Sometimes picking a good restaurant is all about deciding where not to eat.
That's the premise of an alert system developed by researchers at the University of Rochester that tracks Twitter postings to flag restaurants where patrons are getting sick.
The idea is to give diners a more timely way to identify problem kitchens -- and avoid getting food poisoning -- than relying on annual health department inspection reports.
"You don't really know what's happening between inspections," said Adam Sadilek, who led the development of the system while a graduate student at the university and who is now a researcher at Google.
"Inspectors find no problems, and they leave," he said. "But let's say a cook comes in the next day and forgets to wash his hands and makes customers sick. This is an opportunity to identify the cook with the dirty hands."
The system that Mr. Sadilek and three colleagues created is called nEmesis. The name is a play on the word "emesis," the medical term for vomiting.
The researchers used machine learning and crowdsourcing techniques to analyze 3.8 million tweets from some 94,000 smartphone users over a four-month period in New York City.
GPS data allowed them to identify tweets from people in restaurants. The system then tracked those people's tweets for 72 hours, looking for key words indicating symptoms of food poisoning such as "tummy ache," "throw up," "Mylanta" and "Pepto Bismol." The system was "trained" so that irrelevant tweets, such as "I'm sick of school," were ignored.
Individually, the tweets were unreliable, but together they were revealing, said Henry Kautz, chair of the university's computer science department and nEmesis collaborator.
In other words, he said, "a seemingly random collection of online rants becomes an actionable alert."
The researchers identified roughly 23,000 restaurant visitors and found 480 reports of likely food poisoning. They then ranked the restaurants, giving them a "health score" based on the incidence of people becoming ill after eating there.
Their research paper, "nEmesis: Which Restaurants Should You Avoid Today?" will be presented at the Conference on Human Computation and Crowdsourcing in Palm Springs, Calif., in November.
"This is an interesting use of social media that will have to be studied further," the New York City Department of Health and Mental Hygiene said of nEmesis in an email Thursday. Since 2010, the department's bureau of food safety has used a scoring system to assign restaurants letter grades, which they must post on their doors, to better inform the public about how restaurants are performing during inspections.
The Allegheny County Health Department, which handles restaurant inspections in the county, declined to comment on nEmesis.
Mr. Sadilek said roughly one-third of the nEmesis results correlated well with New York health department inspection reports, meaning restaurants that got poor nEmesis scores also tended to have a higher number of food safety violations.
Still, there were significant discrepancies.
Mr. Sadilek said both monitoring systems had flaws. Health inspections are random snapshots of conditions, while nEmesis only focuses on people who use Twitter.
But if the methods were used together, it could lead to fewer people getting sick, he said.
For example, health departments could use nEmesis to better target their inspections. The system could generate a list of restaurants each day, sorted in terms of health risk, so inspectors could visit those first, he said.
Researchers also plan to develop an app that would allow people looking for a place to eat to pull up a map of restaurants in their area that were color-coded based on nEmesis health scores.
Mr. Sadilek acknowledged there were possible legal implications of publicly labeling restaurants as posing a health risk based on patrons' tweets.
One solution would be working with health departments to confirm the information, he said.
"We say, 'We think this restaurant is dangerous today.' So an inspector drives out and verifies" there's a problem, he said.
"nEmesis: Which Restaurants Should You Avoid Today?" is available at http://www.cs.rochester.edu/~sadilek/publications/Sadilek-Brennan-Kautz-Silenzio_nEmesis_HCOMP-13.pdf
Patricia Sabatini: email@example.com or 412-263-3066.