Researchers at the University at Buffalo have developed a new AI tool called SCAI (Semantic Clinical Artificial Intelligence) that outperforms most human physicians—and all prior AI models—on the U.S. Medical Licensing Examination (USMLE). By integrating formal medical knowledge from peer-reviewed sources, clinical guidelines, and pharmacological data, SCAI enables AI to reason like a physician.
In testing, SCAI-enhanced large language models scored up to 95.1% on Step 3 of the USMLE, surpassing GPT-4 Omni. The study, published in JAMA Network Open, suggests AI like SCAI could support physicians in clinical decision-making—not replace them, but significantly enhance their diagnostic capabilities.