OpenAI's Reasoning Model Reaches Human Doctor Level in Clinical Diagnosis, Performs Better in Complex Scenarios

Recently, a research team from Harvard Medical School and Beth Israel Deaconess Medical Center published a study in the journal Science, testing an OpenAI reasoning model's performance in real clinical scenarios. The model's ability to diagnose patients and develop treatment plans under incomplete and unstructured information conditions matched or exceeded that of human doctors.

In one case, a patient with pulmonary embolism had a recurrence after treatment. Doctors suspected the medication was ineffective, but the AI suggested the patient might have a lupus history based on electronic health records, which was later confirmed correct. The study covered emergency triage to hospitalization stages, and the model outperformed two experienced doctors under the same information constraints.

However, the model only relies on text medical records and does not process non-verbal information like images, which is a limitation. The study emphasizes that AI is not a replacement for doctors but a clinical decision support tool, especially suitable for scenarios like emergency rooms with incomplete information, and is expected to reshape medical practice.