AI Could Help Detect Schizophrenia From People's Speech

Geralt/Pixabay

A recent study published in the Proceedings of the National Academy of Sciences (PNAS) shows the potential of using artificial intelligence (AI) language models to find speech signatures in patients with schizophrenia.

“These findings shed light on the neural basis of semantic representation in schizophrenia,” wrote lead author Matthew Nour, MD, PhD, of the University College London (UCL) Queen Square Institute of Neurology, along with colleagues Daniel C. McNamee, Yunzhe Liu, and Raymond J. Dolan.

An estimated 24 million people live with schizophrenia, according to the World Health Organization (WHO). It is one of the top-15 leading causes of disability worldwide, and the average potential life lost for Americans with schizophrenia is over 28 years according to estimates from the National Institute of Mental Health (NIMH) of the National Institute of Mental Health (NIH), a part of the U.S. Department of Health and Human Services.

Schizophrenia is a serious and chronic mental disorder characterized by hallucinations, delusional beliefs, and disturbances in perception, behavior, and thought. The word schizophrenia comes from the Greek words ‘schizo’ (splitting) and ‘phren’ (mind). The term was coined by Swiss psychiatrist Eugen Bleuler (1857–1939) in 1908.

“Human cognition is underpinned by structured internal representations that encode relationships between entities in the world (cognitive maps),” wrote the researchers in the new study. “Clinical features of schizophrenia—from thought disorder to delusions—are proposed to reflect disorganization in such conceptual representations.”

The study tasked 52 participants, half of whom had schizophrenia, with two five-minute fluency tasks. For the category task, participants were asked to name as many words as they could that belong to the category of “animals”; the other task was to name as many words as possible that start with the letter P.

The researchers’ approach was to test whether an AI algorithm could predict the words the participants provided, and to gauge the difference in predictability compared to those with schizophrenia. The scientists hypothesized that the answers provided by those with schizophrenia would be less predictable for the AI algorithm.

The team used an open-source library called fastText, a pretrained Natural language Processing (NLP) word-embedding model for building efficient sentence classification and learning of word representations. Facebook Artificial Intelligence Research (FAIR), now part of engineering at Meta, developed fastText. In under five minutes, fastText can classify a half-million sentences among 300,000 categories.

The answers from control-group participants were more predictable by the AI algorithm than those produced by participants with schizophrenia. The more severe the schizophrenia, the greater the difference in predictability. “At a behavioral level, patients with schizophrenia showed reduced semantically guided word sampling during a verbal fluency task (a marker of “looser” conceptual organization),” the scientists reported.

The researchers hypothesized that the mental representation of cognitive maps may explain their finding. To test this hypothesis, the researchers conducted magnetoencephalography (MEG) scans on the brain regions associated with learning and storing cognitive maps while participants were tasked to learn the sequential relationship between eight task pictures and during a post-task, five-minute resting-state scan.

Artificial Intelligence Essential Reads

“In line with our hypothesis, the influence of semantic similarity on behavior was reduced in schizophrenia relative to controls, predicted negative psychotic symptoms, and correlated with an MEG signature of hippocampal ripple power (but not replay),” reported the scientists. “The findings bridge a gap between phenomenological and neurocomputational accounts of schizophrenia.”