Each year, 1 in 6 Americans will get sick from foodborne illnesses, resulting in as many as 128,000 hospitalizations and 3,000 deaths, according to CDC estimates. FINDER, a new tool developed by Google and Harvard’s T.H. Chan School of Public Health, allows health inspectors to pinpoint the sources of illness faster thanks to advancements in machine-learned epidemiology.
Typically, foodborne illness is traced to a restaurant either during a routine inspection, when sick customers file a complaint, or when healthcare workers report symptoms. Meanwhile FINDER (Foodborne IllNess DEtector in Real time) uses aggregated Google search and location data to help health departments target which restaurants to inspect.
First FINDER identifies queries related to foodborne illness such as “vomiting” or “stomach cramps,” and then it identifies restaurants recently visited by the user searching those terms. It then calculates how many users searched for foodborne illness-related phrases for each restaurant. Local health departments then use this list of restaurants to plan their inspections while the user is never identified and their data remains anonymous.
In 2016 and 2017, FINDER was piloted in Las Vegas and Chicago, known to have one of the most robust illness monitoring programs in the country. In the end, researchers discovered that FINDER-initiated inspections were three times more likely to find unsafe restaurants versus those that were routine or complaint-based. That’s in part because FINDER flags the last few restaurants visited by the user whereas someone calling in a complaint is likely to only cite the last restaurant visited, which does not take into account the incubation period of most illnesses.
And while Chicago supplements its traditional foodborne illness monitoring approach with social media mining—that is using computer programs to comb through Yelp review and Tweets for key words—FINDER still provided more effective results. “Relying on social apps substantially reduces the number of users whose data can help identify food safety issues at specific restaurants because the user needs to actively post a review or Tweet from the restaurant to let his or her location be known,” says Evgeniy Gabrilovich, senior staff research scientist at Google and co-author of the study.
While FINDER is poised to provide a huge boost to local health inspection efforts in terms of precision, scale, and latency, it does have some limitations. For one, a search term such as “diarrhea” does not always suggest foodborne illness so for better accuracy, FINDER also factors in a user’s response to the results (e.g. websites visited) while preserving user privacy. Additionally, users’ location data—the same that permits real-time traffic updates on smartphones—must be turned on, the user must use Google for web searches, and he or she must enter a specific symptom query search to initiate FINDER.
“One of the reasons we initiated this research is that many people use Google search for health-related queries, some of which could be a signal for foodborne illness,” says Gabrilovich. “As we discovered, there is ample data that indicates possible venues where there might be food safety issues. This machine learning-based model is still in early stages of utilizing its full potential. Our approach only suggests venues for inspection, but does not identify the specific safety issues, contaminants, or pathogens that might require further investigation or laboratory work.”