mv

The project has obtained several technical advancements in the development of models for the spread of an influenza epidemic, in the analysis of relative data, and in the assessment of intervention options.

The main achievements of the project can be summarised as follows:

- the development and improvement of different techniques to estimate generation interval and reproduction number from data from various sources (outbreak investigations, FF100s, general surveillance, sequence analysis);

- a quantification of the role of transmission within schools, and within classes; and thus an assessment of the potential role of school holidays or closures in shaping the infection impact, and possibly mitigating it;

- the collection of available data on past pandemics (1889 to 1968) and the re-estimation of parameters, using a unified approach;

- the analysis of data on seasonal influenza, that has provided some of the first estimates for the strength of seasonal forcing, and for the reporting rate in surveillance systems;

- the development of an individual based model of pandemic influenza transmission for Europe, and of techniques to make it useful in real-time;

- a retrospective analysis of the factors determining the observed pattern of spread of the 2009 H1N1 pandemic in Europe, demonstrating  its partial predictability;

- the preparation, on the basis of data from POLYMOD project, from EUROSTAT and from their modelling, of  contact matrices between age classes that provide an accepted standard for modelling through patch models;

- the comparison between different large scale computational modelling approaches, thus providing the first step towards their integration;

- the development of a novel algorithm to approximate optimal allocation patterns for vaccines and for social distancing, on the basis only of available information up to that time;

- the collection of a large dataset on planned and actual behavioural responses during an influenza pandemic in the general population of 4 European countries. Thanks to running two surveys, the first one when news of the pandemic were widespread but the infection spread was extremely limited in Europe (except for UK), the second one the summer after the main pandemic wave, it has been possible to connect stated intention with actual behaviour, and to test the effect of information on this. These information can be essential for planning communication strategies, and are now being tested in the first epidemic models that include behavioural responses.

Moreover, the activities of the researchers of the project have greatly contributed to the analysis of 2009 A/H1N1 pandemic, showing the capability of an early assessment of epidemiological parameters, the potential of real-time modelling, as well as the aspects that need to be improved. The on-going retrospective analysis of the epidemiological data relative to the pandemic strain from early summer 2009 to winter 2010-11 is going to improve our understanding of the factors most influential in determining the patterns of spread of the infection, and should help in better designing surveillance, planning and responses.

WP 1 To improve characterisation of population contact and travel patterns in models

The survey carried out by the POLYMOD project has provided the first data for several European countries on the pattern of contacts among people of different ages (Mossong et al., 2008). Data from that project have been obtained, and integrated with information from EuroStat and other international sources by the UK HPA partner into a unified database and repository (warehouse) for the use of the consortium, thus completing deliverables of the project. POLYMOD data have been re-analysed to obtain more accurate estimates of age-specific contact patterns, that have been compared with those obtained through independent methods, namely the modelling outputs of agent-based models parameterized through socio-demographic data and Time Use data. The contact matrices produced through these different methods are very similar to each other, giving thus support to their use in predictive models.
Still, some differences can be seen when testing the different matrices against available serological data for Varicella (VZV) and ParvoVirus (B19), with VZV data supporting the use of POLYMOD matrix, while B19 that of synthetic matrices (Iozzi et al, 2010). Thus there is support for the idea that, depending on the transmissibility level of the infection, either the number of different contacts, or repeated exposure, may be the key factor for transmission. Available sources of information on population movement data, in the form of origin-destination flows, have been identified (precisely census data for Denmark and UK; the result of a survey on transportations used for the Italian region of Lombardy) and, as long as licensing issues allowed, added to the database or anyway their use made possible to the consortium. These commuting travel data have been integrated into models of epidemic spread to test the feasibility of using simple ‘gravity’ models in epidemic models. The data warehouse will be kept beyond the project end, at least as far as compatible with license issues, and could support other EU-related activities.

