Multi-context linguistic analysis as an objective diagnostic adjunct for PTSD in earthquake survivors

Background

Post-traumatic stress disorder (PTSD) profoundly disrupts narrative organisation, emotional regulation, and communicative fluency. Despite its high prevalence following large-scale disasters, PTSD remains under-detected due to the resource demands of structured clinical evaluation[1]. Computational linguistics offers a promising avenue for objective, scalable, clinician-independent screening [1],[2],[3]. The present study investigated whether automatically extracted speech features reliably differentiate PTSD patients from healthy controls and whether these features independently predict diagnosis beyond demographic and clinical variables.

Methods

311 adults were recruited following the February 2023 Kahramanmaraş earthquakes (PTSD= 154; controls= 157). PTSD diagnoses were confirmed via the Clinician-Administered PTSD Scale for DSM-5 (CAPS-5). Speech samples were elicited across three ecologically distinct contexts: spontaneous daily conversation, structured Thematic Apperception Test (TAT) narratives, and CAPS-5 questionnaire. Speech recordings were transcribed using OpenAI Whisper large-v2, a transformer-based automatic speech recognition (ASR) system with demonstrated robustness in Turkish. Transcripts were processed with BERTurk, a BERT-based language model pre-trained on large-scale Turkish corpora, to compute contextual embeddings used in within-discourse semantic similarity estimation[3]. Fourteen linguistic variables were extracted per participant using automated NLP pipelines, including type-token ratio (TTR), filler-word rate, rule-based grammatical error count, mean perplexity (a proxy for syntactic unpredictability), high-perplexity sentence count, within-discourse semantic similarity, noun/verb/adjective/adverb ratios, and pause-rate metrics. Psychological assessments included State-Trait Anxiety Inventory (STAI), and Beck Depression Inventory (BDI). Group comparisons were conducted using independent-samples t-tests or Mann-Whitney U tests depending on distributional properties. Binary logistic regression with hierarchical entry and ROC analysis evaluated the diagnostic accuracy.

Results

PTSD patients generated substantially longer utterances but exhibited considerably lower lexical diversity (TTR: 0.407 vs. 0.550, p<.001), higher syntactic unpredictability (perplexity: 563 vs. 149, p<.001), greater grammatical error density (143.5 vs. 29.0, p<.001), and lower within-discourse semantic coherence (p <.001). In spontaneous speech, PTSD patients paused more frequently (p =.007), showed slower speech rate (p =.040), and used more filler words (p = .001). Adjective usage was significantly lower in the PTSD group (p <.001), whereas verb ratio (p <.001) and filler word rate (p <.001) were significantly elevated. A three-variable logistic regression model (TTR, semantic similarity, onset rate) achieved 90.4% classification accuracy. Sensitivity was 90.9% and specificity 91.1% at the optimal Youden threshold. After controlling for age, education, and BDI in a hierarchical model, TTR (OR = 0.028, p <.001) and semantic similarity (OR = 0.065, p <.001) remained as robust independent predictors (final model accuracy 91.3%, AUC = .964). Split-validation confirmed stability with negligible performance drop (test-set AUC = .946, sensitivity 97.6%, specificity 83.0%).

Conclusions

Automated speech analysis yields highly accurate, clinician-independent PTSD markers in earthquake survivors. Reduced lexical diversity and disrupted semantic coherence robustly predict PTSD status even after controlling for depression and demographic confounders. Consistent with prior meta-analytic evidence linking word count and negative-affect language to PTSD severity[2] , our findings suggest disorder-specific alterations in discourse organisation rather than general distress. These results support NLP-based tools as scalable, cost-effective adjuncts to conventional PTSD assessment. This research was funded by The Scientific and Technological Research Council of Turkey, grant number 323S218.

References

1. Quillivic, R., Gayraud, F., Auxéméry, Y. et al. Interdisciplinary approach to identify language markers for post-traumatic stress disorder using machine learning and deep learning. Sci Rep 14, 12468 (2024).
2. Yu, Z., Gu, Z., Shen, Y., & Lu, J. (2025). The relationship between language features and PTSD symptoms: a systematic review and meta-analysis. Frontiers in Psychiatry, 16, 1476978.
3. Schweter, S. (2020). BERTurk — BERT models for Turkish (v1.0.0) [Software]. Zenodo. https://doi.org/10.5281/zenodo.3770924