Swiss TPH

Formulation of such models is complicated by many uncertainties. Some of these are statistical, generated by the difficulties acquiring sufficient data. For example, it is not easy to quantify natural variation in exposure to mosquito bites. Other uncertainties, especially those that relate to the human immune response to the parasite, arise because we have not yet been able to measure the relevant variables, or do not know what they are. The lack of a proxy measure for protective immunity is a problem both for vaccine development and for modelling. An important simplification in our strategy so far has been to avoid predicting those intermediate variables whose quantitative relationships with epidemiologic outcomes are very uncertain.

 

Our simulations of Plasmodium falciparum epidemiology use 5-day time steps, starting with hypothesised values for the parasite density at each point in time. The densities are based on a description of the time courses of parasite densities in neurosyphilis patients who received therapeutic P. falciparum infections. These simulated infections are introduced into populations of simulated hosts at a rate dependent on field-based measurements of the EIR. Each host is characterised by a vector of time-varying variables including age, health and immune status. This enables a more realistic consideration of the stochastic interactions between individual hosts and pathogens than the use of compartment models. Further stochastic elements of the model determine whether the simulated host becomes ill or dies.

 

The uncertainty inherent in disease modelling needs to be minimised by ensuring that all elements of the model fit the data as well as possible, consistent with parsimony in the model structure. We fitted different components of the model to a wealth of datasets from many different ecological and epidemiologic settings from across Africa. This approach led to implicit statistical models requiring many repeated simulations to make approximate parameter estimates. We were able to fit these using a simulated annealing algorithm, distributing simulations across our local computer network, using up to 40 processors at any one time.

 

We used this model to investigate the epidemiological effect of the RTS,S/AS02 malaria vaccine in a field trial in Mozambique, and then applied it to predict the likely epidemiological impact of pre-erythrocytic malaria vaccines. To make predictions of cost-effectiveness, we incorporated costing data and a model for the health system currently in place in a low-income country context, based largely on data from Tanzania. This enabled us to use our model to predict the cost-effectiveness of such vaccines.
We found that a vaccine of this kind will be most effective at low transmission intensities, but that alone it is unlikely to reduce transmission very much, except possibly when transmission is already low. Such a vaccine might be most effective if deployed simultaneously with vector control measures. Simulations with a 20-year time horizon of a vaccine introduced via a well-functioning Extended Program on Immunisation into a typical African transmission setting suggest that effectiveness in preventing mortality would be about one-third of the efficacy measured in trials. At a unit vaccine price assumption of US$ 10 per dose, and under other baseline assumptions used in the model, the cost-effectiveness results (US$ 96 per DALY averted, US$ 3,521 per death averted) suggest potential value for money for sub-Saharan African ministries of health. The details of these first-phase models and predictions were published in 15 papers as a supplement to the American Journal of Tropical Medicine and Hygiene (volume 75, supplement 2). The results were also used by the Boston Consulting Group in 2005 for estimates of the global demand and cost of potential malaria vaccines.

 

Vaccines are not the only new innovation in malaria control, and a broad selection of novel malaria control tools are entering the field. Insecticide-treated nets are now in widespread use, but there is controversy about the best way to deliver them. Strategies for intermittent preventive treatment in infants (IPTi), pregnancy and children are also being field tested. New drugs are being developed, but are challenged by the rapid evolution of resistance. Given all these alternatives, analyses are needed to determine the likely impact and cost-effectiveness of an entire range of individual and integrated control strategies.

 

Following the completion of our modelling of pre-erythrocytic vaccines in January 2005, we received support from the Swiss National Science Foundation to extend the modelling to vector control and, within the Bill & Melinda Gates Foundation-supported IPTi Consortium, to model IPTi. Finally, we received overall support from the Bill & Melinda Gates Foundation to integrate all these efforts into comprehensive modelling of a full range of interventions and integrated control programmes in settings characterised by a wider range of transmission patterns. This involves: (i) detailed modelling of health systems, including different delivery systems for interventions; (ii) improving our model, for example by embedding a dynamic model for the characteristics of individual P. falciparum infections, and considering effects of heterogeneity in transmission and in host responses; (iii) comparing predictions of different models where there is substantial uncertainty about aspects of the epidemiology (e.g. decay of immunity); (iv) carrying out probabilistic sensitivity analyses of the results. All these results will feed into a full economic analysis and a comprehensive assessment of uncertainty presented in a policy-relevant form.

 

The second phase of this project will involve fitting many different models via iterative processes, investigating their predictions and carrying out sensitivity analyses. Such modelling is extremely computer intensive, requiring repeated simulations of large human populations with a diverse set of parameters related to biological and social factors. This is particularly the case when relatively rare events (such as death) are included in stochastic simulations and when sensitivity analyses are required.
This power is being provided by volunteers who provide spare computing capacity over the internet via a platform that we set up in collaboration with CERN in Geneva (www.malariacontrol.net). This platform uses the Berkeley Open Infrastructure for Network Computing (http://boinc.berkeley.edu/). Based on prior experience, we expect to complete in a few months — using thousands of volunteer PCs — a volume of computing that would have taken several decades with the computing power previously available to us.

 

In the same way that the field of climate modelling is moving towards making predictions from ensembles of different models, we will be most confident about the predictions if we see that they remain substantially unchanged across a range of different sets of assumptions. To this end, we hope to use the malariacontrol.net platform to implement models proposed by other research groups, taking full advantage of the massive computational power that is available.