this installation, visitors will observe the evolution of the 2009 H1N1 influenza pandemic since its early origin in Mexico, and will discover how the air traffic and commuting flows determined the spreading patterns of the flu worldwide.

Moreover, through a large screen coupled with a touch-screen interface, visitors will be able to explore different scenarios of propagation of the H1N1 pandemic according to different initial conditions (what if the flu started in Edinburgh? what if it were more contagious?) and different intervention measures. In this way, they will learn how the flights ban or the early distribution of vaccines could have affected the spreading of the pandemic.

The venue of the Epidemic Planet exhibition will be Hawthornden Court, in the National Museum of Scotland. Check out the programme of the Festival!

Come and visit the Epidemic Planet in Edinburgh!

Physics World, February 2010 issue.

The Flu Fighters.

A special feature is dedicated to infectious diseases, how they rapidly spread in our modern society, and what weapons we have nowadays to fight them.  The article titled The Flu Fighters by Vittoria Colizza and Alessandro Vespignani describes the contribution of physicists to an interdisciplinary area where complex systems are the main ingredients, and modeling human behavior and biological contagion is the ultimate challenge. The cover of the special issue shows an illustration by B. Goncalves et al. of the multiscale worldwide mobility networks used in the GLEaM model.

PNAS cover image.

In the issue of December the 22nd of the Proceedings of the National Academy of Sciences, we publish a paper that discusses the interplay of human mobility patterns like those between local metropolitan commuters and long-range airline travelers during a global epidemic. The image of the worldwide mobility network constructed in our paper has been featured in the cover of the journal.

Multiscale mobility networks and the spatial spreading of infectious diseases.
D.Balcan, V. Colizza, B. Gonçalves, H. Hu, J. J. Ramasco, A. Vespignani
Proc Natl Acad Sci U S A 106, 21484-21489 (2009).

In the paper we detail the definition of the worldwide multiscale mobility network at the core of the Global Epidemic and Mobility (GLEaM) model and discuss the data integration process and the statistical analysis that allow its construction.

Have a look at this video presenting an overview of the GLEaM model.

We also tackle some general theoretical questions that concerns the basic understanding of the spatial spread of infectious diseases on the large scale:
i)    Is there a most relevant mobility scale in the definition of the global epidemic pattern?
ii)    At which level of resolution of the epidemic behavior a given mobility scale starts to be relevant and to which extent?

In order to fully consider the effect of multiscale mobility processes we first integrate data of commuting patterns in five different continents with the airline transportation database and then develop a time-scale separation technique for evaluating the force of infection due to different mobility couplings and simulate global pandemics with tunable reproductive ratios. The results obtained from the full multiscale mobility network are compared to the simulations in which only the large scale coupling of the airline transportation network is included. Our analysis shows that while commuting flows are, on average, one order of magnitude larger than the long-range airline traffic, the global spatio-temporal patterns of disease spreading are mainly determined by the airline network. Short-range commuting interactions have on the other hand a role in defining a larger degree of synchronization of nearby subpopulations and specific regions which can be considered weakly connected by the airline transportation system.

It also is possible to show that short-range mobility has an impact in the definition of the subpopulation infection hierarchy. In other words, global disease outbreaks tend to touch down at major travel hubs, generally major airport locations and spread out like a wave that follow local commuting patterns. The findings of the paper open the path to quantitative approximation schemes that calibrate the level of data resolution and the needed computational resources with respect to the accuracy in the description of the epidemics.

The techniques developed here allow for an understanding of the level of data integration required to obtain reliable results in large scale modeling of infectious diseases and have already have contributed to the improvement of the computational model we use to provide estimates and projections of the H1N1 pandemic.

