Schizophrenia is a severe mental health disorder characterized by delusions, hallucinations, and speech disorganization. While delusions and hallucinations often dominate discussions surrounding the illness, the speech disorganization aspect, also known as thought disorder, captured the attention of Tufts cognitive science graduate student Samer Nour Eddine.
Working alongside Tufts postdoctoral scholar Victoria Sharpe and collaborators from McGill and Western universities in Canada, Nour Eddine, who is also a medical doctor, embarked on a novel approach to measure speech predictability as a marker for thought disorder. In schizophrenia, thought disorder manifests as a loosening of associations and tangentiality in speech, with individuals frequently jumping from topic to topic.
Their investigation centered around the use of artificial intelligence: They employed an advanced language model to analyze speech predictability in both healthy adults and patient populations. To do so, they fed the artificial intelligence (AI) truncated speech samples from both populations, asking it to predict the next word, and then providing it with more words from the sample, asking for a prediction again, and so on.
While no significant difference in predictability manifested with small speech segments, with larger speech segments, patients exhibited notably lower predictability compared to healthy controls.
“Imagine walking into a room in which someone is speaking, and that person is either a patient or a healthy individual,” Nour Eddine said. “If you entered and then left quickly, you wouldn’t know whether the person was a patient or not. But if you stayed in the room longer and you heard enough speech, you would start to get a sense that, with a patient, the speech is less predictable than that of a healthy adult.”
This discrepancy in predictability was more pronounced in patients experiencing more severe symptoms of schizophrenia, indicating a direct correlation with the disorder. “The project shows that language predictability can be a useful biomarker for early detection of the disorder,” Nour Eddine noted. “If we can detect it earlier, we can intervene earlier, possibly offering therapy—which has good outcomes for the disorder—earlier. In addition, we could use speech predictability to monitor how the disease progresses over time.”
On a deeper level, knowing that schizophrenia impacts the predictability of speech provides valuable insight into the nature of the disorder. “We can use what we know about prediction and language processing and look at patient data through that perspective,” said Nour Eddine. “It’s not just a useful tool to identify schizophrenia early and detect it over time; it’s also a window into the nature of the disorder itself.”