A longitudinal model for the spread of bovine tuberculosis

Bovine tuberculosis (bTB) is an important disease of cattle and badgers with substantial socio-economic impact in the UK, currently costing the exchequer over £100 million per year in surveillance and compensation and also resulting in costly movement and trade restrictions for farmers. Despite intensive controls, disease incidence is still increasing. Currently herds are monitored for the disease through slaughterhouse surveillance and through regular skin testing. The frequency of routine testing for an individual herd is based on localised incidence of the disease, which acts as a proxy for risk of infection, but does not account for individual herd-level characteristics or cattle movements. Recent bTB research has focussed on examining potential underlying causes for this, including environmental contamination (e.g. re-infection from local wildlife reservoirs), insensitivity to the surveillance test and the impact of large-scale cattle movements. It is the purpose of this proposal to extend our recent work identifying markers for the persistence of infection in individual herds into a dynamic longitudinal framework in order to quantify the mechanisms of transmission in the GB national herd and to test the utility of our results as an aid to risk-based surveillance. The dynamics of transmission of bTB infection can be represented by a model with transmission driven by chance processes, with an observation process that is governed by an imperfect test procedure (or slaughterhouse identification of visible lesions), leading to partially hidden infection. Herds that contain one or more reactors are classified as breakdowns, which then have movement restrictions and more rigorous testing imposed until the herd tests clear. Testing and cattle movement information is available through several large national datasets. Recent mathematical modelling approaches have been developed using these data and, while these will provide useful information on population-level parameters, they average out some detailed information available at the individual herd level. Also, they were not designed to predict disease recurrence at the individual-herd level. Here we propose to build a dynamic, statistical, individual-herd level model, based on continuous surveillance data, which we will fit to the data using a likelihood-based approach. The main methodological challenge will be to deal with the hidden states (infection) and the movement of animals between the herds. Recent advances in statistical methodology, such as 'data-augmented' and 'reversible-jump' Markov chain Monte Carlo allow the joint distribution of the observed and hidden states to be estimated simultaneously along with key infection related parameters. We will explore an exciting alternative called 'sequential filtering'. The main challenge is that these statistical techniques are computationally intensive, especially given the large scale (approx. 130,000 premises) and long time frame (6+ years) of the datasets. However, advances in computer processing technology, such as architectures for running algorithms in parallel on graphics cards, provide an exciting and cost-effective way to approach this problem. The focus here is on bTB, but these sorts of models and the estimation issues that we will address are relevant to a wide range of infectious disease systems, and the methodology developed in this project would be applicable to a range of disease systems. It is the aim of this project to elicit information about the hidden states of the system from the test observations using robust statistical methodology, in a way that allows us to identify high-risk herds based on the past history of infection, as well as on localised incidence and connectedness to other premises. This information would have a practical use in terms of targeting specific herds with more stringent or more regular testing.

Currently the incidence of bovine tuberculosis in Great Britain is increasing, despite over £100 million per year being spent in surveillance and compensation. Herds are monitored for the disease through slaughterhouse surveillance and through imperfect routine skin testing. Our proposal is to build a longitudinal statistical model to account for within- and between-herd transmission of bTB using information from the VetNet, Vebus and Cattle Tracing System data sets. The aim is to understand better the mechanisms of transmission and ultimately to be able to identify herds with a high-risk of harbouring undetected infection, based on the past history of infection as well as localised incidence and connectedness to other premises, and also to characterise areas or premises of high- or low-risk relative to that due to the cattle network. The system is characterised by a hidden dynamic non-linear infection process overlaid with a probabilistic observation mechanism. We intend to explore the development of data-augmented Markov chain Monte Carlo or sequential filtering methods to fit the model. Localised (environmental) transmission will be dealt with using spatial and non-spatial random effects. A surveillance model of this nature would have a practical use in terms of targeting specific herds with more stringent or more regular testing, and if successful could have substantial impacts on slowing the spread of the disease. However, these methods are computationally intensive, given the large scale and long time frame of the datasets. Recent advances in parallel processing technology, both on multi-core CPUs as well as on graphics processing units (e.g. CUDA/OpenCL) provide an exciting and cost-effective way of approaching this problem. Although the focus here is on bTB, the computational problems of fitting models to large-scale data sets in a robust manner are relevant to a wide range of infectious disease systems.

The main aim of this project is to develop a longitudinal statistical risk-based surveillance model for bovine tuberculosis (bTB) in the UK, which will aid better understanding of the mechanisms of bTB spread in the GB national herd, and could be used to help better optimise targeting of available control resources. Given the huge socio-economic costs of the disease and the increasing incidence, this information may help to increase the efficacy and cost-effectiveness of control, and hopefully might help slow the spread of infection in the national herd. As such it would be of immediate potential benefit to both Defra and Animal Health (AH), and in the longer term to farmers and all those affected by the disease. Animal movements will be modelled explicitly and localised environmental transmission will be modelled through the use of spatio-temporal random effects. While this does not model explicit mechanisms for localised spread, for which there is little or no data available nationally, it will instead highlight areas or herds for which there is a higher or lower risk of infection or persistence than can be explained by cattle movements alone. This in itself would be useful information that would inform further studies in specifically identified areas in order to ascertain the reasons behind these observed differences. The efficacy of the model as a predictive tool will be assessed using both simulated and real data. If these results show potential practical worth, then regular dialogues with Defra and AH will help to hone the optimal way in which the model could be used in practice. A user-friendly interface will be developed (probably web-based, using open-source software) in conjunction with Defra, VLA and AH. A requirement would be that new information could be entered into the model as data becomes available, with the results output in a manner that is straightforward to implement in the field. This technology and the science behind it would be disseminated at various different workshops such as those run by AH. This work involves development of cutting-edge statistical methodology for fitting complex non-linear dynamic models to large scale data sets. With the impacts of infectious diseases becoming more widely realised and technology for data collection improving rapidly, this research has potentially important consequences for developing and fitting statistical models to a wide range of dangerous infectious disease systems, and importantly could help to add to the suite of real-time modelling tools available for modelling epidemic outbreaks.