Performance of GPT-3.5 and GPT-4 on standardized urology knowledge assessment items in the United States: a descriptive study

Abstract

PURPOSE: This study aimed to evaluate the performance of Chat Generative Pre-Trained Transformer (ChatGPT) with respect to standardized urology multiple-choice items in the United States. METHODS: In total, 700 multiple-choice urology board exam-style items were submitted to GPT-3.5 and GPT-4, and responses were recorded. Items were categorized based on topic and question complexity (recall, interpretation, and problem-solving). The accuracy of GPT-3.5 and GPT-4 was compared across item types in February 2024. RESULTS: GPT-4 answered 44.4% of items correctly compared to 30.9% for GPT-3.5 (P>0.0001). GPT-4 (vs. GPT-3.5) had higher accuracy with urologic oncology (43.8% vs. 33.9%, P=0.03), sexual medicine (44.3% vs. 27.8%, P=0.046), and pediatric urology (47.1% vs. 27.1%, P=0.012) items. Endourology (38.0% vs. 25.7%, P=0.15), reconstruction and trauma (29.0% vs. 21.0%, P=0.41), and neurourology (49.0% vs. 33.3%, P=0.11) items did not show significant differences in performance across versions. GPT-4 also outperformed GPT-3.5 with respect to recall (45.9% vs. 27.4%, P<0.00001), interpretation (45.6% vs. 31.5%, P=0.0005), and problem-solving (41.8% vs. 34.5%, P=0.56) type items. This difference was not significant for the higher-complexity items. Conclusion: s: ChatGPT performs relatively poorly on standardized multiple-choice urology board exam-style items, with GPT-4 outperforming GPT-3.5. The accuracy was below the proposed minimum passing standards for the American Board of Urology's Continuing Urologic Certification knowledge reinforcement activity (60%). As artificial intelligence progresses in complexity, ChatGPT may become more capable and accurate with respect to board examination items. For now, its responses should be scrutinized.

Ähnliche Arbeiten