Can symptom surveys predict the next Covid-19 hotspot?

How useful is symptom data to public health authorities?
How useful is symptom data to public health authorities?
Image: Reuters/Brian Snyder
We may earn a commission from links on this page.

Dry cough. Fever. Chills. The symptoms of Covid-19 have become global public knowledge. And as countries try to get a handle on the disease, you may have already been asked to disclose your own online.

A growing number of online studies, apps, and trackers are polling the public about their health. Facebook and Google have both partnered with academics to launch surveys of symptoms, an effort that could reach billions of people. Israel, the United Kingdom, and Norway are surveilling for Covid-19 symptoms, as are some US states including Florida, Rhode Island, Colorado, and Alabama

The simplest versions of the apps simply help people keep track of their own symptoms, and direct them toward testing centers if necessary. But if enough people use the tools, their collective data could be used by public health authorities to forecast hotspots. A rise in reported Covid-like symptoms in an area that lifted its lockdown order, for example, could help hospitals prepare for more Covid-19 patients in the following weeks, or institute a robust isolation and contact-tracing program.

“Tracking symptoms definitely has a role, and I do think it can be an early warning sign of when something is about to occur,” said Philip Chan, an associate professor in the Department of Medicine at Brown University. Similar symptom-tracking tools have been used to pinpoint outbreaks of Ebola, SARS, H1N1, and Zika.

One notable effort is a smartphone app known as Covid-19 Symptoms Study, developed by a group of scientists and doctors at institutions including Massachusetts General Hospital, Stanford University School of Medicine, and Kings College of London. The research team published a study in Nature on May 11 that analyzed the symptoms of more than 2.6 million participants from the US and the UK.

By comparing reported symptoms to those in patients who tested positive for Covid-19, the study predicted that 17% of untested participants were likely to have the disease. More to the point, the app was able to predict spikes of Covid-19 infections in Southern Wales: “Users reported symptoms that predicted, five to seven days in advance, two spikes in the number of individuals reported by public health authorities to be confirmed with Covid,” wrote the study’s authors in a separate paper for Science. “Conversely, a decline in reports of symptoms preceded a drop in confirmed cases by several days.”

Several experts emphasized that symptom surveillance is no substitute for clinical testing. But in aggregate, symptom-tracking could give hospitals and public health officials a heads-up that something is amiss. Facebook is updating maps of its symptom surveys in real time to predict surges. 

That kind of prediction becomes more important when not everyone who experiences symptoms gets tested. “One of the biggest problems we have as a society is that to understand the incidence of disease, we are reliant on testing,” said Benjamin Rader, a graduate research fellow at Boston’s Children’s Hospital who has been working on Covid Near You, a crowdsourced map of coronavirus symptoms in the US. In countries that neglected or have been unable to execute widespread Covid-19 testing, monitoring symptoms can both further research and sound the alarm on a potential outbreak.

“Self-reporting based on symptoms allows you to see results more quickly,” says Elya Tagar, vice president of business development at Diagnostic Robotics, an AI firm that used self-reporting to identify hotspots in Israel. The company also helped Rhode Island launch a symptom tracker last month, though the state isn’t using the data to identify hotspots. But other states like Alabama and Florida plan to do so.

Some of these tools could eventually use machine learning approaches to identify emerging hotspots. Such tools would analyze large amounts of data, such as symptom reports and number of confirmed cases, and make predictions about where the virus will spread.

But AI and machine learning approaches, in the case of Covid-19, have one major weakness: a lack of training data. Covid-19 is still new, and the amount of information researchers have on it is scant compared to other infectious diseases. “In general, the more people that participate in these surveys, the better information these AI tools are able to spit out,” said Rader.

When the training sets are small, they’re also more likely to bake bias into algorithms. The authors of the smartphone-based symptom studies acknowledge that people who use their app are likely very different from the population at large; they risk missing data from older and low-income populations who are most likely to die of the virus.

“It is impossible to generalize the population of app users to the public,” said Joseph Amon, the director of global health at Drexel University. “They are likely to be quite different in terms of many characteristics that also influence exposure to infection and severity of disease.” Israel’s digital surveys are not reaching its Bedouin communities, for example, many of whom don’t have access to the internet.

“There are at the time of writing still not sufficient data to build AI models that can track and forecast its spread,” wrote Wim Naudé in an April 28 article in the journal AI and Society.

And, as always, AI is no substitute for common sense. “Every day, I see images of people in communities where there is ongoing coronavirus transmission crowded into restaurants, working in meat packing plants, or being housed in overcrowded jails, without masks or other protections,” wrote Amon. “It doesn’t require AI to identify these as hotspots and places for future outbreaks.”