Performance of Open-Source Large Language Models in Psychiatry: Usability Study Through Comparative Analysis of Non-English Records and English Translations

Résumé fourni par la source

Background: Large language models (LLMs) have emerged as promising tools for addressing global disparities in mental health care. However, cloud-based proprietary models raise concerns about data privacy and limited adaptability to local health care systems. In contrast, open-source LLMs offer several advantages, including enhanced data security, the ability to operate offline in resource-limited settings, and greater adaptability to non-English clinical environments. Nevertheless, their performance in psychiatric applications involving non-English language inputs remains largely unexplored. Objective: This study aimed to systematically evaluate the clinical reasoning capabilities and diagnostic accuracy of a locally deployable open-source LLM in both Korean and English psychiatric contexts. Methods: The openbuddy-mistral-7b-v13.1 model, fine-tuned from Mistral 7B to enable conversational capabilities in Korean, was selected. A total of 200 deidentified psychiatric interview notes, documented during initial assessments of emergency department patients, were randomly selected from the electronic medical records of a tertiary hospital in South Korea. The dataset included 50 cases each of schizophrenia, bipolar disorder, depressive disorder, and anxiety disorder. The model translated the Korean notes into English and was prompted to extract 5 clinically meaningful diagnostic clues and generate the 2 most likely diagnoses using both the original Korean and translated English inputs. The hallucination rate and clinical relevance of the generated clues were manually evaluated. Top-1 and top-2 diagnostic accuracy were assessed by comparing the model's prediction with the ground truth labels. Additionally, the model's performance on a structured diagnostic task was evaluated using the psychiatry section of the Korean Medical Licensing Examination and its English-translated version. Results: The model generated 997 clues from Korean interview notes and 1003 clues from English-translated notes. Hallucinations were more frequent with Korean input (n=301, 30.2%) than with English (n=134, 13.4%). Diagnostic relevance was also higher in English (n=429, 42.8%) compared to Korean (n=341, 34.2%). The model showed significantly higher top-1 diagnostic accuracy with English input (74.5% vs 59%; P<.001), while top-2 accuracy was comparable (89.5% vs 90%; P=.56). Across 115 questions from the medical licensing examination, the model performed better in English (n=53, 46.1%) than in Korean (n=37, 32.2%), with superior results in 7 of 11 diagnostic categories. Conclusions: This study provides an in-depth evaluation of an open-source LLM in multilingual psychiatric settings. The model's performance varied notably by language, with English input consistently outperforming Korean. These findings highlight the importance of assessing LLMs in diverse linguistic and clinical contexts. To ensure equitable mental health artificial intelligence, further development of high-quality psychiatric datasets in underrepresented languages and culturally adapted training strategies will be essential.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.