Modeling the critical care demand and antibiotics resources needed during the Fall 2009 wave of influenza A(H1N1) pandemic
D Balcan, V Colizza, AC Singer, C Chouaid, H Hu, B Gonçalves, P Bajardi, C Poletto, JJ Ramasco, N Perra, M Tizzoni, D Paolotti, W Van den Broeck, A-J Valleron and A Vespignani.
PLoS Currents: Influenza. 2009 Dec 4, RRN1133.

The most common symptoms of influenza are generally mild for the majority of people infected. However, in a small portion of clinical cases infected with influenza, the disease can lead to complications of increasing severity requiring medical attention, antibiotics, and, in more serious situations, hospitalization and intensive care. Given the limited capacity of health care providers and hospitals and the limited supplies of antibiotics, it is important to predict the potential demand on critical care to assist planning for the management of resources and plan for additional stockpiling.

In the knol, we introduce a model that considers the development of influenza-associated complications and incorporate it into a GLEaM to assess the expected surge in critical care demands due to viral and bacterial pneumonia. More specifically, the new compartmentalization adds to the basic structure of the influenza dynamics a set of compartments and transitions taking into account the possible evolution of the complications associated to an influenza infection, including viral and bacterial pneumonia, and different speed of progression and stages of severity of the disease. It includes home treatment, hospitalizations, and admission to intensive care unit (ICU). Patients in each stage of pneumonia complications are assumed to be treated with antibiotics, with a preferred empirical antibiotic regimens based on the guidelines issued by the British Thoracic Society. A sketch of the complete compartmental model can be seen in the figure below (click on the left panel to zoom in).

Sketch of the compartment relations for the new epidemic model

Time evolution of the ICU occupancy in a set of countries, measured as the predicted need of ICU beds per 100,000 persons.

Based on the most recent estimates of complication rates, we predict the expected peak number of intensive care unit beds and the stockpile of antibiotic courses needed for the current pandemic wave. The effects of dynamic vaccination campaigns (see this post), and of different length of staying in the intensive care unit are also explored. The right panel of the figure (click on it to expand it) shows the predicted ICU occupancy as a function of time for four countries of the Northern Hemisphere. The three profiles per each country refer to the predicted ICU occupancy in the baseline case when no intervention is implemented, and in case dynamic vaccination campaigns with distribution rates rv=0.1% and rv=1% are considered. Solid curves correspond to the median profiles and the shaded areas to the 95% reference range obtained from 2,000 stochastic simulations. The average ICU length of staying is assumed equal to 7 days.

Tables 1 displays further details on these ICU predictions for a baseline situation and when a vaccination campaign with distribution rates rv=0.1% is considered. The predicted need of ICU beds at peak are typically moderate even when the baseline scenario without intervention is considered, ranging approximately from 5 to 7 ICU beds per 100 000 inhabitants, well below the average national capacity of ICU beds per 100,000 inhabitants. Such numbers can be reduced if measures as vaccination campaigns are taken into account.

Table 1: Predicted need of ICU beds in a scenario with a vaccination campaign covering 0.1% of the population per day until end of vaccine stockpile. The 95% reference range (RR) of the daily number of occupied ICU beds per 100,000 is reported at its peak for several countries in the Northern Hemisphere.

Table 2 reports the number of antibiotics courses needed daily at the peak of the requests, and the total size predicted to be used at the end of the pandemic wave, based on the empirical guidelines of the British Thoracic Society and broken down by the stage of severity of pneumonia. A single course of antibiotics is defined as the combination of antimicrobial drugs considered in the treatment regimen for the suggested duration (see the knol for additional details). The total size of antibiotics courses predicted to be used in the current Fall 2009 pandemic is in the range of [6,337-7,149] per 100,000 for the set of countries explored, which needs to be compared with the available stockpiles of antibiotics courses to cover high-risk groups. Many countries however do not possess nation-wide antibiotic supplies, as antibiotics are generally available through short supply chains able to fulfill average just-in-time requests. The estimates contained in Table 2 can therefore be considered as guidelines to assess the expected needs during the remaining evolution of the pandemic wave with respect to the present usage pattern and available resources.