WP 2 To evaluate behavioural responses to epidemics and social acceptance of restriction measures

Analysis of the data from the 1918 Spanish flu in different US cities showed that public health measures affected the size and the duration of the epidemic (Bootsma and Ferguson, 2007). In order to be able to likely assess the impact of public interventions, it becomes important trying to predict the acceptance of control measures, as well as the behavioural changes that may occur in response to epidemics, in particular lethal ones. To this aim, a survey of the general public was planned in the FluModCont project with questions concerning seasonal influenza, and hypothetical scenarios of a pandemic.
The emergence of A/H1N1 pandemic ("swine flu") has prompted us to rewrite the questionnaire in terms of the current influenza pandemic, giving the added opportunity to address the general public in real time during a pandemic. The survey was carried out in the period June-August 2009 in four European countries (Finland, Italy, Romania, UK) through telephone interviews to adults aged 18 years or more. It has be noted that in Finland, Italy and Romania there had been a lot of media attention about pandemic flu, but very few cases had actually been observed, so that answers could not be based on personal experience; on the other hand, the survey in UK was carried out after a first wave had occurred, with specific recommendations issued by health authorities.
A second survey was then carried out in the summer 2010 to assess the population behaviours during the main pandemic wave, and their attitudes afterwards. Carrying these two surveys has offered a unique opportunity of comparing behavioural intentions to actual actions, and changes in attitudes during a pandemic, although privacy issues have prevented us from contacting the same individuals in the two surveys.
Results of the first survey, showed some differences in previous behaviours related to seasonal flu and beliefs and on the level of knowledge about 2009 pandemic influenza between participating countries, while no substantial differences were detected among the four countries with respect to the actual expected behaviours in case of spread of swine flu. However, results from the second survey, showed differences within and among the four participating countries. When comparing results of the two surveys conducted in 2009 and 2010 we observed a significant reduction in those who reported to be willing to vaccinate their children against swine flu if free of charge and recommended by health authorities. Moreover, a significant reduction with respect to the use of antivirals as precautionary measures was reported in two of the participating countries, while the proportion of individuals reporting to taking time off from work/school in case of mild symptoms increased in the remaining two.
In case of need, most of respondents to the 2009 survey from the participating countries would seek health advice about swine flu from the local GP/nurse and other local and national health authorities. However, about half of respondents from UK and Finland also reported the media and internet as a source of information, thus partly explaining the fact that they perceive themselves well informed about A/H1N1 more frequently than respondents from Italy and Romania. Subsequently, during the 2010 survey, respondents from Finland, Romania and UK reported less frequently local GPs and Nurse as the main source of information about swine flu for the benefit of media (proportion of reporting media: 26.9% vs 77.0% in Finland; 7.2% vs 55.8% in Romania; and 16.3% vs 67.6% in UK in 2009 and 2010, respectively). In 2010, the proportion of respondents reporting to be willing to take the antiviral drugs as a preventive measure decreased in Italy, Romania, and UK. More than 80% of respondents, during the 2009 survey, would be available to stay at home for 7-10 days and about 80% to keep children away from large gatherings in case the new influenza would spread, with no significant changes in 2010. During 2009, more than 60% would take time off from work/school. These proportions significantly increase in Romania and UK in the 2010 survey. The two surveys conducted were very useful in showing differences between behavioural intentions and actual actions related to the 2009 pandemic influenza. As reported from other studies from UK and Hong Kong the worry about the possibility of catching swine flu was strongly associated with increased likelihood to perform protecting behaviour with regard to social distancing, and vaccination in both 2009 and 2010 surveys except for Italy, where the actual behaviour reported in 2010 showed how the mild virulence of the pandemic changed the risk perception and consequently the behavioural intentions reported in the 2009 survey.  

WP 3 To develop a suite of models for the spatio-temporal spread of a new influenza pandemic

