In 2011, one in six Americans, or 48 million people, suffered from some kind of foodborne illnesses, according to the Centers for Disease Control and Prevention. A team of computer scientists from the University of Rochester believe these high numbers are completely preventable. That's the motivation behind nEmesis -- a system that can track food poisoning by parsing mountains of geo-tagged data from Twitter.
The foundation of nEmesis is an algorithm (based on machine learning) that can single out tweets that suggest the tweeter is sick. Using crowdsourced workers in Amazon’s Mechanical Turk program, the algorithm is further trained to spot tweets from people likely to have contracted foodborne illnesses. Running with this constantly fine-tuned "brain," nEmesis "listens" to tweets tagged at verified restaurant locations and tracks the same users' tweets for 72 hours. If any of these users tweet about feeling ill, nEmesis captures this information and traces it back to where the person had eaten.
Just how effective is nEmesis at detecting risky restaurants? Over a four-months period, the system mined 3.8 million tweets from 94,000 unique users in New York City. From that data, it tracked 23,000 restaurant visitors and determined 480 reports of likely cases of food poisoning. From this information, the researchers ranked the restaurants by how likely the customer is to catch a foodborne illness there. This ranking was then compared to the same type of ranking based on the New York City Department of Health's inspection results. The researchers found that between the two rankings, there was an overlap of one-third.
On the surface, it might seem that the remaining discrepancies point to nEmesis's inaccuracy. But according to Adam Sadilek, who started the project as a postdoctoral fellow at Rochester, the differences are the most exciting part. That’s because they give nEmesis the opportunity to contest and verify inspection data, which Sadilek said in a press release, "isn’t perfect either."
Currently, the city’s Department of Health conducts unannounced inspections of restaurants at least once a year, a process that Sadilek compares to a "roll of a dice." Because nEmesis works in real-time, if a restaurant were to have a bad batch of chicken one day, the system can alert inspectors to inspect that restaurant in a timely manner. WIthout this "adaptive inspection" process, a "bad chicken" incident may cause some customers to get sick yet still pass through the next random inspection unacknowledged. Sadilek says the big goal is for nEmesis and health agencies to work together to produce the most accurate and updated ranking, so people can make informed decisions.
Sadilek sees nEmesis assuming the form of a mobile app, like a food-poisoning version of GermTracker, a previous project of his that uses similar technology to show users who around them has the flu.
Although Sadilek has wrapped up research at the University of Rochester, he says the project will carry on under former advisor Henry Kautz and new student researchers. In the meantime, we should probably keep up with those restaurant check-in’s -- they could be saving others from a ruined night out.