Table 2: Predicted usage pattern of antibiotics in the scenario with the previous vaccination campaign. The 95% RR of the daily number of administered antibiotics courses per 100,000 at its peak is reported, along with the total amount predicted to be administered by the end of the pandemic wave. Results are shown for several countries in the Northern Hemisphere, broken down for different stages of influenza-associated complications. Pneumonia stages I, II and III corresponds to home-treatment (or supervised outpatient treatment), hospital wards and ICU, respectively.

Estimate of Novel Influenza A/H1N1 cases in Mexico at the early stage of the pandemic with a spatially structured epidemic model
V Colizza, A Vespignani, N Perra, C Poletto, B Gonçalves, H Hu, D Balcan, D Paolotti, W Van den Broeck, M Tizzoni, P Bajardi, JJ Ramasco. PLoS Currents: Influenza. 2009 Nov 11:RRN1129.

Determining the number of cases in an influenza epidemic represents a great challenge. Reliable figures for the actual number of cases are key to properly assess several parameters, such as mortality, morbidity or hospitalization rates. This is particularly relevant during the early phase of an outbreak, in order to better inform decision making process. Surveillance of cases inevitably suffers of several biases overall leading to an underascertainment of influenza cases. For example, many cases showing mild symptoms might not seek for medical attention, and would therefore not be included in the counting of cases. Moreover, the monitoring of cases is expected to change with time. After the initial alert at the beginning of an outbreak, an enhanced surveillance system is expected to have a large monitoring capacity. However, this situation changes with time, due to the large increase in the number of cases and the majority of resources being dedicated to the severe patients. This leads to an ascertainment mainly focused on the most severe cases, with the number of confirmed cases being a gross underestimation of the actual number of people infected by influenza. A significant example is provided by the time evolution of the monitored cases in Mexico during the first months of the epidemic: two studies [1] [2] have assessed the size of the epidemic in the country and both of them found a significant underascertainment of the confirmed cases as reported by the mexican authorities. [3]

In order to address this point we use GLEaM to simulate the epidemic spreading and compute the number of cases in Mexico at the end of April and the beginning of May. We calibrate our model with a maximum likelihood estimate of the infection parameters, fitting the empirical arrival dates of the first infected in the countries seeded from Mexico (a detailed description of the estimate procedure is reported here). This allows us to provide an ab-initio calculation of the number of cases in Mexico independently by the estimation precedure. In the Table below the number of infectious individuals obtained in our simulations is reported, comparing it with the same number obtained in Refs [1][2] and with the confirmed cases reported by mexican authorities.

	Number of symptomatic cases in Mexico (Apr. the 30th)
Simulation Results	[121,000 - 1,394,000]
Lower bound range of Ref. [2]	113,000-375,000
Estimate of Ref. [1]*	2,000 – 280,000
Mexican official report [3] (confirmed cases)	3,350

Predictions of GLEaM for the size of the epidemic in Mexico on April 30 in thousands of cases and comparison with other approaches and with empirical data. The simulations are obtained with our infectious parameter estimates (a detailed description of the estimate procedure is reported here) and show the 95% reference range over 2,000 stochastic realizations. The results are compared with the lower bound estimate range in [2], the estimate provided in Ref. [1] and the number of confirmed cases given by official reports [3]. *The interval provided for Ref.[1] is obtained by merging the results reported in the paper under different assumptions and including the 95% CI.

Despite the different approximations used here and in Refs. [1][2], the three approaches are providing support to the possibility of a reporting ratio of infected cases in Mexico as low as 1 in 100. This finding is important when evaluating the massive amount of data which are now being collected in a large number of countries around the world. We can easily imagine that the reporting rate as well as the estimate of the cumulative attack rate in most of the countries could be easily underestimated in similar cases.

