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

Area of Science:

• Computational Linguistics
• Clinical Psychology
• Artificial Intelligence

Background:

• Post-Traumatic Stress Disorder (PTSD) is frequently under-detected in clinical settings.
• Automated detection methods offer a potential solution for identifying at-risk individuals.
• Clinical interview transcripts are a rich source of data for PTSD assessment.

Purpose of the Study:

• To evaluate various natural language processing (NLP) approaches for binary PTSD classification.
• To compare embedding-based methods, transformer models, and large language model (LLM) prompting strategies.
• To assess the performance of these methods on the DAIC-WOZ dataset.

Main Methods:

• Utilized the DAIC-WOZ dataset containing semi-structured interviews and psychological assessments.
• Compared SentenceBERT/LLaMA embeddings with logistic regression.
• Evaluated general (BERT/RoBERTa) and mental health-specific transformer models.
• Assessed LLM prompting strategies including zero-shot, few-shot, and chain-of-thought.

Main Results:

• SentenceBERT embeddings with logistic regression achieved the highest performance (AUPRC=0.758±0.128).
• This method outperformed domain-specific models like Mental-RoBERTa (AUPRC=0.675±0.084).
• Few-shot LLM prompting also showed competitive results (AUPRC=0.737).
• Performance was higher for severe PTSD cases and those with comorbid depression.

Conclusions:

• Embedding-based NLP methods demonstrate significant potential for scalable PTSD screening.
• LLMs offer a viable approach for automated PTSD detection in clinical settings.
• Further research is needed to improve the detection of nuanced PTSD presentations and comorbidities.