Psychiatrists may soon be able to diagnose mental disorders, including psychosis and mania, by analyzing patients’ speech patterns. The same technique also promises to someday allow clinicians to spot neurological disorders such as Parkinson’s disease, Alzheimer’s, and Huntington’s disease earlier than is now possible.
Psychiatrists will be able to diagnose mental disorders, including psychosis & mania, through patient speech patterns
At a recent presentation at the National Institute of Mental Health in Bethesda, Maryland, Dr. Guillermo A. Cecchi of IBM’s Thomas J. Watson Research Center discussed how he and other researchers are using computational linguistics to quantify psychiatric conditions using recordings of speech samples. Cecchi’s lecture explored the theoretical underpinnings of this research, reviewed his recent experiments, and highlighted possible applications and implications for the future of psychiatric medicine.
Cecchi noted that researchers today have the advantage of working with powerful algorithms and large data sets that were unavailable to previous generations of scientists. “Those two things are allowing us to think that it might be possible to quantify aspects of mental dysfunction, and they point to applications outside of psychiatry as well,” he explained.
While computational analysis of large data sets is commonly used to unearth meaningful linguistic patterns in everything from email to Shakespeare’s Hamlet, psychiatry has yet to embrace its potential for aiding in clinical diagnosis. It is easy to understand why, though. Although modern computers are capable of mimicking human cognitive processes, they have been less successful at replicating emotional states and complex processes like habits and beliefs.
Psychiatry has yet to embrace potential of computational analysis of large data sets for aiding clinical diagnosis
In their first double-blind pilot experiment, Cecchi and his colleagues sought to determine whether they could chart semantic bias—a tendency to choose words that reveal one’s underlying thought processes. They did this by conducting brief interviews with regular users of the psychoactive drug MDMA (also called Ecstasy). Participants were first given either a placebo or a low or high dose of MDMA, and then were invited to talk about a person who was close to them. Ecstasy is known to induce feelings of empathy, compassion, and forgiveness, so Cecchi’s team reviewed transcripts and tallied the number of words related to those emotions. They used those tallies to predict whether each individual had taken the placebo, the low MDMA dose, or the high dose. The researchers guessed correctly 88 to 92 percent of the time.
Encouraged by the results of the pilot experiment, Cecchi then turned his attention to identifying the linguistic characteristics of schizophrenia and mania, each of which have their own distinctive patterns of thought. Patients and control subjects were asked to recount a recent dream. Cecchi plotted 100-word segments, assigning each word a node and connecting consecutive words with lines. The resulting graph of the speech patterns of manic patients showed many more “loops” than control subjects, as they frequently repeated themselves. The speech patterns of schizophrenic patients, on the other hand, resembled starbursts; their sentences would launch off in many directions without referring back to earlier statements. Diagnoses based on these graphs largely matched those made using conventional psychiatric diagnostic tools such as the Positive and Negative Syndrome Scale and the Brief Psychiatric Rating Scale.
Given these successes, Cecchi and his colleagues wanted to know if graph analysis of semantic characteristics could be used to predict the onset of psychosis, not simply confirm it. They used recordings of interviews with people in their late teens to late twenties who had no prior psychotic episodes. A subset of subjects experienced their first psychotic episode within 2.5 years of the interview. Surprisingly, Cecchi’s team found that the approach did not offer a high level of predictive accuracy. However, when the team analyzed whole sentences, rather than single words, they were better able to capture the “flight of words” that characterizes psychosis. It became clear that the speech patterns of subjects who eventually experienced a psychotic episode were less logically complex than those who didn’t. These findings suggest that language patterns may be able to be used as a means of diagnosing and predicting psychosis.
Speech patterns of subjects who experienced a psychotic episode were less logically complex than those who didn’t
These results have opened doors for diagnostic applications in many fields. For example, IBM is using the technique to identify early markers of Alzheimer’s. Researchers in Spain, Argentina, Colombia, and the United States are attempting to plot the unique language characteristics of childhood and adolescent post-traumatic stress disorder and Parkinson’s disease. Researchers at Northwestern University are looking into applications for chronic pain diagnosis.
Cecchi believes that as our understanding of the underlying science improves, the tools will eventually become available to the general public. A member of the audience asked whether we could expect to see self-diagnosis smartphone apps within the next decade. Cecchi didn’t hesitate. “I expect that within three years, not in ten,” he said with a chuckle. “I hope.”
Additional Resources
IBM Thomas J. Watson Research Center
IBM Research Panel: Neuroscience and Cognitive Systems with Dr. Guillermo Cecchi, IBM Research; Anil Ananthaswamy, New Scientist magazine; Dr. Gary Marcus, New York University; and Dr. Wayne Gray, Rensselaer Polytechnic Institute
Scale-Free Brain Functional Networks, a letter by Cecchi in Physical Review Letters
