Thank you for visiting nature.com. You are using a browser version with limited support for CSS. To obtain
the best experience, we recommend you use a more up to date browser (or turn off compatibility mode in
Internet Explorer). In the meantime, to ensure continued support, we are displaying the site without styles
and JavaScript.
You go to see a psychiatrist a decade from now. Alongside the usual questions about how you’re feeling, your work and your relationships, they look at biometric data from your phone about your sleep and speech patterns, examine the results of a blood test or a brain scan and even check your postcode to see whether your neighbourhood has high levels of pollution. An algorithm then compares that information with data from thousands of other people, determines which types of treatment have worked best for people similar to you and uses its findings to suggest a treatment regime.
That’s a vision of the future called precision psychiatry. And it’s a world away from most people’s experiences today. For all the progress that has been made across the medical sciences, psychiatry remains a frustratingly imprecise discipline. Doctors use someone’s symptoms and clinical history to make a diagnosis and prescribe treatments. But mental-health diagnoses are broad and can be vague — a particular therapy might transform one person’s life, but offer little help to someone else with apparently the same disorder.
As a result, people with mental-health conditions often find themselves cycling through treatments and doses, searching, mostly by trial and error, for one that works.
To improve things, psychiatrists want to divide broad diagnoses into more useful subgroups and match treatments to individuals using a data-driven approach. Some researchers are looking at genes and brain circuits; others are turning to data on speech, sleep or social circumstances. Psychiatrists working on TheDiagnostic and Statistical Manual of Mental Disorders (DSM), the discipline’s dominant classification system, are even considering how socio-economic, cultural and environmental factors might be incorporated in the future.
“I think precision psychiatry is certainly part of the future,” says Tom Insel, a psychiatrist and former director of the US National Institute of Mental Health (NIMH), who now co-directs several mental-health-focused companies. “We’ll be able to say from the get-go, this patient needs this treatment and that’s where they’re going to do best.”
But some clinicians say that the personalized approach has yet to prove itself. Advances in genetics, brain imaging and biomarker research for psychiatry have failed to deliver on their promise before, and studies continue to reveal just how heterogeneous many psychiatric disorders are. “Psychiatry has this model that the white knight is going to ride over the hill and solve all our problems,” says Kenneth Kendler, a psychiatric geneticist at Virginia Commonwealth University in Richmond. “And we’ve been repeatedly disappointed.”
Classifying with precision
Personalized psychiatry is a catchy phrase, but it’s not a new idea.
Psychiatry has been trying to add precision to diagnosis and treatment for decades. In 1980, the third edition of the DSM introduced specific diagnostic criteria for many conditions. The aim was to boost reliability: a diagnosis of major depression or schizophrenia should mean roughly the same thing for everyone, even when issued by different psychiatrists.
But the DSM relies on symptoms to define conditions, and that proxy introduces complexities. The same symptoms appear in many conditions, and one person might have several diagnoses. Moreover, two people who have been diagnosed with the same condition can manifest it in quite different ways.
To meet DSM criteria for major depressive disorder, for example, a person must, in a two-week period, experience at least five symptoms out of a list of nine. The five must include a depressed mood or a loss of interest or pleasure; the remaining criteria include insomnia or excessive sleep, weight loss or gain, feelings of guilt or worthlessness and suicidal thoughts.
“It’s possible to have two people have the same diagnosis without a single required symptom in common,” says Vikram Patel, a psychiatrist at Harvard Medical School in Boston, Massachusetts.
By the late 2000s, researchers were increasingly questioning whether the approved disease categories were the best way to classify and treat people.
At the NIMH, Insel helped to develop the Research Domain Criteria initiative, which tried to go beyond categories and instead study and measure processes such as reward, fear and cognition, which cut across diagnostic boundaries. The idea encouraged researchers to study behaviours and brain functions that are shared across conditions. Although the programme has helped to shift research towards the underlying mechanisms of illness, it was never intended to replace the symptoms-based diagnostic system, and it does not provide a reliable way to select treatments for individuals.
Another approach to precision is to measure someone’s symptoms and behaviour with more rigour. A psychiatrist assessing psychosis, for example, listens to a person and judges qualities such as coherence, paranoia and grandiosity. That relies on human expertise, but also introduces human subjectivity.
Insel imagines a machine-learning system listening to the same person speaking and assigning more objective, reproducible measurements. “The paranoia score here is 6.1,” he suggests. “The coherence level is 2.2.”
