Machine-learning algorithm predicts psychosis from speech

Machine-learning algorithm predicts psychosis from speech

3150 2100 Peter Stevenson, PhD

Algorithms designed to find hidden clues within everyday language can predict those at risk of developing psychosis, a new study reports.

Using machine learning, researchers from the Department of Psychology at Emory University (Atlanta, GA, USA), were able to perform a deep analysis of speech and determine linguistic indicators of psychosis. “Norms” in conversational language from 30,000 social-media discussions were fed into a software platform which collated words based on usage, similarity to other words, and meaning. This baseline data was then compared to extracts from clinically led interviews of 40 participants.

A key finding was that more frequent use of words associated with sound, for example “voice,” “hear”, “chant”, “loud” – and “sound” itself – was a significant predictor of psychosis (93% accuracy). In addition, low semantic density, or vagueness, was another strongly-linked linguistic marker for psychosis.

“It was previously known that subtle features of future psychosis are present in people’s language, but we’ve used machine learning to actually uncover hidden details about those features,” notes Phillip Wolff, senior author of the study.

Using machine learning in this way may offer extra power to clinicians who usually rely on structured interviews and cognitive tests to predict psychosis. The use of machine learning bolsters capabilities, offering the chance to detect more nuanced linguistic warning signs at a much more sensitive level. “[For clinicians], trying to hear these subtleties in conversations with people is like trying to see microscopic germs with your eyes,” commented Neguine Rezaii, who is first author of the paper.

“The automated technique we’ve developed is a really sensitive tool to detect these hidden patterns. It’s like a microscope for [the] warning signs of psychosis.”

Psychotic disorders such as schizophrenia have a typical onset in the third decade of life, with some early warning signs, or “prodromal syndrome”, beginning as early as 17 years of age. While there is no current cure for psychosis, earlier detection of risk could offer more hope to sufferers. “If we can identify individuals who are at risk earlier and use preventive interventions, we might be able to reverse the deficits,” added co-author Elaine Walker.

“There are good data showing that treatments like cognitive-behavioural therapy can delay onset, and perhaps even reduce the occurrence of psychosis.”

Going forward, the researchers are now working on larger data sets, and plan to test their methods in neuropsychiatric diseases 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,” summarised Professor Wolff.

“Machine learning technology is advancing so rapidly that it’s giving us tools to data mine the human mind.”


Rezaii N, Walker E, Wolff P. A machine learning approach to predicting psychosis using semantic density and latent content analysis. npj Schizophrenia. 2019;5(1):9.

The whisper of schizophrenia: Machine learning finds ‘sound’ words predict psychosis. Neuroscience News. June 13, 2019.

Brainwork is supported by unrestricted grants from: