This page was translated automatically from Turkish. Read the original

This page is an AI-assisted Turkish summary of the study presented at the 26th World Congress of Psychiatry. The study belongs to the presenting authors; please refer to the original source when citing it.

EV0305

Predicting PTSD Risk from Trauma Narratives

PREDICTING PTSD-RISK FROM TRAUMA NARRATIVES

Trauma narratives analyzed using natural language processing distinguished individuals at high risk of PTSD significantly better than chance (pooled AUC 0.734), whereas neutral and negative narratives and subjective ratings were not predictive.

The study aimed to examine, using natural language processing (NLP), whether brief narratives of participants’ most traumatic life events could distinguish individuals at high risk of posttraumatic stress disorder (PTSD) in a community sample. Of a total of 169 individuals, 150 were assessed using the PCL-5, DASS-21, LEC-5, SWE, and verbal fluency measures; a PCL-5 cutoff score of 34 was used to define a high-PTSD group (N = 41) and a low-PTSD group (N = 109). Participants produced written autobiographical narratives of their most traumatic, negative but nontraumatic, and neutral life events; the texts were represented using sentence embeddings (deepset/gbert large, 1,024 dimensions) and LIWC word categories, and a linear support vector machine (SVM), logistic regression, and gradient-boosted trees were compared using nested cross-validation. Trauma narratives significantly predicted PTSD symptoms, with a pooled AUC of 0.734, a 95% confidence interval of 0.646 to 0.819, SVM as the best-performing model, and a permutation p-value < 0.002; this performance exceeded that of negative narratives (Δ = 0.159), neutral narratives (Δ = 0.233), and the word-frequency-based TF-IDF baseline model (Δ = 0.152), and remained stable across different PCL-5 cutoffs (AUC 0.680 to 0.734). In contrast, trauma narratives did not predict DASS-21 symptoms (all p ≥ 0.06), despite strong correlations between PCL-5 and DASS-21 scores (ρ = 0.64 to 0.75); none of the DSM-5 symptom clusters was a significant predictor on its own, and negative narratives (p = 0.198), neutral narratives (p = 0.429), and participants’ subjective ratings of their own narratives (p = 0.102) did not provide significant prediction. According to the authors, informative linguistic signals are linked to the construction of trauma memories rather than to general autobiographical language, and the lower accuracy among high-risk individuals may be attributable to overlapping linguistic patterns and the limited sample size.

All WCP 2026 summaries