Scientists in the United States have developed an artificial intelligence (AI) tool, or classifier, that can diagnose posttraumatic stress disorder (PTSD) in veterans by analyzing their voices. Tests showed that the new tool could distinguish between individuals who did or did not have PTSD, with 89% accuracy. With further refinement the tool could potentially be used in a clinical setting to remotely diagnose PTSD, a condition for which the New York University (NYU) School of Medicine-led team acknowledges there is currently no objective test.
“Our findings suggest that speech-based characteristics can be used to diagnose this disease, and with further refinement and validation, may be employed in the clinic in the near future,” said senior study author Charles R. Marmar, MD, the Lucius N. Littauer professor and chair of the department of psychiatry at NYU School of Medicine.
Marmar’s team and colleagues at the Steven and Alexandra Cohen Veterans Center for the Study of Post-Traumatic Stress and Traumatic Brain Injury, and SRI International, describe the AI classifier in Depression and Anxiety, in a paper titled, “Speech-Based Markers for Posttraumatic Stress Disorder in U.S. Veterans.”
More than 70% of adults worldwide experience a traumatic event at some point in their lives, and in some countries up to 12% may suffer from PTSD. The condition leads to severe distress when faced with reminders of the traumatic event. PTSD may also be associated with relationship problems, lower academic achievement, substance abuse, and unemployment, the authors wrote.
Diagnosing PTSD remains “challenging,” they continued, and is commonly based on self-reported assessments or interviews with clinicians. The current gold standard for PTSD diagnosis, the Clinician Administered PTSD Scale (CAPS), is based on a lengthy, structured clinical interview, but it is not ideal, and some patients find it too distressing to discuss past traumatic events and their symptoms. “For these reasons, there is an imperative to develop objective measures for screening and diagnosing psychiatric disorders,” the authors wrote.
One sphere of research is looking for biological markers of PTSD, such as changes to neural structures and function, along with genomic, and immune function markers. While progress is being made, there are drawbacks and, as the authors commented, “… problems in accuracy, cost, and patient burden preclude routine use in clinical practice.”
Speech-based techniques offer an alternative and “attractive” potential approach for diagnosing different psychiatric disorders, they continued. Speech can be measured at low cost, remotely, and non-invasively. “Clinicians have long observed that individuals suffering from psychiatric disorders display changes in speech and routinely use impressions of voice quality as an elemental status examination …” Features of how we speak, such as whether the voice sounds “pressured” may indicate conditions such as bipolar disorder, while characteristics including “monotone,” “lifeless,” and “metallic,” may indicate depression.
While recent techniques developed for automating speech analysis have demonstrated encouraging specificity and sensitivity for some indications, there is relatively little known about changes in speech associated with PTSD. “Speech is an attractive candidate for use in an automated diagnostic system, perhaps as part of a future PTSD smartphone app, because it can be measured cheaply, remotely, and non-intrusively,” commented lead author Adam Brown, PhD, adjunct assistant professor in the department of psychiatry at NYU School of Medicine.
For their study, the Marmar team applied the random forest statistical/machine learning approach, which can learn how to classify individuals, based on example, to CAPS interviews with U.S. military veterans. The team recorded standard CAPS interviews with 53 Iraq and Afghanistan veterans with military service-related PTSD, and another 78 veterans without PTSD. Individuals who had potentially confounding diagnoses, such as history of substance abuse, psychiatric disorders including bipolar disorder, major depressive disorder (MDD), non-PTSD-related depression, recent exposure to traumatic events, or suicidal ideation or attempts, had been excluded. The recordings were then fed into voice software from SRI International, which generated a total of 40,526 different speech-based features, captured in short speech segments.
The AI program analyzed the results to search for patterns of specific voice features that were linked with PTSD. These included less clear speech and a lifeless, metallic tone, which are characteristics that have been reported anecdotally as useful for PTSD diagnosis. “The software analyzes words—in combination with frequency, rhythm, tone, and articulatory characteristics of speech—to infer the state of the speaker, including emotion, sentiment, cognition, health, mental health, and communication quality,” explained Dimitra Vergyri, director of SRI International’s Speech Technology and Research (STAR) Laboratory. The results suggested that the probability of PTSD was higher for markers including “slower, more monotonous speech, less change in tonality, and less activation,” the team wrote.
When used to analyze speech the resulting classifier demonstrated an overall correct classification rate of 89.1%. While the authors acknowledge that their study had a number of limitations, they suggest that the panel of voice markers could be further developed into a clinical tool. “ … we believe that our panel of voice markers represents a rich, multidimensional set of features which with further validation holds promise for developing an objective, low cost, non-invasive, and, given the ubiquity of smart phones, widely accessible tool for assessing PTSD in veteran, military, and civilian contexts,” they wrote.
“The speech analysis technology used in the current study on PTSD detection falls into the range of capabilities included in our speech analytics platform called SenSay Analytics,” Vergyri noted. “The technology has been involved in a series of industry applications visible in startups like Oto, Ambit, and Decoded Health.”
The team plans to continue to train the AI voice tool using additional data, with the ultimate goal of generating a classifier that can be used in a clinical setting.