Although the language of thinking is deliberate—let me think, I have to do some thinking—the actual experience of having thoughts is often passive. Ideas pop up like dandelions; thoughts occur suddenly and escape without warning. People swim in and out of pools of thought in a way that can feel, paradoxically, mindless.
Most of the time, people don’t actively track the way one thought flows into the next. But in psychiatry, much attention is paid to such intricacies of thinking. For instance, disorganized thought, evidenced by disjointed patterns in speech, is considered a hallmark characteristic of schizophrenia. Several studies of at-risk youths have found that doctors are able to guess with impressive accuracy—the best predictive models hover around 79 percent—whether a person will develop psychosis based on tracking that person’s speech patterns in interviews.
A computer, it seems, can do better.
That’s according to a study published Wednesday by researchers at Columbia University, the New York State Psychiatric Institute, and the IBM T. J. Watson Research Center in the Nature Publishing Group journal Schizophrenia. They used an automated speech-analysis program to correctly differentiate—with 100-percent accuracy—between at-risk young people who developed psychosis over a two-and-a-half year period and those who did not. The computer model also outperformed other advanced screening technologies, like biomarkers from neuroimaging and EEG recordings of brain activity.
“In our study, we found that minimal semantic coherence—the flow of meaning from one sentence to the next—was characteristic of those young people at risk who later developed psychosis,” said Guillermo Cecchi, a biometaphorical-computing researcher for IBM Research, in an email. “It was not the average. What this means is that over 45 minutes of interviewing, these young people had at least one occasion of a jarring disruption in meaning from one sentence to the next. As an interviewer, if my mind wandered briefly, I might miss it. But a computer would pick it up.”
Researchers used an algorithm to root out such “jarring disruptions” in otherwise ordinary speech. Their semantic analysis measured coherence and two syntactic markers of speech complexity—including the length of a sentence and how many clauses it entailed. “When people speak, they can speak in short, simple sentences. Or they can speak in longer, more complex sentences, that have clauses added that further elaborate and describe the main idea,” Cecchi said. “The measures of complexity and coherence are separate and are not correlated with one another. However, simple syntax and semantic incoherence do tend to aggregate together in schizophrenia.”
Here's an example of a sentence, provided by Cecchi and revised for patient confidentiality, from one of the study’s participants who later developed psychosis:
I was always into video games. I mean, I don’t feel the urge to do that with this, but it would be fun. You know, so the one block thing is okay. I kind of lied though and I’m nervous about going back.
While the researchers conclude that language processing appears to reveal “subtle, clinically relevant mental-state changes in emergent psychosis,” their work poses several outstanding questions. For one thing, their sample size of 34 patients was tiny. Researchers are planning to attempt to replicate their findings using transcripts from a larger cohort of at-risk youths.
They’re also working to contextualize what their findings might mean more broadly. “We know that thought disorder is an early core feature of schizophrenia evident before psychosis onset,” said Cheryl Corcoran, an assistant professor of clinical psychiatry at Columbia University. “The main question then is: What are the brain mechanisms underlying this abnormality in language? And how might we intervene to address it and possibly improve prognosis? Could we improve the concurrent language problems and function of children and teenagers at risk, and either prevent psychosis or at least modify its course?”
Intervention has long been the goal. And so far it has been an elusive one. Clinicians are already quite good at identifying people who are at increased risk of developing schizophrenia, but taking that one step farther and determining which of those people will actually end up having the illness remains a huge challenge.
“Better characterizing a behavioral component of schizophrenia may lead to a clearer understanding of the alterations to neural circuitry underlying the development of these symptoms,” said Gillinder Bedi, an assistant professor of clinical psychology at Columbia University. “If speech analyses could identify those people most likely to develop schizophrenia, this could allow for more targeted preventive treatment before the onset of psychosis, potentially delaying onset or reducing the severity of the symptoms which do develop.”
All this raises another question about the nature of human language. If the way a person speaks can be a window into how that person is thinking, and further, a means of assessing how they’re doing, which mechanisms of language are really most meaningful? It isn’t what you say, the aphorism goes, it’s how you say it. Actually, though, it’s both.
As Cecchi points out, the computer analysis at the center of the study didn’t include any acoustic features like intonation, cadence, volume—all characteristics which could be meaningful in interpreting a person’s pattern of speaking and, by extension, thinking. “There is a deeper limitation, related to our current understanding of language and how to measure the full extent of what is being expressed and communicated when people speak to each other, or write,” Cecchi said. “The discriminative features that we identified are still a very simplified description of language. Finally, while language provides a unique window into the mind, it is still just one aspect of human behavior and cannot fully substitute for a close observation and interaction with the patient.”