11:25Machine-translated from Azerbaijani by AIRead in Azerbaijani
Microsoft’s Medical AI System Outperforms General Practitioners in Testing
Microsoft has developed a medical diagnostic system called MAI-DxO that coordinates several AI agents. According to a study published in the journal Science, the system produced more accurate results than general practitioners in difficult medical cases. Microsoft considers this an important step in the field of medical AI. However, the tool has so far been tested only in trials based on data from previous patients.
MAI-DxO works like a team of doctors made up of AI assistants that perform individual tasks. These assistants examine the patient’s complaints and medical history, ask follow-up questions, and select the necessary tests. They then compare possible diagnoses. The system does not receive all the information in advance, instead reaching a decision step by step. Microsoft plans to release SDBench, the test it developed to evaluate this approach, as open source. In a trial involving difficult medical cases, MAI-DxO achieved diagnostic accuracy of 83.2%. General practitioners using reference materials scored 36.1%, while those without these materials scored approximately 22%. In a separate trial using data from emergency departments, the system achieved 80.3% accuracy. The doctors’ average result was 46.8%. The system produced its best results with OpenAI’s GPT-5 and GPT-5.5 models. When operating alone, GPT-5 achieved 78.9% accuracy in difficult cases. The cost per case was calculated at approximately $8,400 (approximately 14,280 manat). MAI-DxO reduced that cost by approximately $3,000 (approximately 5,100 manat). Accuracy also increased by approximately 4 percentage points. The cost figures here do not represent actual payments but rather the estimated value of the tests and procedures selected during the trials. MAI-DxO running on GPT-5.5, released in April 2026, raised accuracy in difficult cases from 72% to approximately 87%. On emergency department data, the model achieved approximately 85% both on its own and within the system. With the Gemini 2.5 Flash model, the figure rose from 46.4% to approximately 69%. Specialist doctors and medical teams did not participate in the study. It is therefore emphasized that the difference could be smaller in real-world settings.
In the coming months, Microsoft will test the system in a real hospital setting with Beth Israel Deaconess Medical Center. Patients will receive standard medical care. MAI-DxO will also operate in parallel during the treatment of patients who give their consent. The system’s recommendations will then be compared with the doctors’ decisions. Before the tool can be used widely, issues involving clinical validation, legal liability, and compliance with medical regulations must be resolved. Microsoft considers this project part of its medical AI plan aimed at surpassing the most capable human-run healthcare systems. The company is developing a separate foundation AI model for healthcare with Mayo Clinic and plans to release it in 2027. The plan is not limited to diagnostics. Treatment decisions and the coordination of medical services are also among its key areas. According to Microsoft, doctors should retain the primary role in patient care, and the system should not replace them.
MAI-DxO works like a team of doctors made up of AI assistants that perform individual tasks. These assistants examine the patient’s complaints and medical history, ask follow-up questions, and select the necessary tests. They then compare possible diagnoses. The system does not receive all the information in advance, instead reaching a decision step by step. Microsoft plans to release SDBench, the test it developed to evaluate this approach, as open source. In a trial involving difficult medical cases, MAI-DxO achieved diagnostic accuracy of 83.2%. General practitioners using reference materials scored 36.1%, while those without these materials scored approximately 22%. In a separate trial using data from emergency departments, the system achieved 80.3% accuracy. The doctors’ average result was 46.8%. The system produced its best results with OpenAI’s GPT-5 and GPT-5.5 models. When operating alone, GPT-5 achieved 78.9% accuracy in difficult cases. The cost per case was calculated at approximately $8,400 (approximately 14,280 manat). MAI-DxO reduced that cost by approximately $3,000 (approximately 5,100 manat). Accuracy also increased by approximately 4 percentage points. The cost figures here do not represent actual payments but rather the estimated value of the tests and procedures selected during the trials. MAI-DxO running on GPT-5.5, released in April 2026, raised accuracy in difficult cases from 72% to approximately 87%. On emergency department data, the model achieved approximately 85% both on its own and within the system. With the Gemini 2.5 Flash model, the figure rose from 46.4% to approximately 69%. Specialist doctors and medical teams did not participate in the study. It is therefore emphasized that the difference could be smaller in real-world settings. In the coming months, Microsoft will test the system in a real hospital setting with Beth Israel Deaconess Medical Center. Patients will receive standard medical care. MAI-DxO will also operate in parallel during the treatment of patients who give their consent. The system’s recommendations will then be compared with the doctors’ decisions. Before the tool can be used widely, issues involving clinical validation, legal liability, and compliance with medical regulations must be resolved. Microsoft considers this project part of its medical AI plan aimed at surpassing the most capable human-run healthcare systems. The company is developing a separate foundation AI model for healthcare with Mayo Clinic and plans to release it in 2027. The plan is not limited to diagnostics. Treatment decisions and the coordination of medical services are also among its key areas. According to Microsoft, doctors should retain the primary role in patient care, and the system should not replace them.