An AI Program Listens to How Emergency Room Patients Describe Their Injuries and Flags Who May Develop PTSD
Researchers at NYU Langone Health in New York built an AI program that reads a patient's own account of what brought them to the emergency room and estimates their risk of post-traumatic stress disorder a month later. In a study of 145 patients, the team says it caught more than 95 percent of those at risk. The study is small, comes from one hospital system, and does not yet say how often the program raised false alarms.
Researchers at NYU Langone Health in New York published a study today in the journal Nature Communications describing an AI program that tries to spot, on the day of an injury, which emergency room patients are likely to develop post-traumatic stress disorder, or PTSD.
PTSD is a condition that can follow a frightening or violent experience, such as a car crash or an assault. People can relive the event, avoid reminders of it, and feel constantly on edge for months or years. It is common after injuries: the hospital notes that about 43 million people a year arrive at American emergency rooms with physical injuries, and past studies suggest more than a quarter of them go on to develop PTSD. Yet only about 28 percent of emergency rooms routinely check for it, and doctors there often have just 15 to 20 minutes with each patient.
The program, called PREVIA, works from something nearly every patient is already asked: what brought you in today? It takes the patient's answer, in their own words, and passes it through several AI programs of the same kind as ChatGPT, each with its own job. One looks for warning signs during and just after the event, such as memory gaps or feeling detached from one's own body. Another weighs risk factors, such as past mental health problems, against protective ones, such as having people to lean on. The results are then combined with basic facts, such as the kind of injury, into a risk score with a plain-language explanation of why.
The team tested it on the stories of 145 emergency room patients and checked the predictions against a standard interview that clinicians use to diagnose PTSD, given about a month later. According to the hospital, the program correctly identified more than 95 percent of the patients who were at risk. In a smaller comparison of 30 patients who went on to develop PTSD, it flagged all 30, while an emergency room clinician who specializes in PTSD identified 89 percent of them.
"PREVIA can serve as an AI screener that identifies at-risk patients and explains why," said Katharina Schultebraucks, the study's senior author. The team built it from AI programs that anyone can download and use, without extra training, so that other hospitals can test it for themselves.
The results have real limits. The study is small, and the head-to-head comparison with the clinician covers only 30 people. All the patients came from one hospital system, and the program has not yet been tried in other hospitals. The hospital's announcement reports how many people who developed PTSD the program caught, but not how many people it flagged who never developed it, which matters for a screening test: a program that flagged everyone would also catch every case. And it is not yet in use with patients. For now, it is a promising early result that other researchers will need to repeat.
Sources
- NYU Langone Health — AI predicts PTSD risk based on stories about what brought patients to the emergency room (news release)nyulangone.org · primary
- Crossref — Nature Communications paper record: LLM ensemble PREVIA predicts future PTSD risk from emergency department trauma narrativesapi.crossref.org · primary
- EurekAlert — the release as distributed to science reporterseurekalert.org
- News-Medical — report on the studynews-medical.net