- Subject
- Nutrition
- Filed
- Source
- Shelly Fan
- Loans
- 369
Microsoft’s MAI-DxO AI Diagnosed Test Cases More Accurately and at Lower Cost
In simulated cases, MAI-DxO outperformed 40 physicians on diagnosis and estimated testing costs. Real-world clinical performance remains untested.

Microsoft’s MAI-DxO system performed better than physicians on simulated diagnostic cases while using less costly testing. The results suggest that AI could help clinicians weigh diagnostic clues and test choices together. They don’t show that the system is ready to diagnose patients in routine care: the evaluation used controlled, text-based cases rather than live clinical encounters.
The study, published in Science, tested 304 complex cases drawn from the New England Journal of Medicine’s clinicopathological series and a similar number of typical emergency department cases from a clinical database. The cases ranged from common conditions to rare disorders. Each was turned into an interactive challenge that began with limited information, such as a patient’s age and main symptom. Participants could request medical history, order tests and revise their working diagnosis as new results arrived.
Diagnosis and testing costs
Forty physicians, each with roughly 12 years of clinical experience, took part. In the complex cases, MAI-DxO reached the correct diagnosis 83 percent of the time. Physicians relying on their own knowledge were correct 22 percent of the time; access to the internet and other resources improved their results only slightly. In the emergency department cases, the system was accurate in 80 percent of cases, compared with 47 percent for physicians.
The researchers also assessed the estimated cost of the tests selected along the way. A cost-conscious version of MAI-DxO kept accuracy close to 80 percent on the complex cases, with an estimated cost of about $1,300 per case. The physicians’ testing decisions added up to more than $4,700 per case, using 2023 hospital pricing data for the estimates. These figures reflect modeled costs within the study, not bills that patients would necessarily receive in clinical practice.
MAI-DxO combines several AI roles, with GPT-5 providing the underlying model. One role proposes possible diagnoses; another suggests questions and tests, then challenges the developing explanation. A cost-focused role considers the price of the diagnostic path, while an orchestrator coordinates the exchange, evaluates test results and decides whether further testing may be worthwhile.
That structure performed better than GPT-5 running on its own. In one example, the system linked a patient’s diet with bleeding gums and tooth loss, which led it to consider scurvy, a condition caused by too little vitamin C. A targeted blood test confirmed the diagnosis after estimated testing costs of just over $1,100. GPT-5 alone instead requested a broad neurological workup that cost nearly 7.6 times as much in the study’s estimates.
Why the setup matters
Many earlier AI evaluations have tested whether a model can identify the right answer when a case is already fully described. That can measure medical knowledge, but it leaves out a central part of diagnosis: deciding what information to seek, which tests to order and when the evidence is sufficient to change course.
In the MAI-DxO study, the system had to gather clues over multiple steps, as clinicians do. The researchers used the same case format for the physicians and the AI, and their benchmark tracked two outcomes: whether the final diagnosis was correct and how much the chosen diagnostic route was estimated to cost.
AI has also shown strong results in selected tasks such as interpreting mammograms and identifying signs of skin cancer, cancer spread and heart disease. Those findings don’t settle how well AI can manage a whole diagnostic process. A clinician must work with incomplete histories, interpret images and other clinical information, and adapt when early assumptions fail. The study’s simulations did not reproduce all of those conditions.
The results are also specific to this comparison and its case sets. They don’t establish that MAI-DxO will perform as well across different hospitals, patient groups or conditions that change over time. Chronic illnesses and evolving diseases may require repeated assessments that a short case simulation cannot capture.
Clinical use still needs study
A diagnosis depends on more than selecting a test on paper. Patients may decline a procedure, and a useful system would need to account for their goals and preferences while helping clinicians consider other options. Study author Eric Horvitz said patient understanding and individual preferences must be part of the process.
The Microsoft team is working with Beth Israel Deaconess Medical Center in Boston to evaluate the system across hospital settings. The planned work includes examining how MAI-DxO might fit into clinical routines, how patients interact with it and how its assessments compare with doctors’ judgments.
For now, the clearest finding is that a coordinated AI system can outperform the doctors in these simulations on both accuracy and estimated testing costs. Whether that advantage holds up amid the time pressure, varied records and patient decisions of everyday care remains an open question.