Testing the use of local large language models to extract trauma identification and contextualize posttraumatic stress symptoms from self-report

Abstract

Abstract Accurate contextualization of trauma is critical for assessing posttraumatic stress disorder (PTSD) in behavioral health contexts, yet standard self-reports often fail to link symptoms to specific index traumas. This study evaluated the feasibility of using a local, privacy-preserving Large Language Model (LLM) to extract trauma exposure types from free-text narratives and evaluate their association with PTSD symptoms for description and prediction. Participants ( N = 109) recruited online via Prolific as part of a larger study completed an extended Life Events Checklist (LEC) with up to three free-text trauma descriptions, and the PTSD Checklist for DSM-5 (PCL-5). A local LLM was prompted to extract trauma types, which were compared with trauma categorizations generated by two expert clinical psychologists. Results indicated variable agreement; the LLM demonstrated high specificity (> 85% for most trauma types) and strong agreement for more frequently reported experiences, such as sexual assault (κ = 0.68–0.74). Cluster analysis based on LLM-derived features revealed significant differences in total PCL-5 scores ( p = .04) across clusters. Findings suggest local LLMs can effectively extract clinically relevant features from brief trauma descriptions. While refinement is needed, this approach offers preliminary evidence for a scalable and secure method to augment PTSD screening in behavioral medicine.