Interdisciplinary modelling and forecasting of dengue

Understanding the past, current, and future dynamics of dengue epidemics is challenging yet increasingly important. To date, many techniques across statistics, mathematics, and machine learning have provided us with quantitative tools for studying dengue epidemics. Here, using data from provinces in northern Peru across 2010 to 2021, we provide a new interdisciplinary pipeline that draws on a new and existing techniques to provide comprehensive understanding and robust prediction of dengue epidemic dynamics.

Wavelet analyses can unveil spatiotemporal patterns in epidemic dynamics across annual and multi-annual time periods. Here, these included climatic forcing and greater spatial similarity in large outbreak years. Space-varying epidemic drivers included climatic influences and shorter pairwise distances driving greater epidemic similarity in more northerly coastal provinces. Then, using a Bayesian model, we can probabilistically quantify the timing, structure, and intensity of such climatic influences on Dengue Incidence Rates (DIRs), while simultaneously considering other influences. Recognising that a single model is generally sub-optimal for any forecasting task, we demonstrate how to form trained and untrained probabilistic ensembles for forecasting dengue cases in settings reflective of real-world conditions. We introduce a suite of climate-informed and covariate-free deep learning approaches that leverage big data and foundational time series, temporal convolutional networks, and conformal inference. We complement these modern techniques with statistically principled training and assessment of ensemble frameworks, while explicitly considering strong benchmark models, computational costs, public health priorities, and data availability limitations. In doing so, we show how ensemble frameworks consistently outperform individual models across space and time, and produce sharp and accurate forecasts with robust, reliable descriptions of uncertainty. We report interpretable classification metrics for detection of outbreaks to communicate our outputs with the wider public and public health authorities.

Looking forward, whether the objective is to understand and/or to predict epidemic dynamics, our modelling pipeline can be used in any dengue setting to robustly inform the decision-making and planning of public health authorities.