Study finds associations between particular speech patterns and particular schizophrenia symptoms

The speech of people with schizophrenia has discernable patterns associated with symptoms that might be eventually used to monitor patients more closely and possibly with artificial intelligence, according to results reported today in JAMA Network Open.

Positive symptoms of schizophrenia, which include delusions and confusion, were associated with more negative sentiment and longer utterances. Negative symptoms, such as apathy and social withdrawal, were associated with a “flatter acoustic profile" and reduced verbal output. Hallucinatory behavior was associated with reduced loudness and less variation in pitch.

Corresponding author Silvia Ciampelli, M.Sc., of the Center for Clinical Neuroscience and Cognition at University Medical Center Groningen in the Netherlands, and her colleagues noted that previous research has linked linguistic markers to fluctuations in psychotic symptoms. They explained that the new study maps linguistic attributes to specific symptoms in a quantifiable way.

The study showed that “psychotic symptoms left measurable traces in language,” Ciampelli and her colleagues wrote in the study’s conclusion. “Quantifying these traces during brief, open-ended tasks enabled objective, interpretable and real-time assessment of symptom dynamics at the single-feature level,” they wrote. If validated, speech-based tools could support “precision monitoring” of patients and help prevent relapses and undesirable outcomes, they wrote.

An invited commentary on the study by Hamilton Morrin, MRCPsych, M.B.B.S., of King’s College London, and Matthew M. Nour, of the University of Oxford, echoed some of the optimism of the investigators. “Remotely captured brief speech tasks could help teams identify deviations from a patient’s own baseline and prompt timely review,” Morrin and Nour commented. But they also called out some potential pitfalls, ranging from patient fears that the recordings would be used for surveillance to the need for more subtle attributes of language such as narrative cohesion and appropriate use of metaphor to adjustments for cultures and differences in language. Morrin and Noor also noted that it will take prospective trials designed to assess clinical outcomes to prove the clinical utility of computational speech analysis in psychiatry.

The study done by Ciampelli and her co-authors included a Dutch cohort of 356 study participants and a smaller U.S. one of 72. Some of the U.S. authors of the study reported related findings about sentiments expressed in language and symptoms of psychosis last year in an article published in Cognitive Neuropsychiatry.

The Dutch participants were recruited from the University Medical Centers in Utrecht and Groningen and the U.S. participants from Zucker Hillside Hospital in the Glen Oaks neighborhood in the borough of Queens in New York. The U.S. cohort included people with bipolar 1 disorder, often characterized as the most severe form of the condition, while the Dutch group included only people with schizophrenia spectrum disorders. Speech samples were collected using open-ended questions. In the Dutch study, a total of 773 speech samples were collected at six different times by trained researchers. The U.S. researchers depended on an application on a touch-screen tablet to collect 165 speech samples. The samples were processed to extract various properties of speech, including syntax and sentiment. The researchers looked for associations between those speech components and symptoms of psychosis and their severity. In the Dutch study the symptoms were identified with the Positive and Negative Syndrome Scale, a standard way that researchers have for measuring symptoms of schizophrenia. U.S. researchers used the Brief Psychiatric Rating Scale, a different way of assessing psychotic symptoms that is also among the standard tools used by schizophrenia researchers.

In the Dutch study, Ciampelli and her colleagues found that the detected speech patterns were reliably associated with individual psychotic symptoms with “clinically meaningful accuracy” that was superior to other measures of accuracy. They distilled the speech patterns down to 11 categories, including loudness, sentence length, “sentiment variability” and formants (the native frequencies of a person’s voice). They listed some specific symptoms and the speech patterns with which they were associated. They found, for example, that blunted affect was associated with lower loudness and shorter sentences. Emotional withdrawal, lack of spontaneity and difficulty with abstract thinking were all associated with shorter sentence length.

The automated analysis of the U.S. speech samples yielded eight distinct step components. The researchers found that “thought disturbance” was associated with higher sentence similarity and longer sentences. Withdrawal was associated with reduced loudness, lower formants or timing scores.

Ciampelli and her colleague said one of the strengths of the study was the longitudinal nature of the Dutch study and its clinical diversity. They said one of the limitations was the differences between the Dutch and U.S. cohorts and that the speech “elicitation tasks and analytical pipelines” were comparable but not identical.

In their commentary, Morrin and Nour said the most likely near-term application of speech sampling and analysis would be as a decision-support toll that would be incorporated into existing ways of treating and caring for people with schizophrenia. A patient might be given weekly speech tasks that would “generate a trajectory of interpretable speech components alongside symptom self-report and other measures,” they wrote. “A deviation from that individual’s baseline could prompt a check-in, medication review or targeted assessment of psychosocial stressors and adherence to treatment.”