NYU Langone AI reads ER trauma stories to flag PTSD risk with 95.4 percent sensitivity

Roughly 43 million patients arrive in United States emergency departments each year with physical injuries, and the psychological damage from those events often goes unmeasured. Past studies cited by NYU Langone Health suggest about 27 percent of trauma patients develop post-traumatic stress disorder by one month. The broken bone gets treated and the patient is discharged, but the risk of lasting psychiatric illness travels home with them, largely undocumented and rarely discussed at the bedside.

The screening gap inside emergency medicine is well documented. Only about 28 percent of emergency departments routinely screen for PTSD, and only two-thirds of those use validated questionnaires. Clinicians typically have 15 to 20 minutes per patient, leaving little room for a structured psychiatric interview. Symptoms that best predict later illness, including dissociation and a sense of lost control during the event, are easy to miss in the first hours after trauma.

A study published on 9 November 2026 in Nature Communications argues that the missing signal is already being recorded. Researchers at NYU Langone Health and NYU Grossman School of Medicine developed PREVIA, a clinically informed large language model ensemble that reads a patient's unstructured account of what happened and estimates the likelihood of PTSD one month later. The analysis drew on 145 patients enrolled at two New York City level-1 trauma centers, all within 24 hours of their injuries.

What makes the approach unusual is where it listens. Instead of adding a questionnaire to an already crowded intake, PREVIA analyzes the spontaneous response to one routine care question: what brought the patient to the emergency department. That answer is part of ordinary practice, recorded and transcribed anyway, so the method asks nothing new of clinicians. The question is whether a few hundred words of free speech carry enough signal to act on.

Key Facts

Nature Communications reported on November 9 that PREVIA reached 95.4 percent sensitivity, with a standard deviation of 3.99 percentage points, in cross-validation across three folds. The same runs produced an AUROC of 78.49 percent, plus or minus 5.75 points, and precision of 46.76 percent, plus or minus 8.66 points. The system caught nearly every patient who later met criteria for PTSD, while about half of the patients it flagged did not go on to develop the disorder.

NYU Langone Health, in a release carried by PR Newswire on November 9, reported that a head-to-head test on 30 randomly selected narratives, nine of them from patients who later developed PTSD, favored the model on recall. PREVIA flagged 100 percent of the cases that went on to develop PTSD, while an emergency clinician specializing in PTSD identified 89 percent, listed as 88.89 percent in the paper. The clinician had more than 24 years of emergency medicine experience.

The pipeline is a chain of open-source models rather than a single trained classifier. Llama is prompted to emphasize PTSD risk signs, Mistral to highlight protective and contextual factors, and DeepSeek reconciles their eight separate assessments into structured summaries. The two models' scores are averaged with equal weighting, and a final assessor combines that summary with age, sex, trauma type and known PTSD prevalence to output a 0 to 100 risk score with a short plain-language explanation. Unite.AI reported on November 9 that the system uses zero-shot prompts, with no parameter training or fine-tuning on the study sample.

Participants were English-speaking, medically stable adults aged 18 to 80, enrolled within 24 hours of trauma and given up to three minutes to describe the event while an iPad recorded video. Narratives averaged about 374 words, and Whisper transcribed the speech before native English-speaking research staff reviewed the transcripts. Diagnoses at one month were established with the Clinician-Administered PTSD Scale for DSM-5, the gold-standard clinical interview. The work was funded by National Institute for Mental Health grant R01 MH129856 and ran on NVIDIA A100 GPUs.

News-Medical.net reported on November 9 that the paper appears in Nature Communications volume 17 as article number 11118, credited to Mueller V. and colleagues. A secondary result covered subthreshold PTSD, where PREVIA reached 89.7 percent sensitivity and a 69.4 percent F1 score. A public repository provides the pipeline and prompts, which the authors say lets other hospitals test the approach.

Analysis

The 95.4 percent sensitivity figure reads like a diagnosis, but it describes a screening property, not a clinical verdict. With precision at 46.76 percent, more than half of the patients PREVIA flags would not develop PTSD, an acceptable trade only if the follow-up is cheap, brief and humane. What this really means is that the tool works best as a triage filter that directs scarce clinical attention, not as a replacement for the CAPS-5 interview or for a trained clinician's judgement.

The head-to-head comparison is encouraging but thin. Thirty narratives and nine positive cases is a small stage on which to crown a model, and the clinician's 88.89 percent sensitivity amounts to a single missed case out of nine. One case swinging the other way would erase the gap. The authors are explicit that the result needs validation in larger, independent cohorts, and the paper's confidence intervals tell the same story.

The architecture carries both promise and risk. Building on open-source models with zero-shot prompts means hospitals can run the pipeline without retraining anything, a real advantage for reproducibility and for under-resourced settings. But performance then depends heavily on prompt wording and on the linguistic fit of the cohort. Participants were English-speaking, medically stable adults between 18 and 80, and medically unstable patients were excluded, removing a group that may carry the highest risk of later illness.

The bigger picture here is that the study turns a routine clinical artifact, the patient's own account of what happened, into a measurable risk signal without adding a minute to intake. That is a genuine shift in how emergency psychiatry might be delivered, because it treats narrative as data rather than as background color. The caveat is that narrative is also culture, language and context, and a model tested on a narrow population needs careful recalibration before it travels.

Why It Matters

At the scale of 43 million injury-related emergency visits a year in the United States, even a modest improvement in who gets followed up would reach a very large number of people. Only 28 percent of departments screen at all, and two-thirds of those rely on validated instruments that consume time. A tool that reads what patients already say, and explains its reasoning in plain language, could fit workflows that have no slack for anything longer.

The false-positive burden is the part health systems will feel first. A positive screen would typically trigger a brief assessment rather than an invasive intervention, so the main costs are staff time and patient anxiety rather than unnecessary treatment. Still, in departments running on 15 to 20 minutes per patient, even a short extra conversation competes with something else. Whether PREVIA saves effort or shifts it depends on how the follow-up pathway is designed.

Explainability is the quieter contribution. A final layer produces a plain-language rationale alongside the 0 to 100 score, and the prompts are public, so clinicians and outside researchers can inspect what the models weighted. That transparency matters in a field where opaque risk scores have often been adopted first and audited later, especially when the input is a patient's spontaneous account of a traumatic event.

Next Up

The immediate next step, as the authors frame it, is validation in larger and independent cohorts, including populations that differ in language, insurance status and geography. Prospective deployment would also need to show that flagging patients changes outcomes, not just that the model predicts them. The public repository lowers the barrier for other groups to attempt replication.

Beyond replication, the open questions are regulatory and operational. A risk score that informs care is not a diagnosis, and hospitals will need clear rules for consent, data retention and clinician override. The study opens a path; whether emergency departments walk it depends on the next cohort, not on the first headline.