Microsoft researchers have developed an artificial intelligence system designed to diagnose medical conditions the way clinicians often do: one careful step at a time. In a study published in Science, the tool reportedly diagnosed difficult medical cases more than twice as accurately as general physicians.
Instead of reviewing every symptom, lab result, and scan at once, the system—called the Diagnostic Orchestrator, or DxO—works through a case progressively. It asks relevant questions, reviews findings, orders tests, interprets the results, and revises its possible diagnoses as new information emerges.
How the AI Diagnostic System Works
The technology uses an “orchestrator” to coordinate multiple AI agents that behave like a simulated team of clinical specialists. Together, they decide what patient-history questions matter, which physical-exam findings deserve attention, and which tests should come next.
That structure differs from many current medical AI tools, which often process a complete set of patient information and produce a diagnosis in a single pass. Microsoft’s approach is intended to mirror the judgment involved in a real diagnostic workup.
Stronger Results Than Human Doctors
Researchers tested the system on more than 300 unusually complex medical cases. DxO achieved about 83% diagnostic accuracy, while generalist physicians using textbooks and online references reached roughly 36%.
The AI also performed better on emergency-department cases drawn from a large hospital admissions database. In that test, DxO reached about 80% accuracy, compared with roughly 47% for physicians.
Lower Costs and More Careful Testing
Microsoft says the system did more than improve accuracy. It also designed diagnostic workups that cost less than those created by human doctors in the study.money.usnews+1
Harsha Nori, head of AI at Microsoft AI Health, said the research measured not only whether the AI reached the right answer, but also how carefully and efficiently it moved through the diagnostic process.
What Happens Next
Microsoft plans to work with Beth Israel Deaconess Medical Center in Boston to test whether the tool can help triage patients waiting in emergency departments.
The findings are promising, but they do not prove the system is ready for everyday clinical use. The study used controlled, text-based cases, and researchers say further real-world validation is still needed.
OpenAI Uses Thousands of AI Agents for Brain Research
In a separate development, OpenAI partnered with PerturbAI and Scripps Research to show how large groups of AI agents could speed up biological research. The team used thousands of GPT-Rosalind agents to analyze experimental data from millions of mouse brain cells in just 32 minutes.
Rather than relying on one model to examine everything, the researchers assigned each agent a specific genetic change to investigate. The agents studied how brain cells responded and searched scientific literature for possible explanations.
What the AI Agents Found
The project produced what researchers call a “reasoning atlas,” a collection of connections between genes, cellular changes, and existing scientific knowledge. The team estimates that reviewing the same material manually could take months or even a year.
One notable lead involved DCTN1, a gene linked to Perry syndrome, a rare neurodegenerative disorder that can cause Parkinson-like symptoms. When researchers disrupted the gene, they saw increased activity in genes involved in cholesterol production, suggesting a possible connection worth further study.
The agents also found unexpected differences between Tsc1 and Tsc2, two genes that normally help regulate cell growth and metabolism. Although the genes work together, disrupting them produced different effects in certain cerebellar neurons, the brain region involved in movement coordination.
Why Many AI Agents Beat One
The researchers compared the multi-agent approach with a single powerful AI system handling the same workload. At the largest scale, one agent cited only 11 scientific references, while the distributed agents cited more than 700.
Human experts also reviewed the reports without knowing which method produced them. They preferred the multi-agent reports 75% of the time, suggesting that dividing a large research problem into smaller tasks can produce broader and more useful analysis.techrepublic+1
What This Means for Medicine
These projects do not mean AI has cured Alzheimer’s, Parkinson’s disease, or other neurological conditions. The OpenAI findings are early research leads that scientists still need to test in the lab.openai+1
Still, both studies point to a similar idea: AI may help doctors and researchers handle information more efficiently. Microsoft’s system aims to make diagnosis more structured and cost-effective, while OpenAI’s agent network shows how AI can uncover patterns hidden in massive biological datasets.