Detecting PTSD in Clinical Interviews: A Comparative Analysis of NLP Methods and Large Language Models

Post-Traumatic Stress Disorder (PTSD) remains under-detected in clinical settings, presenting opportunities for automated detection to identify at-risk patients. This study evaluates natural language processing approaches for binary PTSD classification from clinical interview transcripts using the DAIC-WOZ dataset, which contains semi-structured interviews with standardized psychological assessments. We compared embedding-based methods (SentenceBERT/LLaMA with logistic regression), general and mental health-specific transformer models (BERT/RoBERTa), and large language model prompting strategies (zero-shot/few-shot/chain-ofthought). SentenceBERT embeddings with logistic regression achieved the highest overall performance (AUPRC=0.758±0.128), outperforming domain-specific end-to-end fine-tuning models like Mental-RoBERTa (AUPRC=0.675±0.084 vs. RoBERTa-base 0.599±0.145). Few-shot prompting using DSM-5 criteria and two examples yielded competitive results (AUPRC=0.737). Performance varied significantly across symptom severity and comorbidity status with depression, with higher accuracy for severe PTSD cases and patients with comorbid depression. Our findings highlight the potential of embedding-based methods and LLMs for scalable screening while underscoring the need for improved detection of nuanced presentations.