John Torous, a psychiatrist at Beth Israel Deaconess Medical Centre in Boston and at Harvard Medical School, works on a related approach known as digital phenotyping, in which information from smartphones and wearable sensors is used to study behaviour. The term emerged around a decade ago, when researchers proposed that sensors already carried around in smartphones could provide continuous measures of various behaviours, such as physical mobility and communication, that could inform diagnosis1.
Small studies have since found links between activity levels and depressive and manic symptoms in bipolar disorder2, and between location data and symptom severity in schizophrenia3.
The difficulty has been turning these data into something that is useful in the clinic. Medical education has not generally trained clinicians to interpret this information, Torous says, and researchers have not yet developed standards and protocols for using it. “So, we are in the early days in that sense.”
Artificial intelligence could make the difference, he suggests, because it can analyse a range of data streams and package them into useful clinical information.
The approach does come with concerns about privacy: personal data on location, movement, sleep and communication can hold intimate information and could even be used to identify individuals. Furthermore, a systematic review published this year4 found that digital-phenotyping studies showed wide variation in terms of the devices and analytical approaches they used. This could limit their reproducibility and potential for clinical translation.
Biological markers
Digging into biology more deeply could offer another route to precision. As genetics and neuroscience continue to advance, some researchers hope that the insights they provide will help to separate broad psychiatric syndromes into distinct diseases, with different underlying biological mechanisms and, eventually, different treatments.
This hasn’t happened yet, and some argue that it’s a futile exercise. Whereas biological insights helped researchers to divide diabetes into two separate types nearly a century ago, “we’ve never approached that in psychiatry”, says Kendler. Mental-health conditions could be “heterogeneous a long way down”.
Still, genetics has managed to produce one of psychiatry’s more promising potential biomarkers. By looking at the genomes of thousands of people with schizophrenia, in a type of study called genome-wide association, researchers have gleaned some idea about which gene variants confer risk.
For most health conditions, it is the combined small effects of many variants that add up to susceptibility. Geneticists call this collective estimate of risk a polygenic risk score. People towards the top of the polygenic-risk distribution for schizophrenia, says Kendler, can have around a tenfold increase in risk, although the size of the relative increase varies depending on the study. Published estimates5, for example, put risk at about six times higher for the top 1% than for everyone else. But the scores cannot predict who will develop a condition.
“Is it at such a stage now that I would get that score and immediately change what I’m doing diagnostically? No,” Kendler says.
In five or ten years, he says, it is conceivable that a score distinguishing schizophrenia from bipolar disorder could help in a difficult case in which the distinction would affect treatment.
Researchers have also investigated levels of hormones and metabolites as psychiatric biomarkers. But Kendler says that such measures have “almost always” proved to be non-specific. An average biological difference between groups is hard to translate into a specific test that can flag what is wrong with a particular person, or which treatment to offer.
Brain imaging offers another version of the same problem. Magnetic resonance imaging (MRI) studies have reported large numbers of associations between features of the brain and psychiatric symptoms, but many of these have been difficult to reproduce.
A 2022 analysis6 of around 50,000 people found that associations between behavioural or mental-health traits and differences in brain structure or function were generally much smaller than previous studies had suggested.
Helen Mayberg, a neurologist at the Icahn School of Medicine at Mount Sinai in New York City, says that scans can still provide clinically useful information if researchers ask specific-enough questions.
She has spent decades mapping brain circuits involved in depression, focusing on a region of the cortex called the subcallosal cingulate. This is tucked beneath the corpus callosum, the dense bundle of fibres that connects the two sides of the brain.
In one study7, researchers used functional MRI to scan the brains of 122 people with depression. They then assigned the individuals at random to one of two treatment groups: the first received 12 weeks of cognitive behavioural therapy, and the other received antidepressant medication for the same period of time.
The researchers found that connectivity between the subcallosal cingulate and three other brain regions differed according to the outcome of the treatment. People for whom psychotherapy worked and medication didn’t typically showed one pattern, whereas people with the opposite responses showed another.
In other words, this pattern of connectivity had the potential to act as a biomarker that can help assign treatment. “A scan done at baseline could actually help you to make that decision,” Mayberg says.
But there is a considerable distance between a potential biomarker and a routine clinical test. Mayberg says she has been told repeatedly that “nobody’s ever going to do a scan to decide how to treat people”, because it is too costly and time-consuming.
Enjoying our latest content?
Log in or create an account to continue
Access the most recent journalism from Nature's award-winning team
Explore the latest features & opinion covering groundbreaking research