Harvard Study: AI Diagnosis Accuracy Beats ER Doctors in Triage

A Harvard Medical School study published in *Science* reports that AI diagnosis accuracy can outperform emergency room doctors in specific clinical scenarios, especially during early ER triage where information is limited. Researchers compared OpenAI large language models—o1 and 4o—against two attending physicians. In a test set of 76 real ER cases at Beth Israel Deaconess Medical Center, diagnoses produced by OpenAI’s o1 and 4o were later re-evaluated by two other physicians who did not know whether the source was human or AI. The AI diagnosis accuracy results were strongest at triage. The o1 model matched the exact or very close diagnosis 67% of the time, versus 55% and 50% for the two attending physicians, representing an estimated 12–17 percentage point improvement. The study did not pre-process the data; models received the same text information available in electronic medical records at the time of each diagnosis. The authors stress this is not an endorsement for AI to replace clinicians in life-or-death decisions. They call for prospective trials and note key limitations, including that the research used only text-based inputs and that there is currently no formal accountability framework for AI diagnosis errors. Lead author Arjun Manrai and co-lead author Adam Rodman highlight both promise and risk: AI could function as decision-support to reduce diagnostic errors, but integration into clinical workflows must preserve human oversight and trust. For traders, this is a healthcare/AI milestone rather than a direct crypto catalyst.