Toward global estimates of spatial and temporal transmission of Chagas disease

Estimating spatiotemporal variation in disease exposure is critical to developing cost-effective and targeted strategies to reduce the burden of the disease. As The WHO is expecting to eliminate Chagas disease as a public health problem by 2030, being able to measure the local burden of the disease and the progress made thus far is critical. However, obtaining such information when there is no dedicated surveillance system set up to monitor incidence can be challenging. Cross-sectional prevalence studies provide information on past exposure but cannot be used directly to evaluate the epidemiological situation, especially for long-lasting diseases such as Chagas disease. However, the Force-of-Infection (FoI), i.e., the yearly per-susceptible rate of disease acquisition, can be estimated using age prevalence data and provide insight into the local temporal pattern of the disease. Such methodology, relying on localised surveys, can inform the dynamic of transmission locally, but extrapolating FoI estimates from them to assess the burden across the country requires robust statistical methods. In this thesis, we develop and implement a modelling process to, first, predict FoI in space and time from cross-sectional studies and, estimate the burden of the disease, while appropriately propagating uncertainties. Where such an approach has been used, typically mean or median FoI estimates are used as a dependent variable, ignoring the uncertain nature of such FoI being estimates rather than observations. Therefore, the first objective of this thesis was to account for such uncertainty both when fitting models and evaluating their predictive ability. We implemented a set of comprehensive analyses to assess the impact of this uncertainty on performance, by characterising the ability to estimate accurately the central trends while correctly characterising the level of uncertainty. We, then, compare the implementation and performance of this framework to Machine Learning methods to optimise the methodology. Finally, we, propose a modelling process where the predicted FoIs at a fine spatial resolution are used to estimate the burden of Chagas disease. The process, applied to the 76 serosurveys conducted in Colombia, showed a substantial risk of overconfidence when using median estimates to fit and evaluate models, instead of accounting for the uncertain nature of estimated FoI. Machine Learning methods provided a more flexible and reproducible framework while recentring the uncertainty and are thus better suited to provide good burden estimates. Implementing the modelling pipeline, we estimated that the FoI varied considerably across Colombia, but temporal changes were less marked. Relying on predicted current and past exposure, 506,000 (95%CrI: 395,000-648,000) people were estimated to be infected by T. cruzi in Colombia in 2020, representing a 1.0% (95%CrI: 0.8%-1.3%) prevalence in the general population and leading to an estimated 2,400 (95%CrI: 1,900-3,400) deaths. We estimated a substantial increase in the burden of Chagas disease over time, resulting from the interplay between exposure and demography: a slight decrease in exposure was overcompensated by the large increase in population size and the gradual ageing of the population.