Performance Evaluation of Large Language Models in Multilingual Medical Multiple-Choice Questions: Mixed Methods Study

Abstract

Background: Artificial intelligence continues to transform health care, offering promising applications in clinical practice and medical education. While large language models (LLMs), as a form of generative artificial intelligence, have shown potential to match or surpass medical students in licensing examinations, their performance varies across languages. Recent studies highlight the complex influence and interdependency of factors such as language and model type on LLMs’ accuracy; yet, cross-language comparisons remain underexplored. Objective: This study evaluates the performance of LLMs in answering medical multiple-choice questions quantitatively and qualitatively across 3 languages (German, French, and Italian), aiming to uncover model capabilities in a multilingual medical education context. Methods: For this mixed methods study, 114 publicly accessible multiple-choice questions in German, French, and Italian from an online self-assessment tool were analyzed. A quantitative performance analysis of several LLMs developed by OpenAI, Meta AI, Anthropic, and DeepSeek was conducted to evaluate their performance on answering the questions in text-only format. For the comparative analysis, a variation of input question language (German, French, and Italian) and prompt language (English vs language-matched) was used. The 2 best-performing LLMs were then prompted to provide answer explanations for incorrectly answered questions. A subsequent qualitative analysis was conducted on these explanations to identify the reasons leading to the incorrect answers. Results: The performance of LLMs in answering medical multiple-choice questions varied by model and language, showing substantial differences in accuracy (between 64% and 87%). The effect of input question language was significant ( P <.01) with models performing best on German questions. Across the analyzed LLMs, prompting in English generally led to better performance in comparison to language-matched prompts, but the top-performing models exceptionally showed comparable results for language-matched prompts. Qualitative analysis revealed that answer explanations of the analyzed models (GPT4o and Claude-Sonnet-3.7) showed different reasoning errors. In several explanations, this occurred despite factual accuracy on the represented topic. Furthermore, this analysis revealed 3 questions to be insufficiently precise. Conclusions: Our results underline the potential of LLMs in answering medical examination questions and highlight the importance of careful consideration of model choice, prompt, and input languages, because of relevant performance variability across these factors. Analysis of answer explanations demonstrates a valuable use case of LLMs for improving examination question quality in medical education, if data security regulations permit their use. Human oversight of language-sensitive or clinically nuanced content remains essential to determine whether incorrect output stems from flaws in the questions themselves or from errors generated by the LLMs. There is a need for ongoing evaluation as well as transparent reporting to ensure reliable integration of LLMs into medical education contexts.