[For more stories about the experiences and costs of war, sign up for the weekly At War newsletter.]
Post-traumatic stress disorder has long been one of the hardest mental health problems to diagnose because some patients try to hide symptoms while others exaggerate them. But a new voice analysis technique may be able to take the guesswork out of identifying the disorder using the same technology now used to dial home hands-free or order pizza on a smart speaker.
A team of researchers at New York University School of Medicine, working with SRI International, the nonprofit research institute that developed the smartphone assistant Siri, has created an algorithm that can analyze patient interviews, sort through tens of thousands of variables in their speech and identify minute auditory markers of PTSD that are otherwise imperceptible to the human ear, then make a diagnosis.
The results, published online on Monday in the journal Depression and Anxiety, show the algorithm was able to narrow down the 40,500 speech characteristics of a group of patients — like the tension in the larynx and the timing in the flick in the tongue — to just 18 relevant indicators that together could be used to diagnose PTSD. Based on those 18 speech clues, the algorithm was able to correctly identify patients with PTSD 89 percent of the time.
“They were not the speech features we thought,” said Dr. Charles Marmar, a psychiatry professor at N.Y.U. and one of the authors of the paper. “We thought the telling features would reflect agitated speech. In point of fact, when we saw the data, the features are flatter, more atonal speech. We were capturing the numbness that is so typical of PTSD patients.” As the process is refined, speech pattern analysis could become a widely used biomarker for objectively identifying the disorder, he said.
For many it can’t come too soon. The Department of Defense, which funded the study, has been on the hunt for a reliable biomarker for the disorder for more than a decade. Fueled by military dollars, research teams across the country have worked to develop brain scans, blood tests and other objective measures that might take some of the uncertainty out of diagnosing PTSD.
The most common method currently used for identifying PTSD relies largely on a subjective process in which patients answer questions about symptoms and clinicians make a judgment. That works well when patients report symptoms accurately, Marmar said, but — consciously or not — patients are notoriously unreliable.
Some active-duty troops, police officers and firefighters downplay symptoms because they fear being sidelined, or don’t want to admit weakness. Others seeking compensation for PTSD caused by the job may exaggerate symptoms to increase their payout. Complicating matters even more, disorders including insomnia and anxiety share many symptoms with PTSD, and the biases of clinicians doing assessments can sometimes shape diagnoses.
At the Department of Veterans Affairs, PTSD claims have tripled in recent years and now make up more than a fifth of all benefits claims, causing some in Congress and the department to question the accuracy of the diagnosis method. It is a common frustration across psychiatry, said Craig Bryan, a former Air Force psychologist and Iraq war veteran who now runs the National Center for Veterans Studies at the University of Utah. “On active duty, I was always worried patients were lying to me because they are afraid of what was going to happen if they admitted they were having issues,” Bryan said. “Speech analysis could potentially solve that problem.”
Bryan, who was not involved in the N.Y.U. research, said his clinic participated in a separate study funded by the Defense Department focused on a voice biomarker that delivered similarly strong results. Those results were published in February. He said speech analysis was just one of a number of objective methods for diagnosing PTSD, including testing blood or saliva for stress hormones, that are on the cusp of being reliable enough to use in a clinical setting. “The advantage of voice analysis is that you don’t need a needle, you don’t need a lab,” he said. “Potentially, all you might need is a smartphone.”
The N.Y.U. study looked at 129 male military veterans all around 32 years old who had significant exposure to combat, screening out any who also had other disorders that might complicate the study, such as depression or alcohol abuse. Researchers conducted a traditional structured interview, known as a Clinician-Administered PTSD Scale, or CAPS, to identify 52 veterans with PTSD and 77 without.
Researchers recorded each interview, then fed the audio recordings through the speech analysis software at SRI International in Menlo Park, Calif. For five years a team of scientists there has been developing speech software that can understand not only what people say, but also what emotions are expressed in how they say it. The team deconstructed the interviews into 40,526 objective speech-based features that documented tone, variation, pacing and annunciation. This same process is used to help automated customer service programs respond to irate customers.
The data was sent back to New York, where Eugene Laska, a statistician in the psychiatry department at N.Y.U., fed it through an artificial intelligence program that searched repeatedly through the thousands of features until it learned which ones best distinguished the patients with PTSD. In the end it settled on just 18.
Patients with PTSD tended to speak in flatter speech, with less articulation of the tongue and lips and a more monotonous tone, the researchers reported. “We’ve known for a long time that you can tell how someone is doing from their voice. You don’t have to be a doctor to know when someone is feeling down,” Laska said. “But this could take some of the fuzziness out of the process, and help clinicians make better decisions.”
Rather than replace traditional diagnostic interviews, he said, potential biomarkers would most likely become a tool to help psychologists make difficult calls. The results of the study are encouraging, and can be built on to create more sophisticated screening tools, said Rachael Yehuda, director of the Traumatic Stress Studies Division at the Mount Sinai School of Medicine, but she cautioned that in its current form, the algorithm may be of limited practical use.
Many patients with PTSD arrive at the psychologist’s office with a host of other disorders, including depression and substance abuse, and it is unclear how well speech analysis would fare under more complicated real-life circumstances, Yehuda said. Also unclear is whether the signals identified by the algorithm are caused by PTSD or pre-existing conditions that make patients more vulnerable to the disorder, she said.
“We should be enthusiastic but sober,” she added. “This is an important effort. But mental health conditions are complicated and I suspect there is more work to be done.”
Marmar agreed that the technology will need more work before it can be deployed to the field, but said his team hoped to eventually apply for approval by the Food and Drug Administration. “The advantage of this approach is it’s noninvasive and it will become low cost, and easy to perform,” he said. “In theory, we could use this assessment on troops anywhere in the world.”