Background: Estimation of the national-level incidence of seasonal influenza is notoriously challenging. Surveillance of influenza-like illness is carried out in many countries using a variety of data sources, and several methods have been developed to estimate influenza incidence. Our aim was to obtain maximally informed estimates of the proportion of influenza-like illness that is true influenza using all available data.
Methods: We combined data on weekly general practice sentinel surveillance consultation rates for influenza-like illness, virologic testing of sampled patients with influenza-like illness, and positive laboratory tests for influenza and other pathogens, applying Bayesian evidence synthesis to estimate the positive predictive value (PPV) of influenza-like illness as a test for influenza virus infection. We estimated the weekly number of influenza-like illness consultations attributable to influenza for nine influenza seasons, and for four age groups.
Results: The estimated PPV for influenza in influenza-like illness patients was highest in the weeks surrounding seasonal peaks in influenza-like illness rates, dropping to near zero in between-peak periods. Overall, 14.1% (95% credible interval [CrI]: 13.5%, 14.8%) of influenza-like illness consultations were attributed to influenza infection; the estimated PPV was 50% (95% CrI: 48%, 53%) for the peak weeks and 5.8% during the summer periods.
Conclusions: The model quantifies the correspondence between influenza-like illness consultations and influenza at a weekly granularity. Even during peak periods, a substantial proportion of influenza-like illness-61%-was not attributed to influenza. The much lower proportion of influenza outside the peak periods reflects the greater circulation of other respiratory pathogens relative to influenza.