Elucidating user behaviours in a digital health surveillance system to correct prevalence estimates

Knowing the true prevalence of an infectious, but not always fatal, respiratory disease like influenza comes with many challenges. Online participatory health surveillance systems can help reach more individuals and capture the mildly symptomatic than hospitals and GPs, but come with their own biases related to human behaviour. Using Bayesian statistical models, we can correct disease prevalence estimates using information on the prior reporting behaviour of the user, and incorporating the effects of having symptoms on self-reporting of those same symptoms.

We find that having symptoms is a strong predictor of reporting these symptoms, but this effect is weaker if it is only another household member and not the respondent that has symptoms. This increase in probability of reporting decays over time, as the weeks since reporting symptoms increases. Accounting for this kind of user behaviour has a substantial effect on estimating disease prevalence.

This framework could be applied to other digital participatory health systems where participation is inconsistent and sampling bias may be of concern.

Project Researchers