Optimizing Large Language Models for Improved Detection of Depression and Anxiety

This study, approved by the Human Research Ethics Committee of the University of Hong Kong (EA240276), developed a generative pipeline to transform case descriptions into clinical interviews. Licensed psychiatrists and psychologists evaluated the quality of these generated interviews, leading to the training of EmoScan, a system with two agents designed for emotional disorder screening and brief clinical interviews. The methodology involved a four-stage data-generative pipeline, synthesizing 1,157 cases from diverse sources into refined dialogues. Quality checks by experts indicated high standards in alignment, consistency, and logical coherence. EmoScan’s performance was compared with established LLMs like GPT-4 and Mistral-7B, using statistical analyses for screening effectiveness and interviewing performance. Metrics such as the weighted F1-score, BERTScore, and Chi-Square tests validated EmoScan’s robustness in screening and interview tasks. Comprehensive documentation and data access ensure replicability of the research findings, supporting advancements in mental health assessment using AI.

Source link