Modelling infectious diseases using public web search data

UCL researchers have developed an online tool, Flu Detector, that estimates daily flu rates based on web search data, which has been adopted by Public Health England.

Prior to COVID-19, the UK Government Cabinet Office Risk Register identified pandemic influenza as highly likely to occur and to have the greatest adverse impact on the country. 

In 2013, the Engineering and Physical Sciences Research Council (EPSRC) funded an Interdisciplinary Research Collaboration in Early Warning Sensing Systems for Infectious Diseases. This enabled Professor Ingemar Cox and Dr Vasileios Lampos from UCL Computer Science to research the next generation of infectious disease surveillance methods based on monitoring aggregated counts of web searches that contained words related to influenza. 

The research improved feature (search query) selection within this context by using developments in statistical natural language processing. That increased the accuracy of estimates of Influenza-like illness rates by between 12% and 28%.

Providing an earlier indicator of epidemics

Subsequently, Dr Lampos and Professor Cox led the development of an online tool, Flu Detector (also known as ‘i-senseFlu’) that displays estimates of flu rates for England based on web search data on a daily basis. The Flu Detector was incorporated in PHE’s suite of syndromic surveillance methods since the 2017/18 influenza season. Data in these reports are used to determine the timing and duration of an influenza epidemic and associated health recommendations, e.g., prescribing antiviral drugs to the elderly. 

In a paper jointly authored with PHE, Dr Lampos and Professor Cox demonstrated that surveillance based on web searches could give an earlier indicator (one to two weeks) of the onset of an influenza epidemic or pandemic. This was relative to a traditional network of sentinel doctors (coordinated by the Royal College of General Practitioners) that report the fraction of patients presenting at practices with influenza-like illness.

Using social media to infer vaccine effectiveness  

A collaboration with PHE and Microsoft Research provided the first estimation using social media data (Twitter) of the effectiveness of a flu vaccination programme. 

PHE initiated a pilot live attenuated influenza vaccine programme for school age children across seven areas in England. Conventional syndromic surveillance systems analysis did not yield statistically significant outcomes for the impact of this vaccination programme.

Dr Lampos and Professor Cox constructed models based on tweets about influenza-like illness (ILI) activity in pilot (vaccinated) and control (unvaccinated) regions. The control regions were used to predict the prevalence of influenza in the pilot regions in the absence of the vaccine. The difference between this prediction and the estimated ILI prevalence in pilot regions was used to infer the vaccine’s strongly positive and statistically significant effectiveness.

This evidence helped PHE to recommend a national vaccination programme for 2- to 7-year-olds (five million children). This is estimated to reduce the prevalence of influenza in the general population (not just vaccinated children) by 20%.

Covid outbreak management  

During the first wave of the COVID-19 epidemic in the UK, PHE requested that the UCL team construct a model based on their previous work on influenza and to submit a weekly COVID-19 surveillance report. The reports were used by policy makers at the national level to make decisions on outbreak management policy. 

Project team: Professor Ingemar Cox, Dr Vasileios Lampos

 

Links

• Professor Ingemar Cox’s academic profile
• Dr Vasileios Lampos’ academic profile
• UCL Computer Science
• UCL Faculty of Engineering Sciences
• UCL Engineering REF 2021 

Image 

• Image credit: iStock / FatCamera