The Toronto Star
August 13, 2013
Researchers have created a way to mine Twitter for outbreaks of food poisoning by tracking the messages of restaurant patrons who become ill—a tool that may help people, and health agencies, identify places that pose health risks.
The Twitter-tracking system called nEmesis was created by researchers at the University of Rochester, who plan on turning it into an app that would provide real-time information about which restaurants to avoid—and inspect.
"Right now, millions of people get sick with foodborne illness," said researcher Adam Sadilek. "Most of these issues are preventable. If people have better information about what's happening around them in real time, they can make informed decisions and not get sick."
Sylvanus Thompson, a Toronto Public Health associate director, says the new tool could be useful for DineSafe, the food safety program that inspects places that serve and prepare food. Inspections result in a pass, a conditional pass or a notice of closure—all results are posted at the front entrance of the establishment.
"I'm interested in any new tool and anything that can help Toronto Public Health to identify risk factors and places that are more likely to contribute to foodborne illness," said Thompson, who's responsible for the food safety program.
According to a 2009 Toronto public health report, it's estimated that about one in six Toronto residents—about 437,000 people—get a foodborne illness each year. But the majority of cases are never reported.
Toronto Public Health already receives complaints about food poisoning from various sources, but nEmesis is "a new tool that could have potential uses for us."
Details about this new tool are contained in the report” nEmesis: Which Restaurants Should You Avoid Today?”
It will be presented at the Conference on Human Computation & Crowdsourcing, in November in California.
The system listens for geo-tagged tweets sent from restaurants. This can be tracked because people often send messages from their cellphones, which carry GPS information indicating where the individual was when the tweet was sent.
The person's tweets are then tracked for the next 72 hours for signs of food poisoning.
An algorithm was designed to identify potentially relevant messages, including mention of: stomach ache, throw up, Mylanta and Pepto Bismol.
The algorithm was created, in part, by people hired through Amazon's Mechanical Turk, which crowdsources easy tasks. The workers helped train the algorithm to look for accurate tweets so that messages such as "I'm sick of homework" weren't tracked by nEmesis.
For four months—from Dec. 26, 2012 to April 25, 2013—researchers collected data from Twitter users in New York City. The system analyzed 3.8 million tweets from more than 94,000 unique users. It tracked about 23,450 restaurant visitors and found 479 reports of likely food poisoning associated with a restaurant.
Based on the statistical analysis, nEmesis then ranked the restaurants; according to how likely it is for someone to become sick after eating there.
Researchers compared the results with the current database of restaurant inspections done by the city's Department of Health and Mental Hygiene and found the Twitter-based scores correlated with official inspection data.
A single user complaining about acute food poisoning has a small impact on the behavior of others, note the report's authors.
But an automated system that tracks a large online population can find important patterns, by matching the message with a restaurant.
In other words, they say, "a seemingly random collection of online rants suddenly becomes an actionable alert."
Manual inspections by health officials are good at capturing information about things that customers never see, such as mouse droppings in the kitchen, said Sadilek.
But, he adds, automated food inspection is good at measuring the more subtle things that an inspector may not see—but that diners will eventually feel—and tweet about.
"When you join these two approaches they become really powerful," said Sadilek, who conducted the nEmesis research while at the university, but now works for Google.
The nEmesis app will differ from others on the market because it provides real-time information—not past inspection reports.
"Just because a restaurant had a food poisoning incident five years ago, doesn't tell me much about my lunch today," said Sadilek. "(nEmesis) can literally detect a bad chicken delivery in time for you to make a decision."
Quality News Today is an ASQ member benefit offering quality related news
from around the world every business day.