I’m not feeling very well today, which stinks because it’s Friday and I had some really fun plans tonight. Instead, I’ll probably end up staying home for a quiet night with my husband and daughter and some takeout food, and an early night to bed. I’m not complaining too much, though, because let’s be honest, you enjoy those quiet nights when you have a one-year-old toddler! But a recent article in New Scientist makes me wonder if, had I been paying close enough attention to Twitter, I could maybe have known over a week ago that I would’ve been under the weather today, and save me from having to tell all my girlfriends I’m probably pooping out on them tonight.
Adam Sadilek is a postdoctoral fellow at the University of Rochester in New York. He and his computer science colleagues have devised a way to tell if people will get sick just by using nearby Twitter posts in combination with machine learning and natural language understanding techniques. Sadilek and his team used a machine-learning algorithm to analyze 4.4 million tweets containing GPS tagging data from over 630,000 Twitter users in New York City over a month’s time. They trained the algorithm to be able to distinguish between tweets by healthy people (“I am so sick of this hot weather we’re having!”) and someone who was actually ill and showing signs of having the flu (“I feel so sick today. I think I’m going to have to call in to work”). They created the video below, which shoes a heat map of flu occurrence over the course of a day, based on their findings:
The algorithm performed impressively and was able to figure out when healthy people would get sick up to eight days in advance, with a 90 percent accuracy rate. While it’s a similar idea to Google Flu Trends, which tracks where and how often people are searching for flu-related terms, Sadilek contends on his website that this method is better, and shows “emergent aggregate patterns in real-time, with second-by-second resolution [emphasis his]. By contrast, previous state-of-the-art methods…entail time lags from days to years.”
Sadilek and his team aren’t the only ones delving into social media as a way to predict illness, though. Sickweather is a startup that uses data from Facebook, Twitter and other social networks to figure out what illnesses (common cold, bronchitis, chicken pox) or other maladies (allergies, stress, headache) are going around in your area. As their website says: “Just as Doppler radar scans the skies for indicators of bad weather, Sickweather scans social networks for indicators of illness, allowing you to check for the chance of sickness as easily as you can check for the chance of rain.”
Though impressive and interesting examples of social networking as epidemiological prediction engine, neither Sadilek’s research nor Sickweather are perfect. Sadilek’s system is limited in that it misses many cases of illness because people don’t reliably tweet about their symptoms, and contact or proximity is only one factor that leads to people getting sick. And Sickweather, while in their closed beta, apparently acknowledged that their algorithm was in process and that people might notice some “false positives” from people tweeting about the “sick kicks Benny just bought” or the “sick beats in the club last night.”
Even with the limitations, I think it’s still pretty cool to see how something like social networking, that is so often maligned as an egotistical, oversharing time-waster, has real power toward helping to predict outbreaks of significant disease, like the flu. So now, if you’ll excuse me, I’m going to go compose a tweet and a Facebook post detailing how craptastic I feel (you know, for the greater good), and go make myself a nice hot mug of tea.