Temperate climates regularly witness a seasonal influenza epidemic. However there is considerable variation in the timing and severity of these epidemics. Availability of accurate forecasts would allow epidemiologists to develop appropriate responses ahead of time, potentially saving lives and reducing the burden placed on the health system. Moreover, the ability to forecast from a model based on the underlying transmission process will assist in determination of the effectiveness of various intervention strategies.
Since early use in the Apollo space program of the 60’s, Bayesian filters have been used in weather forecasting, econometrics, robotics, audio/visual signal processing and target tracking, and we are now applying these techniques to influenza surveillance data to produce near real-time forecasts of epidemic activity. In preliminary work, we have used a mechanistic model of influenza transmission (SEIR) with a bootstrap particle filter to (retrospectively) forecast the prevalence of influenza in Victoria (Australia) from 2006–14 based on Google Flu Trends data. We have been able to accurately predict the timing of peak incidence up to 4-6 weeks prior to its occurrence.
In current work – in partnership with Defence Science Technology (Commonwealth of Australia), Office of Health Protection, Department of Health (Commonwealth of Australia) and Department of Health (Victoria) – we have been producing weekly forecasts of the 2015 influenza season in Melbourne (Victoria) based on syndromic and lab-confirmed data. Our algorithms will be extended to provide Australia-wide forecasts in 2016 and beyond, making use of a broad range of data streams that monitor influenza activity.