New York – November 09, 2026 -- An AI tool called PREVIA flagged 100 percent of 30 emergency department patients who went on to develop post-traumatic stress disorder (PTSD), while a PTSD-specialist ED clinician identified 89 percent of the same cases, according to a study led by NYU Langone Health.
PREVIA identifies more than 95.4 percent of at-risk patients from 145 trauma narratives
The study, published November 9 in Nature Communications, ran a set of AI models on the recorded and transcribed trauma narratives of 145 ED patients. The models correctly identified more than 95.4 percent of at-risk patients. Results were validated against diagnoses from the Clinician-Administered PTSD Scale for DSM-5 (CAPS-5), the gold-standard clinical interview.
The tool analyzes patients' answers to one routine question: what brought them to the ED. Senior author Katharina Schultebraucks, PhD, said PREVIA can screen patients "without adding any time to intake in busy emergency rooms."
Only 28 percent of US emergency departments routinely screen for PTSD
About 43 million patients are admitted to US emergency departments each year with physical injuries, and more than a quarter of them develop PTSD, according to past studies. Only about 28 percent of EDs routinely screen for the condition, and only two-thirds of those use validated tools. Questionnaires such as the Peritraumatic Dissociative Experiences Questionnaire and the Posttraumatic Adjustment Scale forecast PTSD effectively, but clinicians typically have only 15-20 minutes per patient and prioritize immediate medical concerns.
The researchers identified the first month after trauma as the key window for finding at-risk patients, because many patients lose contact with the hospital after that.
Three-layer pipeline combines Llama and Mistral models to generate an explained risk score
PREVIA chains several AI tools according to their strengths. Llama tends to emphasize signs of PTSD risk, while Mistral highlights protective and contextual factors.
- The first layer scans narratives for symptoms during the trauma, such as loss of control or amnesia, and early ED symptoms such as feeling disconnected from the body.
- The second layer weighs risk factors, including pre-existing mental health problems and lower socioeconomic status. It also weighs protective factors such as coping skills and strong social support.
- A third model synthesizes these assessments into a structured summary, reducing the chance that any single model dominates the results.
- A final layer adds demographics (age, sex), trauma type and known PTSD prevalence rates to produce a risk score with a plain-language explanation.
Experiments showed the three-layer design consistently outperformed simpler configurations across most performance metrics. Unlike AI that succeeds or fails for unknown reasons, PREVIA grounds its predictions in established PTSD risk factors and explains its reasoning.
Open-source build requires no fine-tuning, allowing other hospitals to test it independently
PREVIA runs on open-source models with no training or fine-tuning. Other hospitals can therefore test it without proprietary software or access to NYU Langone's patient data, which Schultebraucks called essential for widespread adoption. She is associate professor in the departments of psychiatry and population health at NYU Langone and co-director of computational psychiatry at NYU Grossman School of Medicine.
The National Institute for Mental Health funded the study under grant R01 MH129856. Co-authors include Eric Oermann of the Department of Neurosurgery and Stephen Wall of the Ronald O. Perelman Department of Emergency Medicine.