[1] Fraser C, Donnelly CA, Cauchemez S, Hanage WP, Van Kerkhove MD, et al. Pandemic Potential of a Strain of Influenza A (H1N1): Early Findings. Science 2009, 324: 1557-1561.

[2] Lipsitch M, La jous M, O’Hagan JJ, Cohen T, Miller JC, et al. Use of Cumulative Incidence of Novel Influenza A/H1N1 in Foreign Travelers to Estimate Lower Bounds on Cumulative Incidence in Mexico. PLoS ONE , 2009 4: e6895.

[3] Secretaria de Salud, Mexico. Situation actual de la epidemia, Oct 12, 2009. http://portal.salud.gob.mx/sites/salud/descargas/pdf/influenza/situacion_actual_epidemia_121009.pdf

Modeling vaccination campaigns and the Fall/Winter 2009 activity of the new A(H1N1) influenza in the Northern Hemisphere
P Bajardi, C Poletto, D Balcan, H Hu, B Goncalves, JJ Ramasco, D Paolotti, N Perra, M Tizzoni, W Van den Broeck, V Colizza, A Vespignani
Emerging Health Threats Journal 2009, 2:e11.

The paper discusses the effectiveness of the vaccination campaign in mitigating the epidemic in the northern hemisphere, according to the predicted epidemic unfolding. The relative reduction of the epidemic peak activity with respect to the baseline (no-interventions) scenario is measured. Mitigating effects are explored depending on the interplay between the predicted pandemic evolution and the expected delivery and distribution rate of vaccines.

The incidence curves, reported below, show the impact of an incremental vaccination with 1% daily distribution starting on October 15. US and Spain are considered as examples. The model is calibrated using the latest estimates on the transmissibility of the new A(N1H1) influenza, considering as reference the late peak case (the details are reported here). The effect of the vaccination campaign is compared with a combined strategy that includes the systematic treatment of clinical cases with antiviral drugs, where different antiviral distribution rates are considered.

Incidence curves for US and Spain for different intervention scenarios. The gray bar indicates the time period during which the immunization takes effect.

The results show that if additional intervention strategies were not used to delay the time of pandemic peak, vaccination campaigns may not roll out before the pandemic peak is already reached. In the US it is likely that the vaccination campaign will not be able to substantially reduce the epidemic activity, vaccination however will be crucial for the protection of risk groups and healthcare workers. In Europe the activity peak is shifted of a few weeks and the modeling shows that timely vaccination campaigns able to cover 30% of the populations by the second half of November might be effective in mitigating the pandemic. Unfortunately reports of delays in the production and distribution of vaccine have fuelled concern that supplies will arrive too late to make a difference in the number of people that get infected with the new virus. This is again a strong rationale for the prioritized vaccine distribution programs focusing on high-risk groups, healthcare and social infrastructure workers.

The coupled animations show the geotemporal evolution of the swine flu epidemic in the United States, by comparing the observed pattern (left) with the simulated one (right). Simulations refer to the baseline scenario obtained from the maximum likelihood analysis used to estimate the transmission potential of the new influenza A(H1N1) (see this post for more details) [1]. Simulations are not iteratively calibrated, and containment or mitigation interventions are not considered in the model. The maps represent the geographic distribution of the cumulative number of cases at a resolution level of ¼°. Both maps adopt the same color code ranging from a minimum of 1 case per cell (pink) to the maximum number of cases per cell reached in the time frame explored (red). This allows a one-to-one comparison of the two epidemics at each date, highlighting analogies and differences of the geographic spread and of the epidemic activity in terms of number of cases. The data reported on the ‘Real map’ is obtained from official sources and projected at this level of resolution, accessible by the model in the simulations. The data reported on the ‘Simulated map’ is obtained from the average value calculated on 2,000 stochastic realizations of the model. The timeline shown is from May 1, 2009 to July 6, 2009, after which date it was no longer possible to track confirmed cases at this level of resolution.