Artificial intelligence tool can detect PTSD from patient voices, study suggests
A computer program could determine who had PTSD and who didn't with 89 percent accuracy based on voice patterns alone.
A computer program can analyze the voices of war veterans and determine which of them have post-traumatic stress disorder with 89 percent accuracy, according to a study published this week in the journal Depression and Anxiety.
The results suggest voice could be a biomarker for PTSD, providing the basis for an objective test to diagnose the disease.
The preliminary study was conducted by researchers at the New York University School of Medicine and funded in part by the Army.
“This could really change the landscape of PTSD treatment,” said Thea Gallagher, clinic director of the University of Pennsylvania’s Center for the Treatment and Study of Anxiety. Gallagher was not involved in the study but often works with PTSD patients.
“Helping people begins with diagnosis,” she said. “And it’s hard to know how many people we’re missing right now.”
PTSD affects more than 24 million Americans, according to national statistics. Many are military veterans or members of law enforcement, but the disorder also affects people who have experienced assault or abuse, natural disasters, terrorist attacks, or other traumatic events.
Currently, PTSD is diagnosed through a clinician interview in which a mental-health professional determines whether a patient is experiencing ongoing distress because of a triggering event. The accuracy of that diagnosis is largely dependent on how much information the patient shares.
But individuals with PTSD often repress traumatic memories and associated emotions. Many patients hide their symptoms out of fear or shame, said Charles Marmar, coauthor of the study and chair of the psychiatry department at the NYU School of Medicine.
“They may be having nightmares three times a week,” Marmar said, “but do they want to tell you that?”
Some patients worry they’ll seem weak or be disqualified from promotions in military and law enforcement jobs, he said.
Cultural and racial biases of the clinician or patient can also affect diagnosis. An objective test could remove these barriers, Marmar said.
In the study, researchers recorded standard clinical interviews with 53 male veterans from Iraq and Afghanistan who had PTSD, as well as 78 without it. The recordings were then fed into voice software that recognized more than 40,000 speech features, such as tone and frequency of pauses. The content of the interviews was not analyzed.
An artificial intelligence tool was then used to identify 18 speech features that were most useful in determining if a person had PTSD. These included monotonous speech, slower speech, more pauses, and a lifeless, metallic tone — all features associated with numbing, Marmar said.
Numbing, or the blunting of emotions, is often seen in patients with untreated PTSD as they try to avoid emotions that can feel overpowering. It can also be a symptom of depression.
To ensure the computer program was accurately identifying PTSD and not any other mental illness, the researchers excluded participants with a diagnosis of major depression. They also used statistical analysis to confirm the findings weren’t influenced by traumatic brain injury or alcohol-use disorder.
Researchers aren’t sure why PTSD would change someone’s voice. They theorize that traumatic events change brain circuits that process emotion and muscle tone, which in turn affect voice.
“People in active panic or fear sound different than when at coffee with a friend,” Gallagher said. Human beings know that intuitively, she said, which is why clinicians often pay attention to a patient’s voice quality during an interview.
While the technology used in the study is promising, it’s far from being put into practice. Researchers have to test another sample of patients and expand the pool to include female veterans and civilian subjects.
Down the line, they hope to develop an app that could take in audio and screen people for PTSD in minutes.
“It could be a highly efficient, relatively low-cost way of doing population-level screening,” Marmar said.
It wouldn’t provide a definitive diagnosis — that would still be done by a professional — but the app could tell individuals whether they need to follow up with a specialist.
Marmar sees it being used by primary care doctors or mental-health hotlines.
Gallagher said it could be useful as an over-the-phone screening tool for patients who call for appointments at the Penn clinic.
While she worries about people trying to “become their own doctors,” Gallagher said the application could help validate people’s symptoms so they know it’s not in their heads and it is worth seeking treatment. “Anything that gets people closer to seeking treatment is very important,” she said.
The study authors also are working to develop a blood test that could identify PTSD based on more than two dozen biomarkers found in blood.