Computers mine online reviews for signs of food poisoning

Reviews of restaurants on social media can be helpful. They may guide you to a popular eatery, or point you to what dishes are popular. They might also tell you when and where the food left someone feeling ill. Now scientists are training computers to scout online reviews for such signs of sickening food.

Tainted food and drink sicken some 48 million people in the United States each year. That’s according to the Centers for Disease Control and Prevention. Nearly seven in every 10 of those incidents came after dining in restaurants.

Government agencies collect reports of these incidents. They typically learn of them when someone calls in a complaint to the local health department. Not everyone will do that. There is, however, a good chance someone posted about it online. Those posts can be helpful to epidemiologists (Ep-ih-dee-me-OLL-oh-gizts). These researchers study disease outbreaks and their spread.

Online sources “may give you an early signal of something happening,” says Mauricio Santillana. He works at Harvard University Medical School in Boston, Mass. He was not involved in the study. An expert in digital epidemiology, he uses the internet to track disease.

In 2012, scientists at Columbia University in New York City built a computer program to help public-health officials. They taught it to read local restaurant reviews on the social-media site Yelp. Epidemiologists told the computer program what it should look for. They did this by labeling clues in 500 reviews.

The prototype computer program then looked for specific words that might point to illness. These words included sick, vomit, diarrhea and food poisoning. If any of these words appeared, the computer would flag the review. It also checked to see if a review indicated that multiple people had gotten sick. This might signal an outbreak. That’s when there’s a sudden uptick in the number of cases of a particular illness. Epidemiologists with New York City’s health department then read the flagged Yelp reviews. When they spotted a suspected outbreak, they sent out experts to investigate.

“There are too many reviews for health-department epidemiologists to read and manually search,” says Tom Effland. He is a computer scientist at Columbia.

The computer program Effland worked on made what was impossible possible. It allowed health-department epidemiologists to monitor millions of reviews. In doing so, they detected 10 disease outbreaks. They also turned up 8,523 complaints of food poisoning from local restaurants. These cases all occurred between July 2012 and May 2017.

Effland’s group shared its findings January 10 in the Journal of the American Medical Informatics Association.

The restaurant reviews helped point to disease outbreaks that health officials otherwise might have missed. A pilot test of the computer program ran for just a few months. It showed that most foodborne disease had not been reported to city health officials. Only three in every 100 cases had been phoned in to the official complaint center.

Younger people spend much of their life interacting with social media. They are less likely than their parents to report illness through traditional ways, the study’s authors note. That makes scouting for illness on social-media sites ever more useful, they add.

Effland and his colleagues are now working to improve their system. They also are exploring how to apply this approach to other social-media sites, such as Twitter.