NYU's PREVIA AI Flags PTSD Risk From ER Trauma Narratives

NYU Langone Health's PREVIA AI ensemble identified patients at risk of post-traumatic stress disorder (PTSD) one month after trauma with 95.4% sensitivity in a 145-patient study, according to a November 9, 2026 announcement and a paper published the same day in Nature Communications. The experimental screener analyzes emergency-department trauma narratives instead of adding a lengthy mental-health questionnaire to emergency intake.

The numbers cut both ways. PREVIA achieved 78.5% AUROC and 46.8% precision across three cross-validation folds. Sensitivity measures the share of later PTSD cases detected; precision measures the share of flagged patients who actually developed PTSD; AUROC summarizes how well the scores separate the two outcomes. At 46.8% precision, many positive screens would be false positives, adding follow-up staff time, patient burden and resource-allocation costs — although the next step would typically be a brief assessment rather than an invasive intervention.

The context is a screening gap. Only about 28% of US emergency departments routinely screen for PTSD, and only two-thirds of those use validated questionnaires, the study reports. Standard assessments compete with immediate medical needs when clinician time runs roughly 15–20 minutes per patient. The first month after trauma is a critical identification window because many patients are no longer in contact with the hospital by then.

Senior author Katharina Schultebraucks, co-director of computational psychiatry at NYU Grossman School of Medicine, said: "PREVIA can serve as an AI screener that identifies at-risk patients and explains why, without adding any time to intake in busy emergency rooms."

The claim of zero added intake time rests on patients recording a narrative themselves. Whether that workflow survives real emergency departments, beyond research coordinators and iPads, remains untested.

How does PREVIA build a risk score?

Research coordinators at two New York City level-1 trauma centers asked patients to describe the event that brought them to the emergency department. Participants were enrolled within 24 hours of trauma and given up to three minutes to respond while an iPad recorded video. Narratives averaged about 374 words. Whisper transcribed patient speech, and native English-speaking research staff reviewed the transcripts. One-month PTSD outcomes were determined using the Clinician-Administered PTSD Scale for DSM-5, a clinical diagnostic interview.

PREVIA divides each transcript across three layers:

• Llama and Mistral independently assess early post-traumatic stress symptoms, reactions during the traumatic event, risk factors and protective factors.
• DeepSeek reconciles their eight assessments into structured summaries, while the two models' numerical scores are averaged with equal weighting.
• A final assessor combines the summaries with age, sex, trauma type and PTSD prevalence information to produce a risk score from 0 to 100 and a short explanation.

The pipeline uses zero-shot prompts (no parameter training or fine-tuning on the study sample). Cross-validation was used to optimize and evaluate the classification threshold.

The design is a constraint as much as a feature. With no fine-tuning, performance depends entirely on the models' general reasoning about trauma narratives — which helps explain why the paper calls for validation in larger, independent cohorts.

How does it compare against simpler models and a specialist?

The paper reports broadly comparable overall performance between PREVIA and single zero-shot configurations of O3, Llama and Mistral on many metrics. PREVIA's clearest advantage is sensitivity. For subthreshold PTSD — clinically meaningful symptoms below the full diagnostic threshold — PREVIA reached 89.7% sensitivity and a 69.4% F1 score, a combined measure of precision and sensitivity.

A separate comparison used 30 randomly selected narratives, stratified to include nine PTSD cases. PREVIA recorded 100% sensitivity and 84.4% AUROC. One emergency clinician recorded 88.9% sensitivity and 83.9% AUROC; both recorded 50% precision. The clinician had more than 24 years of emergency-medicine experience and specialized PTSD research experience.

The single-rater design leaves open whether the comparison generalizes across experience levels and clinical settings. Beating one expert on 30 narratives is suggestive, not conclusive.

Who was studied — and who was left out?

Participants were English-speaking, medically stable adults aged 18–80. Women accounted for 31% of the sample. Medically unstable patients were excluded, potentially omitting some patients at highest risk of chronic PTSD.

The authors identify future work as:

• testing in other languages
• longer-term PTSD trajectories
• comorbidities such as depression
• integration with clinician judgment and emergency workflows

What stands between PREVIA and deployment?

The authors propose a clinical workflow that pairs a brief audio recording with demographic and trauma variables from the electronic health record, then returns a risk score and explanation without additional clinician input. The study did not establish that such screening improves patient outcomes.

PREVIA ran in a HIPAA-compliant high-performance-computing environment on NVIDIA A100 GPUs. Self-hosting supports control over patient data, but adoption depends on hospitals having suitable computing and governance infrastructure — a nontrivial requirement for a system designed around open-weight models that run locally.

The public PREVIA repository provides the complete pipeline and prompts, plus five fabricated narratives for an end-to-end demonstration. Identifiable clinical narratives and psychiatric assessments remain restricted for privacy and consent reasons. The demo can test operation but cannot reproduce the paper's quantitative results.

PREVIA's immediate contribution is an independently testable screening approach built from a routine emergency encounter and clinician-assessed outcomes. Its clinical value now hinges on independent validation: whether larger cohorts preserve its sensitivity, whether services can manage the false positives, and whether hospitals can incorporate recording, follow-up and secure computing without disrupting emergency care.