A major aim of the project is the development and validation of a European-wide modelling environment that could be used as a support for policy decisions. Building on micro-simulation models accounting for household, school/workplace and community transmission already developed (Ferguson et al., 2005), a European model built using detailed socio-demographic information that stress the diversity of European countries (Merler and Ajelli, 2010).
This same model is being used for a retrospective analysis of the data on the A/H1N1 pandemic, on the basis of the parameter estimates obtained from the early analysis of pandemic data (see below), of the volume of travels from US and Mexico to European countries and of the school calendars in the European countries.
The results (submitted for publication) show that the model explains several features of 2009 pandemic in Europe, in particular the fact that two epidemic waves (the first in early summer, the second one in autumn) occurred in UK, while a single wave in autumn-winter occurred in all other European countries. This pandemic, fortunately extremely mild compared to all the ones of the past century, provides an opportunity to test the models that have been used (in real-time and retrospectively); thus, one can place a better motivated confidence in the use of models in future pandemic events.
Real-time estimation and modelling during the 2009 pandemic has been performed by the partners in United Kingdom, Netherlands and Italy. As for the modelling part, scenario analysis developed on the basis of early parameter estimates and of the available models (homogeneous age-structured or agent-based) were used to inform Ministries of Health, and provide a guidance towards possible responses. Judging retrospectively (Ajelli et al., 2011), the errors in predictions were not large (at most, a few weeks in peak time), despite the scarcity of available data. The situation was more problematic in England, where models fitted well the first (early summer 2009) wave, but, due to ignorance of the real number of infections during the first wave, it was difficult to gauge the strength of the first wave. The knowledge gained during this experience, especially the accuracy shown (see below) by early estimates of crucial epidemiological parameters (generation time, reproduction number, age-related susceptibility), makes it more likely that real-time modelling could be useful in the future, overcoming the understandable reluctance to embrace models without a previous validation. In a different but related spirit, a stochastic meta-population (patch) model of global transmission of pandemic influenza previously developed was revisited using final data about 2009 pandemic in UK. The model uses detailed airline and land travel data to link different populations. Coupling importation scenarios with a spatial model of subsequent transmission within the UK allowed us to assess the impact of importation on the early stages of the epidemic. Comparison of model output and data show that, in early stages of a pandemic, the number of people infected and the rate of disease spread are driven by disease importation. Moving towards a predictive perspective, the length of time for which these importations are central to the dynamics of the disease depends on the strength of the public health response.
More technical achievements in this area concern the comparison of models. Precisely, we have compared a fully stochastic model with a deterministic model with stochastic fluctuations added when numbers are small, showing extremely similar outputs (Lunelli et al, 2009), except for the estimate of the probability of a major epidemic, not a relevant problem is the issue is modelling the possible mitigation of an existing pandemic. We have also provided for the first time a side-by-side comparison of the results obtained with a stochastic agent-based model and a structured metapopulation stochastic model for the progression of a baseline pandemic event in a large and geographically heterogeneous country, Italy. The results obtained (Ajelli et al, 2010) show that both models provide epidemic patterns that are in very good agreement, with differences in peak timing on the order of a few days. The relative difference of the epidemic size depends on the basic reproductive ratio, R0, and on the fact that the metapopulation model consistently yields a larger incidence than the agent-based model, as expected due to the differences in the structure in the intra-population contact pattern of the approaches. The good agreement between the two modelling approaches is very important for defining the trade-off between data availability and the information provided by the models; they also suggest the possibility of building hybrid models combining the agent-based and the metapopulation approaches according to the available data and computational resources.

 The dynamics of infection spread is certainly determined by individual mobility. The analysis of early global spread of 2009 pandemic (as well as of other infection, for instance SARS) show that intercontinental spread follows the links provided by flight passengers. However, it is also apparent that local spread is determined by local factors (it is enough thinking of the difference between Southern and Northern hemispheres in 2009) and short-range individual mobility, for which however existing data are scarce. A solution to this problem is the use of the existing data to model individual mobility for instance using a gravity model based on distance and population size, and use the model when mobility data are not available; the fit of gravity models to mobility data is generally very good (citation?). Work has been completed on analysing how well an epidemic model based on the mobility data is approximated by an epidemic model that uses a gravity model based on the same mobility data; the test has been performed on UK and US census data; overall, the fit is good for the UK, but simple gravity models capture US dynamics less well – in  particular the strong inter cities connections on each of the West and East coasts, but weaker connections between the 2 sides of the continent. More elaborate models have been developed which allow for this heterogeneity.  The analysis shows the potential for the use of gravity models, when lacking detailed mobility data, as well as their limitations. 

 

WP 4 To estimate model parameters and test model adequacy using data on seasonal flu and endemic diseases

