To better predict the spread of COVID-19 in counties across the country, researchers at Carnegie Mellon University are turning to individual’s own reports.
Using surveys conducted through CMU’s Delphi research group and distributed through Facebook and Google, researchers ask individuals to self-report symptoms and possible exposure to the novel coronavirus.

For now, the model can provide data about what is happening in communities right now. Next, the researchers will add historical data to make predictions of what the pandemic will look like weeks later.
Using that data, they hope to help individuals plan their day-to-day behavior, health care systems prepare for a surge and government officials make decisions about reopening parts of society.
“It’s very useful for governments to make decisions about whether to tighten or loosen the mitigation strategies,” said Roni Rosenfeld, co-leader of the CMU Delphi research group and head of the Machine Learning Department. “We’re able to tell them not only what’s going to happen but how will it change if you loosen things now.”

The researchers won’t know how accurate the forecasts are for a few more weeks, Mr. Rosenfeld said.
Read More
But the group hopes it can predict if a community will stay within its health care systems’ capacity when stay-at-home orders are lifted, when people begin to go back to work and bars and restaurants open to the public again.
Among many other data sources, researchers have partnered with Facebook to distribute a survey through the social media site’s NewsFeed, as well as search engine operator Google to release a separate survey and to track how often people search the internet for COVID-related terms.
Since the launch of its survey with Facebook at the beginning of April, CMU has received millions of responses. The extensive survey asks users questions about their own symptoms, those of the people in their household and some demographic information, like how many people live in their home and their ZIP code.
Based on self-reported survey data released from Facebook Monday, most counties had fewer than 2% of people reporting symptoms.
A few regions self-reported higher numbers: Natrona County in Wyoming reported 2.29%, Washington County in Maine had 2.74%, Miller County in Arkansas had 3.21% and Navajo County in Arizona reported 3.7%.
In southwestern Pennsylvania, survey responses showed less than 1% of people reported symptoms, including Allegheny County with 0.68%.
Venango County reported 1.14% of people showed symptoms, Indiana County reported 0.94% and Fayette and Beaver both reported 0.9%. Westmoreland County showed 0.43%.
‘Data for good’
The data is not meant to be used for diagnostic or treatment purposes or for guidance on any type of travel, Facebook said in a statement on its website for the project.
“The world has faced pandemics before, but this time we have a new superpower: the ability to gather and share data for good,” Facebook’s founder and chief executive Mark Zuckerberg wrote in a column for The Washington Post discussing the initiative Monday. “If we use it responsibly, I’m optimistic that data can help the world respond to this health crisis and get us started on the road to recovery.”
Although the survey is distributed through the social media platform, Facebook does not see the results and CMU researchers don’t see the identity of the survey respondent.
The tech giant shares a random ID number with CMU for each survey participant. Once the survey is complete, the researchers then send that ID back to Facebook, without the results.
Facebook and CMU then use that ID number to correct for any sample bias, or to ensure the sample of survey respondents is as representative of the entire population as possible.
“This is work that social networks are well-situated to do,” Mr. Zuckerberg wrote. Since companies like Facebook already know their users so well, they can generate enough “signal” to correct for biases and ensure sampling is done properly.
Facebook is partnering with the University of Maryland to expand the survey worldwide while CMU is working on an application programming interface to make the results available to researchers around the world.
Easier than flu
Using a separate survey, the team at CMU has also gathered data from about 600,000 users daily through Google Opinion Rewards, an app that allows users to answer survey questions and receive rewards, and AdMob applications, Google’s advertising platform.
This survey asks just one question, Mr. Rosenfeld said: Do you know anybody in your community who has COVID-like symptoms?
In the future, CMU would like to partner with other data sources to build its data set and increase the accuracy of its findings. The group, which is rolling out a comprehensive data set later this week, wants to add one or two more data sources every week, Mr. Rosenfeld said.
For faculty and staff at the Delphi research group, forecasting illness is nothing new. The center has been forecasting influenza for eight years, Mr. Rosenfeld said, and was recently named one of two National Centers for Excellence for Influenza Forecasting by the Centers for Disease Control and Prevention.
In the past, the Delphi group used a combination of machine learning and computational statistics as well as the judgments of human volunteers to generate weekly forecasts for influenza.
When the Delphi group started forecasting for the flu, they had 20 years of data to look back on, allowing them to be more accurate in their predictions, Mr. Rosenfeld said. Now, they have to rely on data from just the last few months.
In some ways, it has proven easier to forecast for COVID-19 than the flu. One of the hardest parts of forecasting influenza is determining pre-existing immunity in a community because it depends on so many factors, Mr. Rosenfeld said. With the novel coronavirus, no one was immune.
“From a public health perspective, it’s much worse,” he said. “From a model point of view, it actually makes it easier.”
In about four weeks, the researchers will start to test the accuracy of their predictions.
They’ll look back one week, two weeks, one month to see how their forecasts lined up with what health care systems saw.
“As we move along, we’re going to build a new track record,” Mr. Rosenfeld said.
Lauren Rosenblatt: lrosenblatt@post-gazette.com, 412-263-1565.

This story was originally published with images that are no longer available.





Comments
Welcome to Post-Gazette subscriber-only commenting community. Please review our community guidelines.