Harvard Study: AI Diagnoses ER Patients 45% More Accurately Than Veteran Doctors

A new study shows that an AI system can diagnose emergency room patients more accurately than human doctors. Harvard Medical School and Beth Israel Deaconess Medical Center ran the study. It was published in the journal Science on April 30, 2026.

The researchers tested an AI model called o1-preview. They used real emergency room cases from Beth Israel’s hospital. The AI looked at raw electronic health records, just like doctors do. Nothing was cleaned up or simplified beforehand.

The results were striking. The AI got 67.1% of diagnoses right across 76 real ER cases. Attending physicians scored 55.3% and 50.0% on those same cases. That’s a meaningful gap in a high-stakes setting like an emergency room.

Across 76 real ER cases, AI outdiagnosed human doctors — 67.1% accuracy versus 55.3% and 50.0%.

The AI was also tested on 143 medical cases from the New England Journal of Medicine. It included the correct diagnosis in 78.3% of those cases. When it listed multiple possible diagnoses, it gave a helpful answer 97.9% of the time. Blinded reviewers couldn’t tell whether the diagnoses came from a human or the AI.

The AI was graded at three stages of care: triage, diagnosis, and hospital admission. At each stage, it only used the information a doctor would’ve had at that point. That made the test more realistic. Notably, 72% of physicians believe AI already enhances diagnostic capabilities across medical settings.

Dr. Adam Rodman, one of the researchers, noted that the AI handled messy, real-world data well. Researcher Arjun Manrai said the AI is ready for more rigorous trials, similar to how new medical treatments are tested.

Still, the researchers were careful about what the results mean. They said the AI isn’t ready to replace doctors. It can’t practice medicine on its own. And they stressed that controlled clinical trials are needed before the AI could ever be used in real patient care.

The study’s findings don’t cover every possible outcome either. Real-world results can be more complicated than what shows up in benchmarks. Experts see AI as a tool for profound technological shifts in how medicine is practiced rather than a replacement for human clinicians.

Even so, the researchers believe AI could be a powerful tool to help doctors make better decisions faster, especially in busy emergency rooms where time matters most.