When Algorithms Are Running the Asylum
Psychiatry’s biggest breakthrough in decades might come from machines that don’t need to understand the mind at all.
A few years ago, Adam Chekroud tried something new. The Yale University neuroscientist pulled as much data as he could out of questionnaires that had been given to people in a large clinical trial of drugs used to treat depression. The data included obvious elements such as the study subjects’ race, gender, and education level, but it went much deeper, too. It included the answers that study participants had given to questions about themselves. For example, had they been bothered by aches and pains throughout their body? Did standing in long lines make them fearful and anxious?
Then Chekroud and his colleagues had a machine-learning algorithm look at how these various factors correlated with the patients’ responses to a common depression drug. It turned out that a combination of 25 factors — including the answers to questions like that one about long lines — predicted whether patients would be helped by the treatment or not.
How your reaction to queuing, combined with 24 other apparently unrelated factors, might help explain why a particular drug does or does not improve your brain chemistry is a total mystery. Why those factors and not 25 others? Why lines and not, say, enclosed spaces? But no matter. It’s a clue.
And it’s enough for this kind of research — computational psychiatry — to be a promising new avenue in mental health. Humans haven’t figured out how the squishy mass of goo between our ears creates the vast complexity of human behavior; maybe computers can see something we’re missing.
“Machine learning is our only hope,” says Konrad Kording, a computational neuroscientist at the University of Pennsylvania.
Right now, psychiatry is a mess. Diagnostic categories are controversial. Theories of mental illness are rooted in the same era of psychodynamics that gave us the idea that autism is caused by refrigerator mothers. And nothing — not genetics, not neurochemistry, not anatomical differences — has made it possible to pinpoint why any given treatment works for some patients and not for others.
If computational psychiatry works out, it could make diagnoses and treatments much more precise. But it’s also possible that instead of solving psychiatry’s problems, computers will simply reinforce them.
Each year, nearly 18 percent of American adults experience mental illness, according to the National Institute of Mental Health. That’s 43 million people. Nearly as many—40 million—take psychiatric drugs, according to a 2016 study in JAMA Internal Medicine. Nearly 10 million of those struggle with severe mental disorders that majorly affect everyday life. Many of those people will have to try multiple treatments before they find one that works, if one exists at all. Studies on depression have shown that the first treatment works for only about 30 percent of patients, at best. “Most people will go through medication after medication,” Chekroud says.
Adding to the problem is the fact that it can take up to three weeks for even a successful depression medication to start working, says Evian Gordon, chairman of the Brain Resource Company and the principal investigator of iSPOT, a series of studies analyzing how a variety of factors, such as genetics, brain scans, and performance on cognitive tests, predict someone’s responsiveness to common drugs for depression. The time to do trial and error really adds up.
In a paper published in The Lancet in 2016, Chekroud and colleagues showed that their algorithm doubled the likelihood of a person responding to the first round of treatment, boosting it to 60 percent. After that paper and one in JAMA Psychiatry in 2017, Chekroud felt good enough about the technology to launch a company, Spring Health, to provide the algorithm to doctors in clinics. Doctors who use the algorithm can then decide whether its suggestions are medically appropriate, given a patient’s overall health and other medications.
That’s a rare example of computational psychiatry venturing into the real world. Gordon’s iSPOT system is not yet in clinical use, though he’s optimistic that combining multiple lines of information about patients can make sense of how the individuality of a brain determines which treatment works.
“You can give someone one SSRI, and they respond, and you give them another SSRI and they don’t respond. Why is that?” Gordon says. “It’s really clear that small differences in the brain can be quite significant depending on that person’s configuration of their genetics, the way they were bonded, their early life experience — have they had an early-life trauma? — and what type of cognitive profile they have. That’s where the data analytics gets powerful.”
Other research aims to give human psychiatrists an automated assist at keeping tabs on their patients. At IBM’s Thomas J. Watson Research Center in Yorktown Heights, New York, computational psychiatrist Guillermo Cecchi is developing algorithms that can detect how well a mental health patient is doing on any given day. Cecchi’s technology analyzes the patient’s voice and speech for qualities such as disjointed, fragmented speech, which can indicate that a person with psychosis is struggling.
He hopes clinicians will be able to use these algorithms to check on patients between appointments.
“The idea is, I can speak for two minutes on the phone and my psychiatrist will have a signal that will alert whether things are going good or not so good,” he says. “As a psychiatrist, I can call you up or I can reach you and say, ‘Oh, maybe you can adjust your medication or do this activity,’ or I will send a caseworker because there is indication this patient is about to enter a crisis mode.”
It’s easy to imagine that an algorithm someday could be trained to recognize specific mental illnesses from reading brain scans, like some sort of robot radiologist.
Just as Chekroud’s 25 factors predicting people’s response to depression drugs seem somewhat arbitrary, some of Cecchi’s findings seem inscrutable. In yet-unpublished results, he and his team have found that in some conditions, patients’ voices change — but only in certain frequencies — depending on whether they are medicated. It’s not clear why medication should cause the changes, which are too subtle for the human ear to catch, Cecchi says.
