Identifying Psychiatric Manifestations in Outpatients with Depression and Anxiety: A Large Language Model-Based Approach

作者：Shihao Xu, Yiming Yan, Yanli Ding, Li Feng, Shu Zhang, Haoyun Tang, Chao Bao Luo, Yan Li, Hao Liu, Yu Mei, Wen Gu, Hong Qiu, Yong Wang, Jianyin Qiu, Tao Yang, Z. Wang, Qing Zhang, Haiyang Geng, Yunyun Han, Shao Jun, Nils Opel, Lidong Bing, Min Zhao, Yifeng Xu, Xun Jiang, Jianhua Chen · 发表于：medRxiv · 年份：2025 · DOI：10.1101/2025.01.03.24318117 · 被引用次数：4 · 研究领域：Mental Health via Writing、Machine Learning in Healthcare

Abstract Purpose Accurate psychiatric diagnosis and assessment are crucial for effective treatment. However, while current data-driven approaches emphasize diagnostic outcomes, the process of decoding the underlying symptom expressions in patients’ language and mapping them to well-defined psychiatric terminology has received relatively little attention. This study investigates the potential of Large Language Models (LLMs) to automate the identification of diagnostic categories and symptoms from psychiatrist-patient dialogues, to provide interpretable insights and support automatic diagnosis. Methods We analyzed audio recordings from 1160 psychiatric diagnostic interviews, primarily involving patients with depressive disorder and anxiety disorder. A clinical entities corpus was formed by leveraging clinical annotations in EMRs (e.g., chief complaints, mental status, elements in assessment scales) and widely used assessment scales. LLMs were 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. Results The system achieved 86.9% accuracy for identifying the appearance of clinical annotations and 74.7% (77.2%) accuracy for identifying anxiety (depression) symptoms. Patients with depression and anxiety, diagnosed using ICD-10 codes, were differentiated with an accuracy of 75.5%. Analysis of LLM-generated features shows that depression cases exhibited...