man and woman talking

What Does Trauma Sound Like? AI Tool Mines Voice Cues to Detect PTSD

Every day, approximately 17 veterans die by suicide. Many others suffer the invisible wounds of posttraumatic stress. MITRE and VA Boston Healthcare System researchers are using artificial intelligence to analyze voice data—and help improve the health of those who served.

Retired Colonel Ralph Kauzlarich is painfully familiar with the invisible wounds of war. Known as Col. K to those who served with him, Kauzlarich spent “26 years, two months, and six days” as an Army officer and deployed to a war zone eight times.

In 2007, he commanded an infantry division known as the “2-16,” or “Rangers.” His 800-person unit served at the tip of the spear in Baghdad during the critical first days of Operation Iraqi Freedom. In their 14-month deployment, Kauzlarich lost 14 of his fellow soldiers, or “Ranger buddies”—an experience captured in the best-selling book by Washington Post reporter David Finkel.

For injuries, soldiers were treated and sent back to fight or evacuated for further treatment in a stateside hospital. But, Kauzlarich says, those who suffered from the veiled pain of posttraumatic stress disorder (PTSD) endured a longer path to recovery.

“For some people, it takes decades to get to the point where they no longer feel hopeless, that there’s something to live for, that they shouldn’t need to feel survivor’s guilt or have moral injuries,” he says. “Those are the things that the veterans really struggle with.”

And the road to healing begins with diagnosis. With the help of MITRE-developed machine learning technology, doctors can detect PTSD symptoms earlier.

Reading Between the Lines: Hearing What’s Unspoken

Computational linguistics expert Qian Hu, Ph.D., and her MITRE research team created and patented a way to predict the likelihood of PTSD, using multiple sources of information, including vocal features.

The tool, called Mining Audio Cues from PTSD Interviews (MACPI), provides a new method for analyzing voice data to help improve diagnoses—and get veterans the care they need. Hu says the technology provides “the capability to process and analyze PTSD interview recordings and detect features from both voice and spoken words associated with diagnostic symptoms of PTSD. They can correlate question and response, speech rate, vocal characteristics, and emotional state reflected by the speech.” 

This work was funded by MITRE’s Independent Research and Development Program.

This is the least we can do to pay back to the people who’ve sacrificed so much for our nation.

Qian Hu, Ph.D., computational linguistics expert

Joining Forces Against a Common Challenger

In the initial stages of developing MACPI, Hu drew on her work with other voice-related systems. Then she learned that Brian P. Marx, Ph.D., a researcher and deputy director of the Behavioral Science Division at the National Center for PTSD at VA Boston Healthcare System and Boston University Chobanian and Avedisian School of Medicine, had cataloged hundreds of veteran voice recordings, as part of a Department of Defense–funded study.

The two researchers decided to team up and analyze the recordings to see what they could learn.

“We have several effective treatments for veterans with PTSD, but we first need to quickly and accurately determine their diagnostic status,” Marx says.” MACPI is a breakthrough technology that can aid clinicians in rendering a correct PTSD diagnosis.”

Hu explained her driving theory for the research. “The data from the study of controlled groups inspired me to ask, ‘Can we apply machine learning to identify fundamental signals or high-value features from the voice that can be indicative of PTSD diagnostic symptoms?’”

There's information in not only what the speaker says but how—what they convey without actually stating it.

Human-Machine Pairing is Key

She also noted that the technology doesn’t rely on voice alone for detection. Its multi-layered approach includes input from clinicians.

“Even with different sensors, or different signals, system detection results can be enhanced with human input. That's why the system also includes a human clinician in the loop, who’s familiar with the history of the patient and can select the source of information channels,” she explains. “Then the system can use the algorithm to make the prediction. This offers a human-machine-patterned approach to the prediction.”

Hu says a sense of compassion and desire to do something to support veterans inspired her and her team. “I feel this helps the good people who’ve given their whole life to make this country safer and better. This is the least we can do to pay back to the people who’ve sacrificed so much for our nation,” she says.

Kauzlarich agrees, saying this technology may be a lifeline to veterans. “I think this will allow them to start healing.”

Although MACPI comes more than 20 years after the beginning of Operation Iraqi Freedom, and five decades after Vietnam, Kauzlarich says it can still help a veteran at any season of their life. “As long as there’s a veteran struggling with posttraumatic stress injuries, it’s never too late. We’ve got to get that veteran the help they need.”

For information on MITRE’s MACPI or licensing options, contact the MITRE Technology Transfer Office at techtransfer@mitre.org.