← All research storiesHealth
An AI read ER patients’ stories about their injuries and flagged who was at risk of PTSD a month later
Using only patients’ answers to “What brought you to the ER?”, a team of AI language models picked out more than 95% of people at risk of PTSD in a 145-patient study.
What the scientists did
Post-traumatic stress disorder (PTSD) can develop after a frightening event like a car crash or serious injury. Many injured people pass through an emergency department (ED), but staff there are usually too rushed for long mental-health questionnaires. The study notes that only about 28% of U.S. EDs routinely screen for PTSD.
Researchers asked 145 injured patients at two New York City trauma centers one routine question: “What brought you to the ED?” Their spoken answers were transcribed and given to PREVIA, a team of open-source AI language models. One layer looks for early warning signs in the story, such as loss of control or feeling disconnected from one’s body; another weighs risk factors and protective factors like social support; a third combines these; and a final step adds age, sex and trauma type to produce a risk score with a plain-language explanation.
About a month later, the patients were assessed with the gold-standard clinical interview for PTSD so the predictions could be checked.
Words to know
PTSD (post-traumatic stress disorder)
A mental health condition that can follow a terrifying event, with symptoms that last at least a month.
- Flashbacks or nightmares about the event
- Avoiding places or people that are reminders
- Feeling constantly on edge or easily startled
Language model ensemble
Several AI language models, like the ones behind chatbots, working together with each handling a different part of the job.
- Llama, which in this study tended to focus on signs of risk
- Mistral, which tended to focus on protective factors
Sensitivity
The share of people who truly have a condition that a test correctly flags.
- 100% sensitivity means no true cases were missed
- A test can have high sensitivity but still flag many people who turn out fine
What they found
In its best setup, PREVIA flagged 95.4% of the patients who were at risk. Across the full group, its overall accuracy score (called AUROC) averaged about 78.5%, and about 47% of the people it flagged actually developed PTSD symptoms, according to reports on the paper.
In a head-to-head test on 30 stories that included nine PTSD cases, PREVIA caught all nine, while an experienced emergency clinician who specializes in PTSD caught eight. Both had 50% precision, meaning half of the people each one flagged did not develop PTSD.
Why it matters to you
If tools like this hold up, someone hurt in an accident could be flagged for a mental-health follow-up before leaving the ER, without extra forms. That matters because the first weeks after trauma may be the best time for help. This is a research tool, not a diagnosis. If you or someone you know is struggling after a scary event, talk to a doctor, counselor or trusted adult.
Where this could lead
Quick PTSD screening in the ER
Hospitals could use the patient’s own description of what happened, already collected at intake, to flag who needs a mental-health follow-up.
3–10 years, after larger testsIndependent testing at other hospitals
Because PREVIA uses open-source models with no extra training, other hospitals can test it on their own patients.
1–3 yearsThese are possibilities the research points toward, not promises. Most early findings take years of testing before they reach everyday life, and some never do.
Why scientists care
Mental-health problems after injuries are common but often missed because screening takes time. Showing that one routine question carries enough signal to predict risk, and that an AI can explain its reasoning, could make early screening practical even in crowded emergency rooms.
Keep in mind
This was a small study of 145 patients at two New York hospitals, and the head-to-head comparison used only 30 stories and one clinician. Many people the AI flagged did not develop PTSD. The authors say it needs testing in larger, independent groups, in other languages and over longer time periods before any clinical use.
How far along is the science?
- Peer-reviewedReviewed by independent experts for a journal before publication. Peer review is a quality check, not a grade of evidence strength: confidence still depends on study design, limitations, replication and the wider evidence.
- PreprintA manuscript shared publicly, often before or alongside journal review. Check its separate journal and review status (each story says what we know). Results may change.
- Conference talkPresented to other scientists at a meeting. Usually short and early; a full paper may come later.
- Agency reportPublished by a science agency such as NASA, NOAA or NIH. Reviewed internally, but not by an outside journal.
- Press releaseAn announcement from a university or company. Useful, but always check the research it describes.
How soon could it reach you?
- Could reach you soonCould show up in products, services or advice soon. Our rough editorial guess is about 1 to 3 years, not a forecast from the research.
- A few years outPromising, but needs more testing or engineering first. Our rough editorial guess is several years, with a lot of uncertainty.
- Long-term scienceBasic research that builds knowledge. Real-world uses, if any, are far off.
- Small stepA small update to earlier work. Useful to researchers, but not a change you would notice yet.
Every story links to its original source: the paper itself, or the conference program when no paper is out yet. Explanations are written with AI, usually from the paper’s abstract (each story says what it was written from), then checked against the source. We only use journals, recognized preprint servers, science agencies, scientific conferences and established science news outlets.