Microsoft AI coordinator outperforms doctors in complex diagnosis tests

A new artificial intelligence tool developed by Microsoft’s health division has demonstrated diagnostic capabilities that significantly surpass those of human specialists in specific testing scenarios. The system, referred to as a "coordinator," operates by simulating the collaborative workflow of a medical team rather than analyzing all symptoms and test results simultaneously, which is the standard approach for most current medical AI models.

Mechanism of the coordinator system

The operational process begins with the AI identifying critical questions regarding a patient's medical history and determining which aspects require focused attention during clinical examination. Following this initial assessment, the system selects necessary laboratory tests and imaging studies. Based on the resulting data, the coordinator re-evaluates its working hypotheses and decides whether additional testing is required to reach a definitive diagnosis.

This iterative method aims to mirror real-world medical practice, where information is gathered progressively and diagnoses are updated as new evidence emerges. Researchers noted that traditional benchmarks comparing AI performance against doctors often rely on multiple-choice questions, a format where AI models have already achieved near-perfect scores. Dominic King, Vice President of Health at Microsoft AI, stated that while AI excels at these standardized tests, they do not reflect clinical reality.

"No patient has ever walked into my office with an organized summary of their condition," King said, highlighting the disconnect between academic testing and actual patient care. He added that AI has become extremely proficient at achieving 100 percent success rates on such simplified queries.

Diagnostic accuracy and cost efficiency

Testing involving more than 300 complex and unusual medical cases revealed that the AI tool achieved an accuracy rate exceeding twice that of general practitioners who utilized medical references and internet searches. In separate evaluations based on cases recorded in a large database of emergency department patients, the system also outperformed physicians.

Beyond accuracy metrics, the study indicated that the AI-driven approach reduced diagnostic costs compared to the pathways established by the participating doctors. These findings suggest promising potential for using artificial intelligence to support medical decision-making processes.

However, researchers cautioned that superior performance in controlled research environments does not equate to proven dominance over doctors in daily clinical practice. The comparisons remain limited to the specific parameters of the conducted tests.