THE CARTOON ARCHETYPE of a drunk person is a disheveled mess, with droopy eyelids, an erratic gait, and bubbles coming off their head—for some reason. If only it were so easy to tell if someone’s blitzed out of their mind. If a cop pulls you over, they can only objectively determine intoxication on the roadside with a breathalyzer: In the lungs, ethanol is transferred from the blood into air, so the device can detect alcohol in your exhalation. Even then, one person at the US federal legal driving limit of .08 breath alcohol concentration (BrAC) might act perfectly normal, while another person would be trying unsuccessfully to fish a slice of pizza out of a gutter.
But scientists are working on what might be a new way to determine intoxication, by returning to a stereotypical characteristic of the drunk that’s actually true: that soused walk. No matter how well you think you’re walking when you’re intoxicated—especially if you compare yourself to your friend in the gutter—subtle and not-so-subtle changes in your gait could betray your alcohol level. And if you’re carrying a smartphone, its onboard accelerometer can pick up those changes. In fact, scientists from the University of Pittsburgh just published research showing that in the lab at least, they can use smartphone motion data to detect if a subject is intoxicated, with an average accuracy of 93 percent. It sounds like fun and games—getting people loaded and watching them stumble around for science—but the work could have some serious utility.
Their experiment began with a surprisingly sophisticated drink for the subjects: vodka with lime juice and simple syrup—heavy on the vodka. The subjects had an hour to finish the stiff drink, because the researchers wanted to get them to a peak of .20 BrAC fairly quickly, though of course safely. “We really wanted to have enough data points over the threshold of the legal limit of alcohol, which is 0.08,” says study lead author Brian Suffoletto, then at University of Pittsburgh School of Medicine and now at Stanford University. “So if you’re giving them a lower dose, you may only get one or two time points before the body sort of metabolizes the alcohol.” Also, there are only so many hours in a lab day—no sense in waiting around for the subjects to leisurely build a buzz.
Before the drinking, the researchers had strapped ordinary smartphones to the lower backs of the subjects to get an idea of their baseline gaits. They had the participants walk 10 steps forward, turn around, and walk and 10 steps back. The subjects then repeated this strut while intoxicated, doing it once an hour over the course of 7 hours as their BrAC peaked and waned. All the while, the smartphone was logging motion data.
Because each participant had their own unique sober gait and unique intoxicated gait, Suffoletto and his colleagues used individualized mathematical models that allowed them to compare each person against himself or herself. “We found that, really, what was driving the model is the medial-lateral sway,” says Suffoletto. That means lurching side to side; your typical sober gait is more or less down the line. “Which kind of makes sense when you think about the caricature of the drunk cartoon, a figure who kind of is swaying back and forth,” he continues.
The model of each person’s particular gait allowed the researchers to correctly identify over 90 percent of the time when a subject’s BrAC passed .08. Of course, in the real world, most people don’t strap their smartphones to their lower backs, but the researchers are planning to do further experiments with subjects carrying the devices in their hands or pockets.
Even though this research is in its early stages, it builds on a growing body of evidence showing that a device’s motion data could be an accurate measure of intoxication. “The work is well done and extends some of our group’s work on the same problem,” says Worcester Polytechnic Institute computer scientist Emmanuel Agu. He and his coworkers also used smartphones to measure intoxicated sway, and even built an inebriation-detecting app called AlcoGait.
Law enforcement officers have long used a version of a gait test when pulling over suspected impaired drivers. In the “walk and turn test,” the driver must be able to walk heel-to-toe in a straight line while maintaining their balance and counting the steps aloud. “Gait-sensed impairment is quite accurate, which is why the police have used the field sobriety test for decades,” Agu continues. “One key difference with breathalyzers is that it’s an assessment that can be done completely passively and avoids the stigma of having to carry breathalyzers to parties.”
However, to get a spot-on measure of how much alcohol is in a person’s system, you need to get a blood sample. (Breathalyzers accurately measure the alcohol in a person’s breath, but can be thrown off by other variables, like if the person recently chewed gum. That’s why police officers don’t give you a breath test immediately if they suspect you’re intoxicated.) But here we should recall our pizza-loving friend who doesn’t handle their alcohol so well: Thanks to their metabolism, body weight, tolerance for alcohol, and other factors, one person’s .08 might be another’s .20. The legal BrAC limit for driving in the US is based on a level of intoxication that would impair most people—that is, it’s an average. “This means that some people who have high tolerance would get DUIs even when they are not impaired,” says Agu. “Gait sensing directly senses impairment and may facilitate a more futuristic approach to assessing DUIs.”
Such gait sensing may also play an important role in public health. The predominant model for treating alcohol-related problems has been single-session interviewing to motivate the patient. “There’s enough evidence to say that it’s moderately effective in some people to reduce drinking,” says Suffoletto. “But I don’t care how persuasive a person you are—think of someone like Bill Clinton or whatever—the salience of whatever message you provided in that single episode dissipates with time. And it especially dissipates and becomes less salient when individuals are around peers drinking.”
Previously, counselors trying to treat problematic drinking have had to guess when a person might drink by asking what days over 90 percent of the time they typically imbibe. Then they might send them texts on, say, Thursdays and Saturdays to check in. “You can imagine that this resulted in triggering dialog when it wasn’t needed, that or missing out on opportunities that we just didn’t sense,” says Suffoletto. “The thought started going through my head a few years ago: We’re already interacting with phones in people’s pockets. What if we can use some of the data on those phones to infer when a drinking event is going to happen?”
By monitoring gait, health experts might be able to better predict when a person’s problematic drinking might occur. Suffoletto’s larger goal with this research would be to provide more targeted support to the person during such high-risk periods, helping them reflect on their drinking and set limits. “It made more sense to me than just trying to send people out to the world and expect them to act the way you told them to two months ago,” he says.
We humans aren’t very good at recognizing when we’re drunk. In previous work, Suffoletto measured gait and asked subjects to self-assess how impaired they were. “We found that 50 percent of the episodes where they were intoxicated, they didn’t recognize it. But we were able to identify it based on their gait signature,” he says. “So you can imagine how that would be useful in situations where somebody may have had a couple of quick drinks at a bar or something, and then said, ‘Well, I’m not feeling intoxicated.’”
The intoxicated brain may lie to a person, but smartphone gait data won’t. “So now you’ve got an objective tool,” Suffoletto adds, “where if we’re good enough, we can actually refine it so that we can estimate what the blood alcohol content is just by their gait signatures.”
Researchers have actually been trying to do this another way with wearable alcohol biosensors, which could theoretically monitor the amount of alcohol in a person’s body in real time by measuring the small amount of alcohol excreted through the skin. But that isn’t necessarily a good measure of intoxication. “My issue with this has always been: In the real world, we don’t care as much about the absolute value, as the impairment,” says Suffoletto. “We know that one person’s response to alcohol is different than another’s.”
In other words, the objective measurement of alcohol in someone’s system doesn’t necessarily map to risky behavior. “What drives the public health problems,” he adds, “is largely around acute intoxication causing things like motor vehicle accidents, interpersonal violence, risky sexual encounters, those kinds of things. And each person is going to have their different threshold.”
The old caricature of the stumbling drunk may be over the top, but in the end, there’s a certain truth to it: You may be able to play down your intoxication, but your gait doesn’t lie.