Accurate psychiatric diagnosis and assessment are crucial for effective treatment. However, current diagnostic approaches heavily rely on subjective observations constrained by time and clinical resources. This study investigates the potential of using Large Language Models (LLMs) to identify the symptoms in psychiatrist-patient dialogues and use them as intermediate features to predict the diagnostic labels. We collected audio recordings of 1160 outpatients with depressive disorder and anxiety disorder. LLMs were trained and utilized to identify clinical symptoms, rate assessment scales, and an ensemble learning pipeline was designed to classify diagnostic results and symptoms with 10-fold cross-validation. The system achieved 86.9% accuracy for identifying the appearance of clinical annotations and 74.7% (77.2%) accuracy for identifying symptoms of anxiety (depression). In addition, analysis of LLM-generated features shows that depression cases exhibited prominent markers of anhedonia and decreased volition, whereas anxiety disorders were characterized by tension and an inability to relax.
Ding et al. (Tue,) studied this question.