Subtle differences in word use can indicate psychosis risk, and machine learning can help identify it.
Researchers at Emory University in Atlanta, GA, and Harvard University in Boston, MA, used a machine-learning technique to analyze language in a group of at-risk young people.
They found that they could predict which individuals would go on to develop psychosis with an accuracy of 93%.
A recent npj Schizophrenia study paper describes how the team developed and tested the method.
Senior study author Phillip Wolff, a professor of psychology at Emory University, explains that earlier research had already established that “subtle features of future psychosis are present in people’s language.” However, he noted, “we’ve used machine learning to actually uncover hidden details about those features.”
He and his colleagues devised their machine-learning approach to measure two linguistic variables: semantic density and use of words relating to sound.
They concluded that “conversion to psychosis is signaled by low semantic density and talk about voices and sounds.”
Low semantic density is a measure of what the team refers to as “poverty of content” or vagueness.
“This work,” note the authors, “is a proof of concept study demonstrating that indicators of future mental health can be extracted from people’s natural language using computational methods.”
Machine learning and psychosis symptoms
Machine learning is a type of artificial intelligence in which computers “learn from experience” without scientists having to program the learning explicitly.
A machine-learning system looks for patterns in a known set of data and decides which patterns identify specific features. Having “learned” what these features are, it can then tirelessly identify them in a new set of data.
Machine learning can spot patterns in people’s use of language that even doctors who have undergone training to diagnose and treat those at risk of psychosis may not notice.
“Trying to hear these subtleties in conversations with people is like trying to see microscopic germs with your eyes,” explains first study author Neguine Rezaii, a fellow in the Department of Neurology at Harvard Medical School.
However, it is possible to use machine learning to find certain subtle patterns hiding in people’s language. “It’s like a microscope for warning signs of psychosis,” she adds.
Rezaii began working on the study while she was a resident in the Department of Psychiatry and Behavioral Sciences at Emory University School of Medicine.
Psychosis is a state of mind in which it can be difficult to tell the difference between what is real and what is not.
When a person enters this state of mind, doctors call it a psychotic episode. During such an episode, people experience disturbed perceptions and thoughts. Delusions and hallucinations are common symptoms of psychosis.
During a psychotic episode, a person may display inappropriate behavior or talk incoherently. In addition, they may experience sleep disruption and become socially withdrawn, depressed, and anxious.
In the United States, about 3% of people will experience a period of psychosis during their lifetime, according to figures from the National Institute of Mental Health, which is one of the National Institutes of Health (NIH).
Improving early diagnosis of psychosis risk
Psychosis is a hallmark of schizophrenia and other severe long-term mental health conditions.
The warning signs of psychosis usually begin during the mid to late teenage years with a cluster of psychosis symptoms that doctors describe as prodromal syndrome.
Around 25–30% of teens who develop prodromal syndrome will develop a psychotic illness such as schizophrenia.
From interviews and tests of cognitive ability, doctors with the appropriate training can usually predict which people with prodromal syndrome will go on to develop psychosis with an accuracy of around 80%.
Scientists are trying various approaches to improve this prediction rate and make the diagnostic process more accurate and straightforward. Machine learning is one of these approaches.
Prof. Wolff and his team began their study by getting their machine-learning system to identify the language norms of everyday conversation.
They fed the system online conversations from 30,000 users of Reddit. Reddit is an online news, content rating, and discussion platform where registered users can converse about various topics.
The team used Word2Vec software to analyze individual words in the conversation. The software maps words so that those that have similar meanings are near each other in “semantic space,” while those that have very dissimilar meanings are far away from each other.
The researchers added another program to the system to extend its ability to analyze semantics. Previous studies have confined this analysis to measuring semantic coherence, which looks at how people use words across sentences.
However, semantic density goes a step further and also assesses how people organize their words into sentences. The team suggests that this is a better indicator of the mental processes that people use to form sentences.
After training the machine-learning system to establish a “normal baseline,” the team then fed it the conversations from diagnostic interviews of 40 participants in the North American Prodrome Longitudinal Study (NAPLS).
NAPLS is a multisite, 14-year project that aims to improve doctors’ ability to diagnose young people who might be at risk of developing psychosis and to understand the reasons.
The team then compared the machine-learning analysis of the NAPLS conversations with the baseline data. They also compared it with follow-up data that showed which participants went on to develop psychosis.
The results revealed that participants who later developed psychosis tended to use more sound-related words than the baseline, and they also used words of similar meaning more frequently.
“If we can identify individuals who are at risk earlier and use preventive interventions,” explains co-author Prof. Elaine Walker, “we might be able to reverse the deficits.”
“There are good data showing that treatments like cognitive-behavioral therapy can delay onset and perhaps even reduce the occurrence of psychosis,” she adds.
The team is now putting together more extensive collections of data and plans to test the new machine-learning technique with other brain and psychiatric conditions, such as dementia.
“This research is interesting not just for its potential to reveal more about mental illness but for understanding how the mind works — how it puts ideas together.”
Prof. Phillip Wolff