Conference is now cancelled

Probabilistic dengue forecasting using model super-ensembles

Felipe J Colon-Gonzalez (London School of Hygiene and Tropical Medicine)

Introduction: South-east Asia reports more cases of dengue than anywhere else in the world. Vietnam is particularly affected, with over 90,000 cases reported per annum. Dengue in Vietnam is characterised by strong seasonality and substantial inter-annual and spatial variation. The sensitivity of dengue to climate variations offers potential for forecasting dengue risk ahead of time by using statistical models driven by seasonal climate forecasts. The diagnosis and reporting of cases typically have a delay that may vary across time and space, preventing the timely planning and execution of control measures by health practitioners. Reliable predictions could help decision-makers and planners design and implement interventions in high-risk regions.

Methods: We developed a climate-driven dengue forecasting system that generates monthly estimates of dengue risk across Vietnam at the province level in real time. Multiple dengue model formulations were tested using a spatiotemporal Bayesian hierarchical model framework. The best performing models were selected to form a super ensemble dengue forecasting model. Probabilistic dengue forecasts with lead times up to six months were produced using a state-of-the-art seasonal climate forecasting system.

Results: We tested 2,268 different model specifications. The best performing five models were selected to create the super-ensemble. The use of the model super-ensemble resulted in more accurate forecasts than using any individual model. Compared to a naïve model based on seasonal averages, incorporating forecast climate data improved the predictive performance of the super-ensemble.

Discussion: Probabilistic seasonal dengue forecasts of the onset and peak of the transmission season may contribute to improved preparedness and timely resource allocation. The framework presented here is flexible enough to be applied to other settings and to predict other climate-sensitive diseases.