Using sequential data assimilation to forecast influenza outbreaks from the Royal College of General Practitioners (RCGP), Research and Surveillance Centre (RSC) network data
The Royal College of General Practitioners (RCGP), Research and Surveillance Centre (RSC) has produced weekly reports on the numbers of communicable and respiratory disease in England for over 50 years. Current data comes from around 200 practices from over 2.5 million patient records and over 2 billion consultations. The data is presented in a weekly report that simply presents the collected data from GPs. However, RCGP RSC does not carry out any forecasting. We report how data assimilation techniques concepts from weather forecasting can be used to forecast the number of presentations of influenza like illness (ILI) at practices in the RSC network over a winter influenza season.
We take all first presentations of ILI since August 2003 and apply a particular data assimilation technique, namely particle filtering (PF), for each winter influenza season using the standard susceptible-infection-recovery (SIR) model for communicable disease. For a brief period during the start of the respective influenza seasons, we apply the PF and then forecast for the remainder of the season to produce a series of fan-charts for the expected number of presentations and a 90% probability window. We will also present forecasts made for the winter 2018/19 influenza season made in real-time.
For the past four years, the DA/SIR method was able to successfully forecast the number of ILI cases presented at the RCGP practices within a 90% highest probability interval up to 4 weeks in advance of the peak of the flu outbreak. Post-peak, the forecast variance collapses and the forecast becomes highly reliable. However, for years 2011/12 & 2013/14, there was a double-peak in the flu outbreak and the DA/SIR method was unable to forecast the second peak.
This forecast system is the first step to moving the RCGP RSC’s weekly respiratory and communicable disease reports from a “data-in-data-out” report to a “data-in-forecast-out” report that will help with planning, and the drug supply chain analysis across England. For regular single-peak flu outbreak years, the DA/SIR method can provide reliable flu outbreak forecasts up to 4 weeks in advance of the peak. At peak and post-peak outbreak the forecasts become highly reliable in forecasting the expected number of presentations.