Spatio temporal forecasting of dengue outbreaks using machine learning

Abstract

Dengue is one of the most critical public health concerns in Sri Lanka which imposes a severe economic and welfare burden on the nation annually. Prior work has shown that there are multiple factors that contribute to propagation of dengue, including sociological factors such as human mobility. Therefore, it is a non-trivial task to model the propagation of this disease accurately at a regional level. However, accurate quantitative modeling approaches that can predict dengue incidence for a public health administrative division would be invaluable in allocating valuable public health resources and preventing sudden disease outbreaks. In this study, we make use of large-scale pseudonymized call detail records of approximately 10 million mobile phone subscribers to derive human mobility patterns that can contribute towards disease propagation. We develop 3 distinct proxy indicators for human mobility based on different assumptions and evaluate the suitability of each indicator to accurately model the disease transmission dynamics of dengue. Using the proxy measures developed by us, we go on to show that human mobility has a significant impact on the disease incidence at a regional level, even if the disease is already endemic to a given region. Combining these proxy mobility indicators with other climatic factors that is known to affect dengue incidence, we build multiple predictive models using different machine learning methods to predict dengue incidence 2 weeks ahead of time for a given MOH division. By introducing an automated input feature selection method based on genetic algorithms, we show that we are able to improve the predictive accuracy of our models significantly, with predictive models based on XGBoost yielding the best performance, with an R2 of 0.935 and RMSE of 7.688.