Performance Evaluation of 18 Generative AI Models (ChatGPT, Gemini, Claude, and Perplexity) in 2024 Japanese Pharmacist Licensing Examination: Comparative Study

要旨

Background: Generative artificial intelligence (AI) has shown rapid advancements and increasing applications in various domains, including health care. Previous studies have evaluated AI performance on medical license examinations, primarily focusing on ChatGPT. However, the availability of new online chat-based large language models (OC-LLMs) and their potential utility in pharmacy licensing examinations remain underexplored. Considering that pharmacists require a broad range of expertise in physics, chemistry, biology, and pharmacology, verifying the knowledge base and problem-solving abilities of these new models in Japanese pharmacy examinations is necessary. Objective: This study aimed to assess the performance of 18 OC-LLMs released in 2024 in the 107th Japanese National License Examination for Pharmacists (JNLEP). Specifically, the study compared their accuracy and identified areas of improvement relative to earlier models. Methods: The 107th JNLEP, comprising 345 questions in Japanese, was used as a benchmark. Each OC-LLM was prompted by the original text-based questions, and images were uploaded where permitted. No additional prompt engineering or English translation was performed. For questions that included diagrams or chemical structures, the models incapable of image input were considered incorrect. The model outputs were compared with publicly available correct answers. The overall accuracy rates were calculated based on subject area (pharmacology and chemistry) and question type (text-only, diagram-based, calculation, and chemical structure). Fleiss' κ was used to measure answer consistency among the top-performing models. Results: Four flagship models-ChatGPT o1, Gemini 2.0 Flash, Claude 3.5 Sonnet (new), and Perplexity Pro-achieved 80% accuracy, surpassing the official passing threshold and average examinee score. A significant improvement in the overall accuracy was observed between the early and the latest 2024 models. Marked improvements were noted in text-only and diagram-based questions compared with those of earlier versions. However, the accuracy of chemistry-related and chemical structure questions remains relatively low. Fleiss' κ among the 4 flagship models was 0.334, which suggests moderate consistency but highlights variability in more complex questions. Conclusions: OC-LLMs have substantially improved their capacity to handle Japanese pharmacists' examination content, with several newer models achieving accuracy rates of >80%. Despite these advancements, even the best-performing models exhibit an error rate exceeding 10%, underscoring the ongoing need for careful human oversight in clinical settings. Overall, the 107th JNLEP will serve as a valuable benchmark for current and future generative AI evaluations in pharmacy licensing examinations.

主題