Predicting Malaria Risk from Diverse and Multilevel data
Dr Tim Lucas
Oxford University, UK
Abstract:
Malaria causes hundreds of thousands of deaths annually despite major reductions in incidence. High-resolution, risk maps are a vital tool for policy makers to optimally distribute the resources needed to combat this disease. As many countries are now aiming to eliminate malaria, risk maps are needed in areas of relatively low malaria incidence. However, traditional data and models are ill-suited for estimating risk in these areas.
To model malaria risk in low-incidence areas, we use aggregated data collected at the national and state level by governments and health ministries. To create risk maps, we must learn from this aggregated malaria data but make predictions at a much higher resolution. This task requires a number of techniques with broad applicability outside of epidemiology. We use stacking to combine information from multiple, diverse datasets, disaggregation models to learn pixel level risk from aggregated data and Bayesian, hierarchical geostatistical models to leverage spatial information. Throughout, we have made novel innovations to both the models and how the models are fitted. Overall, this modelling exercise demonstrates the utility of data science in both the academic and policy sphere.
Biography:
Dr Tim Lucas is a research scientist at the Big Data Institute, Nuffield Department of Population Health at the University of Oxford. His current research involves developing geostatistical methods for the creation of malaria risk maps. These results feed into the Global Burden of Disease study run by the Institute for Health Metrics and Evaluation, University of Washington, USA. This work involves novel statistical methods for making high resolution malaria risk maps, from low resolution data and methods for combining different types of malariametric data. During his PhD at University College London, he worked on mathematical models of zoonotic pathogens in bats (such as Ebola, Nipah and SARS). More recent work has included software (zoon) for predicting the spatial distribution of wild animal and plant species, in a reproducible way.