Searching for ADHD in a Million Tweets

After analyzing 1.3 million tweets written by nearly 1,400 Twitter users, researchers at the University of Pennsylvania have created a machine learning model that can predict which of the site’s users are affected by attention deficit/hyperactivity disorder (ADHD).

Researchers monitored Twitter users with self-reported ADHD, comparing them to a set written by age- and gender-matched controls. The team then used various machine learning models to search for patterns in the two sets of tweets.

Words like “hate,” “disappointed,” “cry,” and “sad” appeared more frequently in the ADHD group’s tweets. Lyle Ungar, one of the study’s authors, told Healthcare Analytics News™ that such tweets yielded findings his team did not expect.

“I didn’t realize how common it was for patients to use marijuana to treat their symptoms, so you see people talking more about dope and weed,” he said. “There were also ones that seemed more related to confusion and distraction…They’re swearing more, they’re hedging more and talking more about being tired or self-critical.”

One of the team’s researchers, J. Russell Ramsay, PhD, an adult psychologist, confirmed that the observations were consistent with what he has observed in practice.

Another author, Sharath Chandra Guntuku, told Healthcare Analytics News™ that the methodology was 76% correct in a 5-fold predictive validation test. Because less than 10% of American adults have ADHD, that comparison would produce significantly more false positives were it to be applied to the general population.

Both Guntuku and Ungar stressed that such work could be used to develop interventions for patients with ADHD, through feedback given to clinicians or to the patients themselves.

“We can create a model that points out when a user is maybe posting a lot when they are supposed to be sleeping, or posting a lot about high or low emotions,” Guntuku said. “There are similar sorts of patterns we can summarize well, and then the physician will have a much more personalized insight into patients’ health or behavior.”

“The nice thing about social media is that young people are all on social media,” Ungar said. “They’re using their phones to communicate. It would be nice if the phones could give them some sort of feedback, like, ‘Hey you seem a bit more fragmented and distracted than usual, do you want to talk to someone?’” He would not want such a system to be entirely automated, however, instead preferring an integrated approach that alerted human clinicians of patients who are at risk.

Ungar does admit that there’s a “creepy aspect” to that idea. Twitter feeds are entirely public, which allows researchers to do work like this without having to contact each individual user whose data is analyzed.

“You worry a little about companies and insurers monitoring people’s posts…there’s a negative side,” he said. “The positive side is that it can actually help, we hope.”

Social media is increasingly a source of insight for mental health conditions. Another recent study examined hundreds of thousands of tweets to find if they could be used for depression diagnoses.