Key to the potential utility of models is the quality of the input parameters, and the reliability of model assumptions. The main objectives of the project were the development of statistical methods useful for the analysis of spatiotemporal data on epidemics, and their application to existing datasets, concerning especially seasonal influenza, in order to assess the adequacy of existing models to describe actual patterns in epidemic spatiotemporal spread. The spread of H1N1 pandemic shifted the research focus to the analysis, in strict contact with health authorities, of the first data collected concerning the new influenza strain.  The analysis, carried out by the Imperial College together with WHO and the Mexican health authorities, of the data available by the end of April, provided the first estimates of the reproduction number R, of the serial interval (time between symptom onset in an individual and her/his presumed infector), of the dependence of susceptibility on age (Fraser et al. 2009). These estimates have been substantially confirmed by subsequent analyses of the "first few-hundred cases" in UK (Ghani et al. 2009), of the household studies in US (Cauchemez et al., 2009), of the analysis of epidemic spread first in Southern Hemisphere countries and later in several other countries (White et al. 2009; Hahne et al, 2009; Turbelin et al, 2009; Rizzo et al, 2009; Ajelli et al. 2010).
A re-analysis, performed by the group at UPMC, of all datasets for the 2009 pandemic allowing for an estimation of the serial interval (SI, the time interval between the date of symptoms onset in one case and that in its infector) yields an overall estimate of the mean SI of 3.0 days (CI95% [2.4, 3.6]) that can be used as best estimate for further studies. Similarly, a review (Boelle et al, 2011) of all studies focused on the first few months (March-October 2009) of the pandemic shows estimates at the community level (town, region or country) varying between 1.1 and 3.1, with a median value of 1.6, close to initial estimates. In the Netherlands, as in other European countries (except the UK), the reproduction number was smaller than 1 (R=0.5) in the period considered. The largest estimate (R=3.3) was obtained from the analysis of a school outbreak, and thus not quite comparable.
Serological studies have corroborated (Miller et al., 2009; Rizzo et al. 2010) the hypothesis of a higher immunity in older age classes, and have yielded first information about the effective infection rate in the population (Hardelid et al, 2010).
A simple comparison of surveillance data to school calendar in England during the pandemic, shows the importance of school transmission for epidemic growth but reveals little of the detailed dynamical impact schools have on the intricacies of transmission. In a very important collaborative study between the Imperial College group and the US CDC, Cauchemez et al (PNAS, 2011) analysed very detailed epidemiological and social data collected by CDC (with input from IC on study design) from a school and community outbreak of H1N1 in Pennsylvania in May-June 2009. This allowed the investigators to quantify the strength of infection transmission between students of the same school, or of the same grade, or of the same class, and compare these levels to household transmission. In addition, they were able to examine whether friendship networks predicted transmission using detailed social network data collected during the outbreak, and whether gender predicted epidemiological mixing. Friendship groups did not appear to affect transmission, but assortative mixing by sex did affect transmission – the epidemic in boys preceded that in girls by several days.  This work, as well as related analyses of surveillance data from Hong Kong, is essential in order to assess the potential mitigation of impact that can be achieved through school closures, although the effect of unplanned school closures may differ from those of normal holidays.
Using the parameter estimates from these latter type of studies (and prior such work), during the pandemic, the Imperial College group could be used to mitigate the autumn wave of the 2009 pandemic (see WP5 below). 
Another study undertaken by a researcher employed on FluModCont at Imperial College examined the determinants of seasonal influenza transmission, notably the magnitude of seasonal forcing of transmission in temperate countries (Truscott et al, J R Soc Interface, 2011). This work used approximate Bayesian computation (ABC) methods to estimate from seasonal influenza surveillance data the transmissibility of influenza and the extent to which transmissibility varies seasonally in temperate countries. Results showed that strain structure and age-dependent heterogeneity in host population mixing are necessary to explain seasonal influenza dynamics. The reproduction number (R0) at the time of year where transmissibility is maximal is estimated to be approximately 2 (1.5-2.5) and seasonal variation (peak to trough) in transmission is between 10% and 40%.
A related research aim has been a collection and compilation in a single spreadsheet of the available datasets for past pandemics (one of project deliverables). Values of R and of the serial interval have been re-estimated from the original data; this was really necessary as in the literature, before 2008, widely varying figures could be found: between 1 and 20 for the reproduction number, and between 4 and 6 for the generation time.
Best fitting models of data on serial interval estimate an SI longer (mean 2.8-3.1 days) for pandemic than for seasonal (mean 2·2 days) influenza; basing the estimate of the reproductive number R on the same distribution of serial intervals, the median has been estimated to 1·6 (interquartile range 1·5, 1·7), with values much more similar to each other than what available in the literature. The 1889 pandemic was studied on his own using detailed data obtained from diverse sources: it was estimated that the reproduction number was approximately 2, similarly to other pandemics (Valleron et al., 2010). This study provides much narrower bounds about the values of epidemiological parameters that can reasonably be expected from new pandemic strains.
Progress has also been made on the theoretical aspects of parameter estimation. The THL group developed and implemented a software package for the real-time estimation of key epidemiological parameters during the course of an outbreak. This works expands and unifies earlier work on non-parametric smoothing of effective reproduction numbers from longitudinal data on symptom times. The method allows the use of incomplete contact tracing data, and it takes into account all the possible infection routes that could have generated the observations, thus extending its applicability.
Finally a review of existing viral titre data was performed with a view to defining optimal observational protocols to obtain direct estimates of the duration of latency or infectious period. Challenging volunteers with viruses provides valuable data regarding these, but is very rare. It is essential to have good protocols to analyse such data. Modelling provides such an approach, and was indeed adopted here. In parallel, we also developed a direct approach to interpreting viral excretion as the infectivity profile. This analysis (another deliverable of the project) has contributed to the analysis of existing titre data, and would be a basis for designing an optimal protocol for possible new studies.

