AI predicts later PTSD risk from emergency-room narratives in 145-patient study

A patient’s account of what brought them to the emergency room can contain warning signs of later post-traumatic stress disorder. In a Nature Communications study published November 9, researchers used an AI pipeline called PREVIA to estimate that risk in 145 patients. It detected most future cases, but also flagged many patients who did not receive the diagnosis.

One ordinary question, three stages of analysis

PREVIA starts with answers to a routine question: “What brought you to the ED?” In the study, researchers collected bedside, video-recorded narratives within 24 hours of trauma at two New York City trauma centers. The accounts averaged about 374 words, though their lengths varied considerably.

Rather than asking one language model for an immediate prediction, the researchers divided the work into smaller clinical tasks. PREVIA combines several models in a sequence designed to examine evidence before producing a risk estimate:

1. Assess the narrative. Separate prompts examine early stress symptoms, reactions during the trauma, risk factors and protective factors. These can include distress as well as support that might buffer its effects.
2. Build a structured clinical summary. A second stage brings those assessments together, rather than passing along only one model’s judgment.
3. Estimate future risk. The final stage generates a probability of PTSD alongside a short explanation, making the stated basis for the prediction available for review.

The best-performing configuration used Llama and Mistral to assess the narrative, DeepSeek to assemble the summary, and Mistral for the final assessment. The authors found that PREVIA outperformed the single-model, direct-prompt approaches they tested, including O3, Llama and Mistral.

95.4%Sensitivity

Mean share of patients with a later PTSD diagnosis that PREVIA identified across three cross-validation folds.

46.76%Precision

Mean share of PREVIA’s positive predictions that corresponded to a later PTSD diagnosis.

Catching cases is not the same as confirming them

Researchers checked the predictions against assessments conducted one month after trauma using CAPS-5, a structured clinical interview for diagnosing PTSD. That comparison tested whether the early narratives predicted a later clinical outcome—not simply whether another AI agreed with PREVIA.

The distinction between sensitivity and precision is central to reading the results. PREVIA identified most patients who later had PTSD, but fewer than half of its positive predictions corresponded to that diagnosis. Its high detection rate therefore came with substantial false alarms within this study.

The paper also compared PREVIA with an emergency clinician specializing in PTSD on 30 randomly selected narratives from the same cohort. PREVIA detected all patients in that subset who later received a PTSD diagnosis; the clinician detected 88.89%. Both had 50% precision. This was a small comparison within the original group, not a separate hospital validation.

A screening opportunity before patients leave care

The proposed use is early screening, not replacing the diagnostic interview. The authors point to a practical gap: emergency clinicians must prioritize immediate medical needs, and many trauma survivors are no longer in contact with the hospital a month later. Drawing on prior research, the paper says only about 28% of emergency departments routinely screen for PTSD.

Using a question already asked during care is intended to avoid adding a separate screening questionnaire. But the study’s narratives were collected through recorded bedside interviews. Its results show a predictive signal in those accounts; they do not establish how the full process would perform across routine emergency-room workflows.

In NYU Langone’s account, published by News Medical, senior author Katharina Schultebraucks said PREVIA uses open-source models without training or fine-tuning. She presented that design as a way for other hospitals to test it without proprietary software or access to the team’s patient data. The study’s authors describe the findings as initial and call for larger, independent validation.