A recent feature on real-time tracking of influenza in Nature Outlook discusses how scientists are using social media and online search data to monitor potential outbreaks.
PHE has one of the most advanced public health systems for influenza surveillance in the world, which brings together multiple data sources including GP reports, hospital cases and symptoms reported to NHS111.
Consultant epidemiologist at PHE, Dr Richard Pebody, who leads PHE flu surveillance says “We have been working closely with UCL researchers and are excited to be integrating the i-sense tool into Public Health England national flu surveillance.”
How has i-sense research contributed to this space?
Since the beginning of the i-sense project, our researchers from UCL Computer Science have been focusing on alternative approaches that could complement traditional syndromic surveillance.
They have experimented with content from social media (Twitter) and online search engines (Google, Bing) and have seen that, when appropriately modelled, the latter can be an accurate estimator of influenza-like illness (ILI) rates in primary care in the US or England.
A key outcome of the research is the development of a software system that collects online search data and deploys the previously published machine learning models to compute and visualise influenza rate estimates. This tool, named i-sense flu (formerly known as flu detector), currently provides daily ILI rate estimates for England based on online search data. The model is trained on historical ILI rates obtained by the Royal College of General Practitioners and uses web search data from Google. A dedicated Twitter account (@isenseflu) automatically tweets the most recent estimate on a daily basis.
In an exciting new development for i-sense, these alternative ILI rate estimates will now be included, for the first time, in the weekly UK flu reports by PHE for the 2019/20 flu season. The weekly flu report is a key flu surveillance tool and assists decision making throughout the flu season. PHE has also included i-sense flu’s estimates in their annual flu surveillance report for 2017/18 and 2018/19.
How can i-sense flu complement traditional surveillance systems?
Apart from the timeliness of predictions, including daily availability of outputs and the uninterrupted coverage (unaffected by weekend or holiday closures), online data can represent broader segments of the population who have not consulted the health care system.
The Nature Outlook piece also reports on the potential application of such systems in low- and middle-income countries that may be lacking an established surveillance infrastructure. The existence of historical syndromic surveillance reports is essential in order to train models from online search. Recent research from i-sense researchers at UCL Computer Science has proposed a transfer learning model to map a disease model from one country, where online search and syndromic surveillance data are available, to a country, where only online search data are available.
i-sense Principal Research Fellow at the UCL Department of Computer Science, Dr Vasileios Lampos, says: "The main reason for scepticism about user data-driven methods for estimating influenza prevalence was the inappropriate application of machine learning. Our research has played a key role in re-establishing AI techniques as a complementary tool to traditional methods."
"We are excited about the current developments that highlight the impact of our research. We look forward to expanding our contribution with new models for flu forecasting, in collaboration with PHE.”
As a result of the ongoing evaluation process and given the various indicators of added value, PHE has adopted the current version of i-sense flu so that they can further evaluate how this approach can support their current surveillance systems. This collaboration is an example of how i-sense research outcomes are translated to support active components of England’s public health surveillance system.
- UCL: Prof Ingemar Cox, Jens Groth, David Guzman, Dr Vasileios Lampos, and Dr Bin Zou
- RCGP: Simon de Lusignan
- PHE: Dr Richard Pebody
- Google: Andrew Eland
- Vasileios Lampos, Andrew C. Miller, Steve Crossan, Christian Stefansen (2015). Advances in nowcasting influenza-like illness rates using search query logs. Scientific Reports 5 (12760), doi:10.1038/srep12760.
- Vasileios Lampos, Bin Zou, Ingemar J. Cox (2017). Enhancing feature selection using word embeddings: The case of flu surveillance. The Web Conference, pp. 695–704, doi:10.1145/3038912.3052622.
- Moritz Wagner, Vasileios Lampos, Ingemar J. Cox, Richard Pebody (2018). The added value of online user-generated content in traditional methods for influenza surveillance. Scientific Reports 8 (13963), doi:10.1038/s41598-018-32029-6.
- Bin Zou, Vasileios Lampos, Ingemar J. Cox (2019). Transfer learning for unsupervised influenza-like illness models from online search. The Web Conference, pp. 2505–2516, doi:10.1145/3308558.3313477.
- Sources of UK Flu Data: Influenza Surveillance in the UK – Flu Detector Model