WP 5 To evaluate the impact of intervention options for containing and mitigating a pandemic influenza outbreak

This WP was designed to evaluate the impact of intervention options for containing and mitigating a pandemic influenza outbreak. Clearly, one aspect that comes before modelling is an assessment of which intervention options are actually considered and feasible for health authorities. Thus, a short questionnaire was prepared and sent to EU technical advisors aimed at identifying the specific pandemic influenza interventions that a selection of different EU member states might be considering before the new A/H1N1 threat. Having received the answers to this questionnaire just prior to the emergence of 2009 pandemic, we took the opportunity of running a second survey in 2010, following the 2009/10 influenza pandemic outbreak, in order to assess the changes in pandemic planning suggested by the 2009/10 experience. The results of this second survey, more extensive and overarching the first survey, are presented in a draft manuscript: "Pandemic influenza preparedness and public health measures in EU, Survey II.
Extensive scenario modelling of anti-viral treatment, of vaccination strategies and of social distancing methods (especially school closures) has been performed, confirming the broadly understood principles that Interventions can indeed reduce impact, and especially delay and flatten peak; that different types of interventions may be synergistic, and that timing of interventions is crucial.
Less intuitive results have been obtained by analysing specific questions:
- the efficacy of interventions based on age-prioritized use of antivirals has been evaluated in terms of cumulative attack rate and excess mortality reduction under different scenarios in a stochastic, spatially structured individual-based model;
- a novel algorithm, relying on an approximation of the contact network structure, has been developed to approximate optimal allocation patterns for vaccines during an influenza pandemic, using only information that can be observed during the pandemic.
-During the pandemic, the Imperial College group undertook an analysis for the U.S. CDC to examine the extent to which reactive school closure using absenteeism based triggers could be used to mitigate the autumn wave of the pandemic. Analysis showed that reactive closure of individual schools was unlikely to significantly impact transmission, but that local area closure using absenteeism based triggers could be effective while minimising the economic impact of school closure inherent in any blanket national closure of schools. This work (Pellis 2009) has been submitted for publication. 

WP 6 To develop efficient, extensible and usable individual-based simulation models

Individual-based simulation models are an important modelling tool for pandemic planning, especially when individually targeted intervention measures (such as anti-viral prophylaxis, or reactive school closures) are considered; in future individual-based simulation models might be used – in conjunction with real-time data analysis – for prediction and to refine control policies in the face of an outbreak. Within the project a very efficient code of the European-wide model outlined above has been developed that allows for rapid simulations on a single work-station; the code is available on the project on the web site, and is accompanied by a short manual on its use (one deliverable of the project); as time allows, the documentation will be improved, to help its use and possible modifications.A computational tool has been developed during the project to add computational steering and visualization front-ends to an existing meta-population patch model for Great Britain. The model allows for different spatial scales and for different types of interventions—contact tracing, case isolation and mass vaccination, for example—.
Computational steering and interactive visualization are clearly useful in this application area as:

1. interactive visualization allows immediate examination of the results of the simulation;-

2. drilling down allows the integration of geo-spatial information at a variety of levels of detail;

3. visualization combined with computational steering allows faster analyses of more sophisticated scenarios and disease models;

4. computational steering combined with interactive visualization allows the computational modeller to judge the merits of an intervention strategy before running a full analysis;

5. computational steering combined with interactive visualization allows the comparison between the currently implemented policy, previously considered policies and the baseline scenario;

6. computational steering allows check-pointing and branching and thus the re-playing of the out-break and testing of alternative policies;

7. and computational steering combined with interactive visualization allows an exploration of the parameters that are internal to the model which can aid debugging and training of new epidemiological modellers.

The development of this model with these additional tools makes epidemiological modelling accessible to audiences not previously regarded as important targets. Now a single user-friendly package could be available to a wide variety of users allowing them to explore spatio-temporal epidemiological models for the first time.
The tools developed for this model could be exported to other modelling methodologies, such as the agent-base models.