Photo by Guian Bolisay, via Flickr
Photo by Guian Bolisay, via Flickr
ScienceOnly Human

Why Do Some Teens Become Binge Drinkers? Algorithms Answer.

The first time I got drunk I was 15. It was in a hotel room in Paris, on a trip with my high school French Club, drinking vodka and Orangina from a plastic bottle. I remember looking at my blurry reflection in the bathroom mirror and thinking, So this is what being drunk is. I didn’t hate it. I drank a few more times that year, and then pretty steadily for the next two. I had one blackout night in a friend’s basement. Then came college, where everything escalated. It honestly makes me queasy right now to think about what I put my body through.

But it was fun. And it didn’t lead to anything horrible. I did well academically, went to grad school, found (mostly) gainful employment. I’m 30 now and, knock on wood, don’t have any health problems.

My story is typical. “We tend not to want to say this out loud to teenagers, but most people who tried drugs don’t get addicted,” says Hugh Garavan, a cognitive neuroscientist at the University of Vermont. “Most kids have tried alcohol by age 14, and most kids don’t develop a problem. Same with cigarettes and same with cocaine. But there’s a certain subset who do, and we don’t have a clue what it is about them.”

Scientists have pinpointed lots of factors that increase the risk of alcohol misuse — a bit. Adolescents who are anxious or impulsive, for example, tend to be at higher risk. Same for those who carry certain genetic variants (dubbed ‘SNPs’) in their genome, and for kids who are abused or neglected. But most studies haven’t looked at enough factors, or at enough kids, to make predictions with much oomph. “It’s hard to look at all of it, but we have this luxury,” Garavan says.

In today’s issue of Nature, Garavan and his colleagues present a new predictive model based on an enormous amount of data—brain scans, genetic screens, personality trait tests, and family and medical histories—from 2,400 teenagers in Europe. The model isn’t by any means a crystal ball, but it can guess which 14-year-olds will become binge drinkers by age 16 with odds far better than chance.

Garavan is one of the leaders of the IMAGEN Consortium, a €10 million-plus study following teenagers at eight different sites in Europe with the aim of pinpointing biological and environmental factors that influence adolescent mental health. The Consortium has published several dozen papers related to various relationships between brain activity, genetics, and behavior at age 14. This paper is the first to look at whether data collected from the volunteers at age 14 could predict their behaviors at age 16. Turns out, it can.

Part of the reason this study is powerful is because of its math. The researchers’ task was retrospective prediction: Knowing what happened to a kid at age 16 and looking back at a massive pool of data from age 14 to see what could have predicted it. In this case, though, the massive pool of data posed a problem. “With a gazillion variables that could potentially predict, there’s a real risk that you’ll find associations just by chance,” Garavan says.

To get around this, researchers used a machine-learning method that separated the data into many subgroups of participants. They’d develop a predictive model for one subgroup, then test it on another subgroup to see if the relationships held. And then they repeated the process on another group, and another and another. In the end, the best model relied on several dozen variables, as shown in this chart:

View Images
Whelan et al., Nature 2014 (Click to enlarge)

The variables listed on the left hand side refer to what participants scored at age 14. (For this analysis, the researchers only looked at 14-year-olds who were not drinkers, reporting two or fewer alcoholic drinks in their lifetime.) The factors with the most negative correlation coefficients (that is, the ones with lines furthest to the left) are those that, on their own, most strongly predict alcohol misuse at age 16. The variables with the most positive correlation coefficients (with lines furthest to the right) are most protective against alcohol misuse.

So, according to the model, if a 14-year-old non-drinker has an “extravagant” personality, which is characterized by grandiosity, exuberance and impulsivity, he or she will have a higher risk of becoming a binge drinker by age 16. (The researchers defined binge drinking as having at least three binge-drinking episodes leading to drunkenness.) In contrast, a 14-year-old non-drinker with “conscientious” personality scores will be at a lower risk of becoming a binger.

Under the “Brain” heading, you’ll see “Parenchymal volume,” which is the volume of the whole brain. In other words, kids who have larger brains at age 14 are at a higher risk of binging at age 16. This is intriguing, Garavan says, because the brain typically gets smaller in adolescence, when connections that aren’t used get pruned away. “So their brains seem to be less mature.”

By weighing these several dozen factors together, the mathematical model could correctly classify 66 percent of the binge drinkers and 73 percent of the non-binge drinkers, which is significantly better than chance. (About 45 percent of the 14-year-olds in this model went on to become binge drinkers at age 16.)

“It is a great example of how machine learning can provide novel insight in ways that have big potential for clinical impact,” says Dennis Wall of Stanford University, who was not involved in the study but is using similar techniques to diagnose autism. The new model’s predictive power is somewhat modest, Wall adds, “but even this could have meaningful impact.”

To me, what’s most interesting about the study is that the variables most difficult and expensive to obtain — the genetic markers and brain signatures — are far less important to the model’s predictive power than things like personal history and personality traits. Garavan says his team created a version of the model with only the history and personality measures, and “on their own they do a pretty good job.” That’s good news because it means that doctors, parents or educators might be able to spot high-risk teens without much more than a survey or two.

One big caveat with this particular study is that it stops at age 16. For most teens, drinking does not. So the IMAGEN team brought many of the same teens back into the lab at age 18 and is now working on comparisons of all of this data at age 14, 16, and 18. If the researchers get funding, they’ll look again at age 23.

As I do with most people I interview, I asked Garavan what message he would most like to get across to the general public about this study. He answered with what he doesn’t want to communicate.

“What I don’t want to get across is that we have figured out some secret formula: Give us your kids and we’ll tell you what to do,” he says. Teenage drinking “is not just evil kids choosing to do bad things. There are these preexisting risk factors of vulnerabilities, and we can measure them.”