Hashtag Epidemiology

Search Twitter for the keywords "food poisoning," and you might discover that Geekeratti darling Adrianne Curry thinks she ate a bad salad this morning. Or that there was much suspicion about the freshness of cronuts served at the Canadian National Exhibition in Toronto. Possibly interesting, but probably not so useful.

But take a kajillion geo-tagged tweets, pick out all the messages about bellyaches, map which restaurants those people visited in the last few hours, and presto! You've turned social media into a tool that could stop a surge of Montezuma's revenge in its tracks.

This is exactly what researchers at the University of Rochester were going for when they coded a software system called nEmesis, which is designed to monitor public, geo-tagged tweets. The program then guesses how likely a given user is to come down with a food-borne illness based on this information.

Next, using Mechanical Turk, the highest-risk tweets are sorted by an actual human to make sure the user wasn't just talking about a sick concert they heard last night. Finally, restaurants ailing users recently visited are ranked for possible bugs.

Too Much Information?

"There is no such thing as over sharing," said Vincent Silenzio, a physician and associate professor at the University of Rochester who worked on the project. "There is information that's embedded in these patterns of how and what people are sharing. If they weren't sharing 'too much,' we wouldn't have the opportunity to realize how fabulous of a remote sensing system that we have in our hands."

Silenzio says Twitter data is particularly context-rich, because it can be cross referenced with the location and reporting times of other users.

Out of the 3.8 million tweets researchers analyzed from 94,000 Twitter users in the New York City area, 480 cases of food poisoning turned up. These cases mapped neatly onto the city's health-inspection grading system.

The program was also designed to pick up on colloquial language connected to sickness. It "learned" that verbiage like "#tummyache," "being sick man" and "ughhh cramps" are pretty good flags to look for. Unsurprisingly, expletives were too.

Where Can I Get The App?

You can't. While the nEmesis project could be used as a back end for a killer restaurant review app, it would need some tweaks to prevent users from "poisoning" the data with false claims or gaming keywords to torpedo the reputation of competitors.

For the moment, nEmesis will stay in the laboratory, perhaps to roll out as a tool for public health workers to zero in on trouble spots, or for epidemiologists looking for ground zero of a listeria outbreak.

"I think that these are going to prove to be a set of tools that are really going to become important to a lot of the work we do in epidemiology, not just food-borne illness," Silenzio said.

Researchers have already been able to use similar Twitter-mining models to predict the spread of influenza viruses, and even to track cases of depression based on "sentiment analysis" of language used in tweets.

Read More:

• How Twitter Can Help You Avoid Getting Food Poisoning (Digital Trends)
• Can Twitter Predict Where You'll Get Food Poisoning? (The Atlantic Cities)