Bayesian spatio-temporal modelling for malaria surveillance and residual pockets of transmission identification in Swaziland

Malaria control has been and is in the world spotlight for over 50 years and had been marked by a proliferation of research studies, as interest in finding new control methods increased. Along the full spectrum from malaria control to prevention of reintroduction emphasis on surveillance and response has been made if gains in the fight against the disease are to be realized. International funding for anti-malaria related activities has also been up-scaled and sustained with a significant amount of it focusing on the high burden countries in the sub-Saharan region of continental Africa. Understanding the complex interactions between malaria vectors, parasites and human hosts is key to control and elimination of the disease. Geostatistical methods involving the use of remote sensing (RS) techniques and geographic information system (GIS) tools have proven to be an effective way of estimating spatial and temporal effects of environmental determinants on disease outcomes. They also allow us to produce model-based maps which could be used to predict the disease at explicit geographic scales thus aiding targeted control.
In the context of surveillance, preparedness and response we explored potential methods and tools that could be used for surveillance by malaria control programmes in very low endemic settings like Swaziland. In this country, malaria has drastically declined and the country is currently in its elimination stage as it entered the critical 3-year phase from 2015 to 2018 where it is anticipated that it will receive certification from the World Health Organization (WHO) as a malaria free country. Spatially explicit maps on micro-epidemiological heterogeneities as well as space and time trends and patterns in malaria transmission are needed to aid the country to target and prioritize interventions in this critical phase as it deals with individual episodic cases. Currently achieving malaria elimination remains operationally challenging due to the ever present threat of imported cases from nearby endemic regions and from uncensored immigration. Also the turnaround time from data collection, processing and use for planning purposes is too long for rapid response actions. Therefore a rapid response surveillance system is needed in order to achieve elimination and prevent reintroduction after elimination.
Chapter 1 presents the overall background informing this study including the rationale for undertaking this PhD work. The role of surveillance in malaria control and elimination as well as the importance of rapid response in malaria elimination were also presented. We showed the progress the country has made from the establishment of the malaria control unit in the 1940s to present time. Our study focused on the use of environmental data for disease surveillance. Therefore we detailed the environmental factors associated with malaria transmission and demonstrated how they were interlinked with disease incidence. Such factors included temperature, precipitation and humidity. The use of earth observation (EO) data derived from RS techniques was also presented. Tools that could be used to support surveillance such as GIS and global positioning systems (GPS) are also discussed. We look at the current malaria situation in Swaziland with emphasis on the latest developments following the scaling up of malaria interventions in that country.
In chapter 2 we emphasised the importance of mapping potential vector breeding sites in Swaziland using high resolution remotely sensed data in conjunction with entomological data to aid larval source management (LSM) strategies. We used larva scooping methods to identify potential breeding sites in the country and those identified were fed into a decision tree induction algorithm and a logistic regression to assess which environmental factors characterised larvae presence or absence. Both approaches reliably distinguished between the two set of scenarios of larvae presence or absence and identified the same environmental predictor related to human activity (subsistence farming) as key determinant of potential vector breeding. Models linking presence of larvae with high resolution land-use variables were found to have good predictive ability. Thus we produced a map of predicted potential breeding sites at explicit geographic scales to assist the malaria control programme in planning its LSM budget.
There are many environmental proxies that have been proposed by ecologists and remote sensing experts which have a potential for use in vector-borne disease mapping. However, their uptake by epidemiologists has remained notoriously slow. Therefore in chapter 3 we investigated the litany of available RS variables that could be used in vector-borne disease mapping studies. We reviewed literature on available sources of remotely sensed data and presented a library of supplier processed variables and those that need to be derived by the end-users and processed at different levels before being incorporated into disease mapping studies. We discussed the reasons and criteria used to select the proxies described and presented in our catalogue. Indices investigated were limited to those related to EO data products with continental or global coverage scales, and were grouped according to meteorology, land use/cover, cartography and urban mapping variables which could be used as proxies for disease suitability mapping. We found numerous indices that have been derived by ecologists and remote sensing experts from the various satellite sensors that have been launched over the years. However, they have remained underutilized in epidemiology partly because of lack of remote sensing skills needed to derive them and partly because they were not high demand variables and therefore not provided by remote sensing agents and suppliers of remotely sensed data.
In chapter 4, we explored different scenarios for malaria incidence risk by investigating the environmental effects of weekly distributed lags in Swaziland. A Bayesian geostatistical model based on polynomial distributed lags function was developed to assess how different environmental and socio economic factors influenced malaria incidence in the country. We then produced model based spatially explicit maps of predicted malaria incidence risk which could be used by the control programme to target their control interventions for high impact.
In chapter 5, we evaluated some of the new and potential indices for epidemiological studies by testing and comparing their use in predicting malaria incidence risk in Swaziland. We discussed the inclusion criteria and choice of the selected variables for malaria incidence risk prediction in the country. This was necessitated by the fact that new satellites have been launched with much improved sensor capabilities than previous first generation sensors. Sensor improvements are noticeable in the number of spectral bands, spatial and temporal resolutions, thus presenting unprecedented good image sources for identification of spatial heterogeneities, trends and patterns in disease mapping by epidemiologists. We ended with emphasising the importance of why this research work was carried out including discussing the key findings and overall message that came from this study. The contributions that had been made by this study are also discussed as well as remaining research work that could be undertaken as follow up.