Microsoft Launches New AI for Diagnosis: “Up to 4x Better than Doctors”

Microsoft today presented research that could mark a turning point in the relationship between artificial intelligence and medicine.

A project from the AI ​​working group led by Mustafa Suleyman has demonstrated the ability to correctly diagnose 85% of clinical cases published in the New England Journal of Medicine (one of the most prestigious medical journals in the world) . This, it says, exceeds the accuracy rate of a group of expert doctors by more than four times.

“The Microsoft AI team shares research that demonstrates how AI can sequentially examine and solve the most complex diagnostic challenges in medicine,” reads an article published on the official Microsoft website, “cases that even experienced doctors find difficult to solve.”

The new frontier of sequential diagnosis

The heart of the innovation of MAI – DxO (this is the name of the program) lies in its ability to replicate clinical reasoning step by step. That is, exactly as a doctor would do in real practice. It is not about answering multiple choice questions, but about following a real diagnosis process, starting from the symptoms, asking specific questions, ordering tests and evaluating the results. More than a single analysis, a somewhat accurate replication of clinical reasoning.

“At Microsoft AI, we are working to improve and assess clinical reasoning capabilities. To go beyond the limitations of multiple-choice tests, we focused on sequential diagnosis, which is the basis of real-world medical decision-making,” the site continues on the page presenting the research results.

A concrete example among those proposed in the study. According to the researchers, a patient with a cough and fever could make the system decide to first request blood tests and then a chest X-ray, before arriving at the correct diagnosis of pneumonia, just like a doctor would do.

AI vs. Doctors: Software Beats Human Team 85% to 20%

The team used 304 real clinical cases from the NEJM’s weekly Case Records series, turning them into a benchmark called the SD Bench (Sequential Diagnosis Benchmark). The performance of MAI?DxO, which coordinates large language models such as OpenAI o3, Gemini, Claude, Grok, and others, surprised the researchers, who wrote: “Comparing the results with real cases published weekly in the New England Journal of Medicine, we show that the Microsoft AI Diagnostic Orchestrator (MAI-DxO) correctly diagnoses up to 85% of cases, more than four times higher than a panel of expert clinicians.”

The 21 doctors involved in the evaluation – from the United States and the United Kingdom – obtained an average of 20% correct diagnoses. Microsoft’s system also proved to be more cost-effective, choosing more targeted and less invasive tests. “MAI-DxO arrives at the correct diagnosis more cheaply than doctors,” they add.

Not a replacement for the human, but an enhancement

Microsoft was keen to point out that MAI?DxO is not intended to replace doctors, but to support them in the most difficult cases or when there is a lack of access to specialists.

“The role of doctors is much broader than just diagnosis. They must be able to manage ambiguous situations and build trust with patients and their families, something that AI is not able to do today,” it continues.

The company admits that the system is not yet ready for clinical use, but believes that in the future it could improve access to care, reduce errors and help in particularly complex situations.

Towards Medical “Superintelligence”

There is an expression chosen by Microsoft to define the path: “Path to Medical Superintelligence” , a road to medical superintelligence. The goal: an AI that not only equals but surpasses human capabilities in medical diagnosis, treatment and prediction.

The paper explains that “Taking this level of reasoning—and beyond—could profoundly change the healthcare system. AI could help patients autonomously manage simple aspects of their health and provide doctors with advanced support for more complex cases.”

Microsoft announced that it will make the benchmark public for scientific collaboration and verification. Future tests will focus on more common symptoms to confirm reliability in ordinary situations.

The project comes at a time when global health spending is soaring, with US costs - the Guardian recalls - exceeding 20% ​​of GDP and where up to a quarter of spending is considered inefficient.