Language patterns may predict psychosis

At a Glance

  • In a proof-of-concept study, spoken language patterns predicted which people at risk for psychosis would progress to full psychosis within two years with 93% accuracy.
  • The ability to identify people who may experience psychosis could allow for early intervention strategies.

Psychotic disorders such as schizophrenia can be highly disabling. An episode of psychosis involves experiences that aren’t based in reality. These can include hallucinations and delusions, such as feeling that people are trying to harm you. If researchers could identify when people with psychotic disorders are verging on psychosis, promising methods to delay or stop the process could be tested.

Studies suggest that language patterns may help predict if someone is likely to experience psychosis. Drs. Neguine Rezaii, Elaine Walker, and Phillip Wolf of Emory University tested whether machine learning could help identify such patterns. They used sophisticated computer programs to analyze patterns of speech from 40 people enrolled in a long-term study of youth who are at risk of developing psychosis. The participants were enrolled because of unusual patterns of thought, perception, and communication.

The researchers collected recordings of interviews when participants joined the study. The team used machine learning to rate the complexity of speech and categorize the subject matter used by participants. After the interviews, participants were followed for two years to determine who developed psychosis. The team looked for specific patterns of language that might mark the earliest phase of illness.

The study was funded in part by NIH’s National Institute of Mental Health (NIMH). Results were published on June 13, 2019, in NPJSchizophrenia.

The machine learning system was trained using language samples from 30 of the 40 participants. It found that samples from those who progressed to psychosis had lower semantic density (less meaning conveyed in their sentences) than samples from a large database of normal language. The system’s predictive value was then tested against the remaining 10 participants’ outcomes. This language trait could predict which of the remaining participants would later develop psychosis with 80% accuracy.

The system also found that words related to voices and sounds were used much more frequently in those who later had a psychotic episode than in conversations among 30,000 contributors on Reddit. The social network was used for comparison because the casual language used online closely matched that in the study interviews.

Using the combination of differences in semantic density and word choice, the system could predict which of the remaining 10 participants would develop psychosis with 93% accuracy.

“If we can identify individuals who are at risk earlier and use preventive interventions, we might be able to reverse the deficits,” Walker says. “There are good data showing that treatments like cognitive-behavioral therapy can delay onset, and perhaps even reduce the occurrence of psychosis.”

The results need to be confirmed in larger groups of people at risk for psychosis. This type of machine-learning language approach has the potential to detect emerging mental illness as well as to provide insights into how the human brain normally processes thoughts and ideas.

—by Sharon Reynolds


National Institutes of Health