One routine ER question may be enough: PREVIA's LLM ensemble forecasts PTSD risk a month after trauma
Synopsis
Emergency departments rarely have time for PTSD screening. PREVIA feeds the transcript of a patient's answer to the routine question "What brought you to the ED?" into a clinically informed ensemble of Llama, Mistral and DeepSeek models arranged in three layers: four clinical dimensions are assessed first, synthesized into a structured summary, then combined with age, sex and trauma type into a 0–100 risk score with a short explanation. Among 145 trauma survivors at two New York City level-1 trauma centers, the ensemble reached AUROC 78.49%±5.75% and sensitivity 95.4%±3.99% for one-month PTSD diagnosis; on 30 randomly selected narratives it scored AUROC 84.39% versus 83.86% for an emergency clinician specialized in PTSD.
Interpretation
A single answer to a routine emergency-department question can now be used to forecast whether a patient will develop post-traumatic stress disorder a month later, without extra questionnaires or added clinician time. Existing ED screening relies on instruments such as the PDEQ, a brief ISRC-derived screener and the PAS, which take additional time; about 28% of US emergency departments routinely screen for PTSD, and only two-thirds of those use validated questionnaires. Among 145 survivors enrolled within 24 hours of trauma at two New York City level-1 trauma centers, PREVIA reached AUROC 78.49%±5.75%, sensitivity 95.4%±3.99% and precision 46.76%±8.66% against one-month CAPS-5 diagnosis.
Prediction runs through three layers: two models read the same narrative along four clinical dimensions, a third model synthesizes the eight assessments into a structured summary, and a final layer combines that summary with demographics and population prevalence into a risk score and explanation. Single-prompt, single-model approaches vary substantially and generalize poorly across clinical settings, whereas decomposing the task into well-defined subtasks and combining outputs from several models reduces reliance on any one model. The best configuration used Llama and Mistral in Layer 1, DeepSeek in Layer 2 and Mistral in Layer 3; on 30 randomly selected narratives it reached AUROC 84.39%, sensitivity 100% and precision 50.00%, while one clinician with over 24 years of emergency medicine and PTSD research experience reached AUROC 83.86%, sensitivity 88.89% and precision 50.00%.
Perspective
The demonstrated scope is 145 adult survivors who spoke English, enrolled within 24 hours of trauma at two New York City level-1 trauma centers, with one-month CAPS-5 diagnosis as the outcome; the clinician comparison covers one expert and 30 randomly selected narratives.
The authors describe these as initial findings that warrant validation in larger, independent cohorts; the clinician comparison involves a single expert, and future multi-clinician studies across experience levels and settings are needed; the framework is currently limited to English-speaking populations, and self-hosting open-source models requires high-performance computing.