AI Tool PREVIA Predicts Future PTSD Risk From Emergency Department Narratives With 95% Accuracy

AI Tool PREVIA Achieves Expert-Level Accuracy in PTSD Risk Prediction

Researchers at NYU Langone Health have developed PREVIA, an LLM ensemble that can predict future PTSD risk by analyzing unstructured patient narratives collected during emergency department intake. Published in Nature Communications on November 9, the study demonstrates that this AI tool PREVIA correctly identified more than 95.4% of at-risk patients from a cohort of 145 trauma survivors1

1

. In a direct comparison, PREVIA flagged 100% of 30 cases that developed post-trauma PTSD risk within one month, while an emergency department clinician specializing in PTSD correctly identified only 89% of these cases2

2

.

The tool analyzes patients' spontaneous responses to a single routine-care question: "What brought you to the ED?" These emergency department narratives, recorded and transcribed at bedside, contained an average of 373.68 words and ranged from 51 to 1,622 words1

1

. Senior study author Katharina Schultebraucks, PhD, associate professor in the departments of psychiatry and population health at NYU Langone Health and co-director of computational psychiatry at NYU Grossman School of Medicine, emphasized that this represents the first demonstration that a patient's spontaneous answer carries enough early psychological signals to forecast PTSD risk a month later3

3

.

Addressing Critical Gaps in Emergency Department Screening

Approximately 43 million patients with potential trauma exposure are admitted to emergency departments in the US annually. Among these trauma survivors, about 27% will develop post-traumatic stress disorder by one month post-trauma1

1

. The first month following trauma represents a critical window for early intervention, yet many patients lose contact with the healthcare system during this period. The emergency department often serves as the sole point of contact before symptomatology becomes chronic and challenging to treat2

2

.

Despite the clinical importance of PTSD screening, only about 28% of emergency departments routinely screen for the condition, and only two-thirds of these facilities use validated questionnaires3

3

. This low screening rate reflects overcrowding and limited clinician time, typically 15 to 20 minutes per patient, forcing clinicians to prioritize immediate medical concerns over lengthy mental health assessments. Existing validated tools like the Peritraumatic Dissociative Experiences Questionnaire and the Posttraumatic Adjustment Scale, while effective, add burden to already strained emergency department workflows1

1

.

How the LLM Ensemble Architecture Works

PREVIA employs a multi-layered pipeline that leverages different generative AI models based on their complementary strengths. The Llama model tends to emphasize signs of PTSD risk, while Mistral highlights protective and contextual factors that could buffer against it2

2

. This approach addresses limitations of single-prompt and single-model configurations, which struggle with generalizability across diverse clinical settings1

1

.

The first layer scans unstructured patient narratives for peritraumatic symptoms occurring during initial trauma, such as loss of control, amnesia, and dissociation, as well as early symptoms manifesting in the emergency department like feeling disconnected from the body. The second layer evaluates risk and protective factors, weighing preexisting mental health problems and lower socioeconomic status against coping skills and strong social support. A third AI model synthesizes these assessments into a structured summary, preventing any single model from dominating results. The final layer combines this synthesis with demographic information, trauma type, and known PTSD prevalence rates to generate explainable risk scores with plain-language explanations3

3

.

Performance Metrics and Clinical Validation

The AI tool PREVIA demonstrated strong screening performance with an AUROC of 78.49 ± 5.75 and high sensitivity of 95.4% ± 3.99% for identifying patients with future PTSD risk1

1

. Experiments confirmed that this three-layer design consistently outperformed simpler configurations across most performance metrics. Results were validated against diagnoses made using the Clinician-Administered PTSD Scale for DSM-5 (CAPS-5), the gold standard diagnostic interview for PTSD2

2

.

Importantly, PREVIA also forecasts subthreshold PTSD risk, which is highly prevalent, clinically relevant, and associated with increased risk of later PTSD diagnosis. The framework's ability to identify both full and subthreshold cases demonstrates robust predictive validity across the spectrum of post-trauma responses1

1

.

Implications for Scalable Implementation

PREVIA was built on open-source models requiring no training or fine-tuning, enabling other hospitals to test it independently without access to proprietary software or patient data from the original study2

2

. This design choice addresses practical barriers to adoption in diverse healthcare settings. The tool leverages information already embedded within routine emergency department care, adding no additional burden to personnel or patients during intake1

1

.

By identifying at-risk patients at their first healthcare encounter, PREVIA could open a window for timely care approaches when preventive intervention is most effective. The system's interpretability and grounding in established PTSD risk factors distinguishes it from black-box AI approaches, making clinical decision-making more transparent. As emergency departments face mounting pressure to address mental health needs within constrained resources, tools like PREVIA represent a pathway toward scalable, evidence-based screening that could alter post-trauma symptom trajectories before they become entrenched3

3

.