Online searches help predict COVID-19 cases

A new study suggests that tracking online searches for symptoms can help predict peaks in COVID-19 cases.

The research, led by University College London, also known as UCL, Public Health England, and Bar-Ilan University in Israel, looked at online health trends in seven countries, and found that searches for certain symptoms of COVID-19 increased markedly in the weeks preceding peaks in confirmed cases.

On average, a spike in searches occurred 17 days before a rise in positive test results. The researchers said online search data could help health authorities respond more swiftly to outbreaks of the disease caused by the novel coronavirus.

"Our best chance of tackling health emergencies such as the COVID-19 pandemic is to detect them early in order to act early," said Michael Edelstein, a professor of public health at Bar-Ilan University who co-authored the study. "Using innovative approaches to disease detection, such as analyzing internet search activity, to complement established approaches is the best way to identify outbreaks early."

The study, which was published in the journal Nature, said that the frequency of searches for symptoms like loss of the sense of smell, rashes, pink eye, sneezing, and blue face correlated with a future rise or fall in confirmed test results.

The researchers said queries about common symptoms, such as cough, fatigue, and fever, were not among the most correlated or impactful, suggesting that web searches about rarer symptoms may be more informative.

Spikes in searches preceded a rise in cases by different intervals depending on the country. In the United Kingdom, a jump in searches came on average 24 days before a major peak in confirmed cases. In Canada, the average interval was 31 days. The delay was 20 days in the United States; 14 days in Italy; 12 days in Greece; 10 days in France; and six days in Australia.

South Africa presented an outlier, with an interval of 53 days. Researchers said this was perhaps due to limited testing capacity and levels of internet access. When South Africa was excluded from the data, spikes in searches preceded increases in confirmed cases on average by 16.7 days.

"This provides an indication of how much sooner the proposed unsupervised models could have signaled an early warning about these epidemics at a national level," the paper said.

Web search analysis has helped predict outbreaks of diseases in the past. The Google Flu Trends project, which ran from 2008 to 2015, helped predict some regional outbreaks of influenza in the US, and flagged a spike in symptom searches two weeks before the first reports released by the Centers for Disease Control and Prevention during the 2009 flu pandemic.

The algorithm ran well for short periods but began to become less accurate over time, and ran into serious problems in the lead-up to the 2013 flu season, for which its predictions were off. A host of issues contributed to this failure, including the breadth of illnesses that can cause flu-like symptoms and levels of reporting of flu in the media.

The UCL-led study performed measures to reduce media influence on the data, and also noted that symptoms that are common among a range of illnesses, such as a cough, were not the best indicators of an emerging trend.