Your phone houses an ever-expanding record of your questionable decisions, from 2am "u up?" texts to repeated nights of Domino's orders. Now, a new study shows that it can also tell when you've been drinking with 96.6 percent accuracy. It even knows how hammered you're getting.
A research team, mainly from the University of Pittsburgh and Carnegie Mellon University, used phone metadata to determine when their research subjects—regular drinkers aged 21 to 28—were imbibing. They looked at data such as the times users texted friends, decreased spelling accuracy, and diminished neuromotor skills when using a touch screen. Using those variables, it turns out they could detect a drinking episode with close to 97 percent accuracy. They could also distinguish episodes of binge drinking in 90 percent of cases.
"Young adults drink in certain patterns," says Brian Suffoletto, an assistant professor of emergency medicine and clinical researcher at the University of Pittsburgh. "We know when they are likely to go out, what days and times, and then there are a few other indicators that get us to near certainty." The finding could be used to send real-time messages to drinkers that they've crossed a threshold or are failing to meet preset goals of moderation.
Suffoletto has long researched technological interventions to curb dangerous drinking. One challenge of this field is finding alcohol intake detection devices people will actually wear and use. There is the old-school method of personal breathalyzers. A company also recently developed a wristband that detects alcohol molecules through the skin. But both of those devices require commitment from the user that's hard to keep up, especially after your willpower has been eroded by three IPAs. "Most young adults generally don't choose to use those devices," Suffoletto says.
Using a grant from the National Institute of Alcohol Abuse and Alcoholism, Suffoletto involved researchers at Carnegie Mellon University's Human-Computer Interaction Institute and other partners to find a way to use the behavioral data inadvertently collected by our phones to track how much we booze. He was inspired by the work of David Mohr at Northwestern University, who used phone metadata to decipher users' moods.
More from Tonic: Uganda's Moonshine Epidemic
The researchers tested their prediction model on two groups: One of recent emergency room patients, aged 21 to 28, who completed a survey showing they had "hazardous" drinking habits, a standard less serious than alcoholism that has a "pattern of substance use that increases the risk of harmful consequences for the user." (Suffoletto says about 40 to 50 percent of young adults reach this threshold.) The other group consisted of college undergraduate and graduate students, of the same age range and with similar alcohol intake. Suffoletto says he didn't want a subject pool too educated or undereducated compared to the general population.
The researchers instructed participants to download a data-collection app into their phones. Their program used the first drinking episode it detected to establish a user's individual phone use pattern. Accumulating data over time improved the accuracy of the predictors. The model tracked everything from durations of time when the device is locked and unlocked to drain of battery, but the two best indicators of drinking were texts to friends in order to meet up, and diminished neuromotor ability to do basic smart phone tasks.
Comparing data gathered from phones to self-reports, the researchers concluded that the automated tech was able to determine when it had been Miller time for the participants with 96.6-percent accuracy and detected episodes of drinking heavily (more than four drinks of average alcohol content for women and more than five for men) with an accuracy of 90 percent. The high percentages "surprised our whole group," Suffoletto says, though he cautions against drawing broad conclusions from a sample size so small. Also, younger drinkers, who often socialize when drinking, may have more predictable patterns than older ones.
Suffoletto says his next project is to determine what "feedback" provided during the drinking episode is most effective in curtailing the behavior—as in, what your phone should say to you when it realizes you're on a bender. It could simply inform the user how much they've been drinking and the risks associated with being that drunk, or perhaps goals they'd programmed into an app (like a Fitbit—but one that would help you booze less). It could also provide a one-touch connection to a supportive friend, or bring up the Lyft app when it's time to go home.
Read This Next: How Dangerous Is it to Black Out From Alcohol?