Using approximate Bayesian computation and machine learning model selection techniques to understand the impact of climate on seasonal influenza-like illness in Australia

Abstract

Influenza-like illness exhibits a strong seasonal cycle in temperate climates, with a peak of varying intensity appearing each winter. However, the driving force of this seasonal cycle remains poorly understood. We develop stochastic epidemic models and a model selection framework to understand influenza-like illness seasonality, with the basic reproduction number R0 being dependent on climate. We test four Functional Forms of transmissibility based on three different climate variables and select the best Functional Form for climate-dependent transmissibility via modern Bayesian machine learning model selection methods. By analysing a unique dataset comprising ten years of General Practitioner-reported influenza-like illness surveillance data from Adelaide, Brisbane, Perth and Sydney, Australia, we explore the relationship between influenza-like illness transmission and weather across Mediterranean and subtropical climate zones. We find that absolute humidity has the strongest impact on seasonal influenza-like illness, with two different Functional Forms both based on absolute humidity best describing influenza-like illness in Mediterranean and subtropical climates. Finally, we consider the problem of forecasting the timing of peak influenza-like illness using ensemble modelling techniques. We employ two score metrics and four techniques for calculating ensemble weights in a prototype ensemble forecasting framework. By implementing this method to predict the peak week of influenza-like illness in 2014 in each of the four different locations, we find that forecasting peak week from the start of the year is a challenging exercise providing mixed conclusions on the best training methods, with two approaches – traditional and prototype – producing comparable results. We find again that absolute humidity appears to be a strong factor in the seasonality of influenza-like illness, and find that random forests are a useful tool in informing ensemble forecast weights.