Microsoft Develops AI Tool That Diagnoses Diseases Step-by-Step, Much Like a Physician

 Microsoft Develops AI Tool That Diagnoses Diseases Step-by-Step, Much Like a Physician

An artificial intelligence (AI) tool developed by researchers at Microsoft has demonstrated more than twice the accuracy of a typical physician in diagnosing complex medical conditions. A study detailing this AI tool—which analyzes a patient's condition and reaches a diagnosis through a step-by-step process, much like a human doctor—was published last Thursday in the journal Science.

According to the study, the AI ​​model identifies the questions to ask a patient, the physical examinations to conduct, and the necessary laboratory tests, while systematically guiding the diagnostic process. It utilizes a system called an "orchestrator," which coordinates how the diagnostic procedure unfolds. This system facilitates collaboration among a simulated group of AI agents—each acting as a different medical specialist—to diagnose complex health issues.

Most medical AI models currently in use diagnose diseases by analyzing all of a patient's symptoms, laboratory test results, and imaging reports simultaneously.

In contrast, this new AI model—coordinated by the "orchestrator"—conducts the diagnostic process in a step-by-step manner. It begins by determining what questions to ask regarding the patient's medical history and identifying which findings from a physical examination are significant. Based on the information gathered, it selects the necessary tests, analyzes the results, and decides whether further testing is required. According to the scientists who led the research, the AI ​​model continuously updates its initial hypotheses and diagnostic plans regarding a disease throughout the process. In the study, the model was tasked with diagnosing over 300 extremely rare and complex medical conditions. Utilizing a "diagnostic orchestrator," the AI ​​model performed a comprehensive assessment of the patient's condition; it demonstrated more than twice the accuracy of typical physicians who rely on textbooks and the internet to identify diseases.

Even when tested using health scenarios drawn from a large database of patients admitted to emergency departments, the AI ​​tool outperformed human physicians.

Researchers noted that previous comparisons between such AI systems and human doctors often relied on multiple-choice questions. Dr. Dominic King, Corporate Vice President of Microsoft’s AI health division, stated, "AI has reached a level where it can achieve a 100% success rate on such questions. However, the real world doesn't work that way. Patients never come to you with a neatly prepared summary of their medical issues."

The researchers also found that when the "orchestrator" determined the testing and diagnostic pathway, the process incurred lower costs compared to when a human physician managed the same procedure. Harsh Nori, Principal AI Researcher at Microsoft AI Health, said, “We are not just measuring how accurately the system diagnoses diseases; we are also evaluating how judiciously and carefully it carries out the diagnostic process.”