DSM-Guided Large Language Model Reasoning for Depression and PTSD Assessment From Psychiatric Interview Transcripts

Abstract

Large language models (LLMs) have shown remarkable generalization capabilities across diverse natural language processing tasks. Yet, their application in psychiatry remains limited due to the absence of clinically validated reasoning frameworks. To address this gap, we propose a DSM-5 (Diagnostic and Statistical Manual of Mental Disorders, Fifth Edition)-guided prompting framework for mental health assessment. We design a series of experiments that progressively embed clinical knowledge, ranging from direct classification of raw interview transcripts to DSM-guided summarization and few-shot rationale-augmented reasoning. In each setting, LLMs generate clinically meaningful intermediate outputs, which are subsequently evaluated in LLM zero-shot prediction and further used for fine-tuning language models such as BERT and RoBERTa. Multiple LLM-generated outputs per case are used for data augmentation, with subsets selected by either clinicians or the LLM. Experiments on the DAIC-WOZ and E-DAIC benchmarks show that DSM-guided summarization and rationale-augmented reasoning substantially improve transcript-based mental health classification performance. Crucially, clinician-selected outputs consistently yield stronger classification performance, whereas LLM-selected outputs remain less reliable. These results indicate that human validation is indispensable when applying LLMs in psychiatry, even though the process is inherently time-consuming. In that respect, LLMs promise to reduce the human resources required to process large volumes of interviews. They may provide practical value as initial tools for large-scale mental health screening, where scalability is essential and expert supervision can be selectively applied.