An ensemble of language models read 145 trauma narratives recorded in two New York emergency departments and flagged nearly everyone who later met criteria for PTSD. Fewer than half the people it flagged did.

The limiting factor in early post-trauma care is not treatment, it is contact. By the authors' accounting, about 27% of patients with trauma exposure seen in American emergency departments meet criteria for post-traumatic stress disorder a month later, and the emergency visit is often the last time the health system sees them before symptoms settle in. Short bedside questionnaires already carry enough prognostic signal to identify much of that risk. They are mostly not used: about 28% of emergency departments screen routinely, and only two-thirds of those use a validated instrument.
So a team at NYU Grossman School of Medicine removed the questionnaire. Patients at two New York City level-1 trauma centres were asked one open-ended question within 24 hours of injury, "Tell me about the event that brought you to the ED", and given three minutes to answer. The transcribed answer, plus age, sex and trauma type, went to an ensemble of three open-weight language models. Across 145 patients, the system identified 95.4% of those who went on to receive a clinician-administered diagnosis one month later, with an area under the curve of 78.49%. Its precision was 46.76%, so fewer than half the people it flagged turned out to be cases.
Why it matters: The first month after a trauma is when prevention is most tractable and when contact with the patient is most easily lost. A screener that runs off a question already asked at triage adds no bedside time, which is the constraint that keeps validated questionnaires out of use.
Victoria Mueller, Katharina Schultebraucks and colleagues in the Department of Psychiatry at NYU Grossman School of Medicine report the system, which they call PREVIA, in Nature Communications (Mueller et al.). The patients were enrolled in PREDICT, an ongoing NIH-funded longitudinal study, and their narratives averaged 373.68 words, ranging from 51 to 1622.
PREVIA is not one model answering one prompt. Its first layer passes each narrative to Llama-3.3-70B-Instruct and Mistral-Large-Instruct 2411 under four separate prompts, each targeting a dimension the PTSD literature already treats as prognostic: early post-traumatic symptoms, peri-traumatic symptoms such as dissociation and loss of control, risk factors, and protective factors. The second layer hands the resulting eight assessments to a third model, DeepSeek-R1-Distill Llama-70B, which summarises them, while the numeric score is an equal-weight average so neither of the first two models dominates. The third layer takes those summaries plus demographics and published one-month prevalence figures, and returns a risk score from 0 to 100 with a short written justification.
Nothing is trained. The models run zero-shot, and the only quantity fitted to this cohort is the cutoff that turns the score into a yes or no, optimised inside the training folds of a three-fold cross-validation and applied to the held-out fold. That restraint matters at 145 patients, where a fine-tuned classifier would have little room to avoid memorising its own sample.
Sensitivity of 95.4% against precision of 46.76% is not a flaw to be engineered away, it is the design. At a one-month prevalence near 30%, a screener this sensitive necessarily flags roughly twice as many people as will become cases. The authors argue that asymmetry is the right one here, since a false positive buys a brief follow-up assessment rather than an invasive intervention, while a false negative loses the patient altogether. They also state the cost plainly, in follow-up staff time, patient burden and resource allocation.
Two secondary numbers deserve more attention than the headline. The Brier score, which measures how well calibrated the probabilities are rather than how well they rank, was 27.71%, so these scores separate patients better than they estimate anyone's actual risk. And the ablations locate some of the signal outside the narrative: removing the prevalence information barely moved performance, while removing age, sex and trauma type produced the largest drop of any condition tested. Part of what the ensemble knows, it knows from the chart.
The comparison with an expert comes from 30 narratives, stratified to the cohort's prevalence, which yields nine actual cases. One emergency clinician with more than 24 years in emergency medicine and PTSD research read them blinded to outcomes, to PREVIA's predictions and to the cohort's prevalence. The clinician reached an AUROC of 83.86% and the ensemble 84.39%. On nine positives, a half-point gap is one borderline call. It also sets a reader deliberately denied prevalence information against a system whose final prompt is handed it. The authors note the single-clinician limitation and call for multi-clinician studies.
Nobody was treated. The work shows that risk can be forecast from a narrative, not that forecasting it changes any patient's trajectory. The cohort is 145 patients at one New York health system, English-speaking, and 31% women against 40% to 50% in comparable studies. Medically unstable patients were excluded at triage by design, which the authors concede may omit those at highest risk of chronic PTSD. The endpoint is a diagnosis at one month, not a longer trajectory or comorbid depression. And the system needs self-hosted large models on institutional computing hardware, which is what makes its HIPAA-compliant deployment possible and also what will keep it out of most emergency departments.
On the central trade, the authors write that In this context, a model that reliably identifies at-risk patients, even at the cost of lower precision, is clinically preferable to one that misses at-risk cases. That is defensible, and it is a clinical judgement rather than a result.
Does this diagnose PTSD? No. It estimates the risk of a diagnosis a month out, and the diagnosis itself still came from a clinician-administered interview.
Why three models instead of the best one? Splitting the task across models and prompts was meant to limit the influence of any one model's biases, and the full three-layer version beat every reduced version tested.
What is the takeaway? A spontaneous answer to a routine triage question carries enough signal to flag almost everyone who will develop PTSD within a month, at the price of flagging about as many who will not.
Mueller et al. "LLM ensemble PREVIA predicts future PTSD risk from emergency department trauma narratives with expert-level accuracy." Nature Communications, 2026. doi.org/10.1038/s41467-026-78768-3
PubMed PMID: 43003853.
Image: a hospital emergency department (Leopoldina-Krankenhaus). Michael Bemmerl, CC BY 3.0 DE, via Wikimedia Commons.
Weekly research updates, breakthrough summaries, and new articles — straight to your inbox. Free, always.
Comments