“Even the best psychiatrist or the best neurologist in the world will not detect that,” he says. “But it’s meaningful.”
This kind of approach might detect things that even patients themselves aren’t aware of. In a study published last fall, a genomics researcher and a psychiatrist recorded 143 volunteers speaking at various points in their days and analyzed genetic changes in their white blood cells. The subjects were likelier to use certain words indicative of being stressed out — especially adverbs such as “really” and “incredibly” — at times when stress also could be detected by changes in gene expression, even if the study participants didn’t necessarily say they felt stressed. The physical markers appear to be a more accurate gauge than what people report about themselves.
That’s the exciting potential of data-driven computational psychiatry: it can detect patterns that humans never could. If so, we might be able to improve mental health treatments without fully understanding how we’re doing it.
Any improvement in mental health care would be laudable, but too much reliance on computer analysis — leaving algorithms in charge of the asylum, so to speak — could be a dangerous proposition.
For one thing, humans will want to look under the hood of an algorithm’s calculations. It’s easy to stumble into spurious correlations in machine learning, said Penn’s Kording. Small research samples can be misleading, and even algorithms that work like a charm can be functionally useless in the real world. How many primary-care doctors (who prescribe most antidepressants) can send their patients to a high-resolution functional magnetic resonance imaging (fMRI) scanner on a routine basis? How many patients can afford it? It’s important that the measurements used to train an algorithm can be easily gathered from real patients, Chekroud says.
Other potential problems go deeper. Mental disorders can be slippery things. Despite the stolid pronouncements of the Diagnostic and Statistical Manual 5 or the International Statistical Classification of Diseases and Related Health Problems 10, two manuals widely used to make diagnoses, the definition of even familiar disorders like depression is up for debate. There have been more than 280 scales to gauge the severity of depression, says Eiko Fried, a postdoctoral researcher in psychological methodology at the University of Amsterdam. Just the seven most commonly used have 52 unique symptoms between them.
In fact, Fried says, there isn’t great evidence that one rating scale can be swapped out for another. Because mental illnesses can be assessed on different scales, researchers combining multiple data sets might inadvertently be including people in their studies who really aren’t comparable to one another.
How we assess mental illness is also a problem on a deeper level. If researchers are training algorithms to recognize “depression” as the catch-all that defines the word now, they may just be entrenching psychiatry’s problems. What’s the point of predicting remission in depression if the thing we call depression doesn’t match our current definition of it?
Computational psychiatry could also stumble badly if its underlying assumptions about the brain are flawed. The field assumes that people’s brains function similarly enough that the complexities of behavior manifest themselves in physically predictable ways. “We’re all basically carrying around the same system, and are subject to the same failure modes,” says James Kozloski, a computational neuroscientist who works with Cecchi at IBM.
This idea, that mental disorders are the result of something broken in the brain, is standard in neuroscience. If so, it’s easy to imagine a world where an algorithm could be trained to recognize specific mental illnesses from reading brain scans, like some sort of robot radiologist. But it’s debatable whether mental illness always works that way.
One alternative idea is that mental disorders don’t arise from clear-cut and identifiable disruptions in brain function. Instead, symptoms might essentially cause each other in a kind of feedback loop. A person might experience a triggering event, like the loss of a loved one, leading to sadness. Sadness, in turn, could activate insomnia, which could feed into a cycle of guilt and self-reproach. This self-sustaining network of symptoms would be reflected in brain function, of course, as everything in the mind must be. As brain function goes haywire, disruptions like neurotransmitter fluctuations might even cause more symptoms. But the disorder isn’t just the result of an off-kilter brain. Instead, something much more complicated might be happening.
For the first time, we just might have the power to sort through the massive complexity of the brain.
The bottom line of this idea, which is known as the network theory of psychopathology, is that mental disorders that manifest themselves in similar ways actually could have very different origins. If that’s true, computational approaches aren’t likely to generalize beyond small groups of patients or lead to more precise treatments.
To try to tackle that problem, some computational psychologists are testing computational methods against existing psychological theories. At the Laureate Institute for Brain Research in Tulsa, Oklahoma, for example, psychiatrist Martin Paulus has been modeling the processes that could explain anxiety. In his team’s simulations, anxiety disorders emerge when random information (the frown on the face of a stranger) is weighted too heavily over pertinent facts (most people are pretty friendly). These models hint that in real brains, irrelevant stuff gets interpreted as evidence of a scary world. Thus, therapies designed to nudge people with social anxiety into harmless situations that scare them might not be best for people who will interpret the situation as scary no matter how benign it really is.
Developing a deeper understanding of the brain will probably require knitting together all of these computational approaches, something the brand-new field hasn’t managed quite yet. Even so, there is optimism for computational psychiatry. For the first time, researchers say, we just might have the power to sort through the massive complexity of the brain.
“The field is generally on the right path,” Penn’s Kording says. “It’s just that it’s a hard path.”
This story was updated on February 9, 2018, to clarify that not all of the 280 scales for measuring depression are still in use.