Microsoft CEO Satya Nadella announced a new AI healthcare system, MAI-DxO, that simulates multiple virtual doctors collaborating to diagnose complex medical cases.
In tests involving 304 difficult cases from the New England Journal of Medicine, MAI-DxO correctly diagnosed 85.5% of cases, significantly outperforming a group of 21 experienced physicians who achieved a 20% success rate.
MAI-DxO functions by mimicking a panel of medical experts using the Sequential Diagnosis Benchmark (SDBench). It begins with limited patient data, asks follow-up questions, orders tests, and continuously refines its diagnosis, balancing accuracy with cost efficiency.
The AI team includes roles such as Dr. Hypothesis, who tracks likely diagnoses; Dr. Test-Chooser, selecting informative tests; Dr. Challenger, who searches for contradictory evidence; Dr. Stewardship, who manages test costs; and Dr. Checklist, who ensures reasoning consistency.
At an optimal setting, MAI-DxO reached 85.5% accuracy with a per-case cost of $7,184, compared to OpenAI’s standalone model achieving 78.6% accuracy at $7,850 per case. In a more cost-conscious mode, it maintained 80% accuracy while spending 20% less than typical physician costs.
Microsoft highlights that the AI was tested on recent cases it could not have memorized and that the participating physicians worked without external aids to ensure fair comparison.
MAI-DxO operates in various modes, including “Instant Answer,” “Question Only,” “Budgeted,” and “Ensemble,” which run multiple panels for maximum accuracy.
This project is part of Microsoft’s larger effort to integrate AI into healthcare, noting over 50 million daily health-related interactions on Bing and Copilot. The approach aims to augment physicians, not replace them.
While promising, Microsoft researchers stress the system is a demonstration requiring further validation, safety testing, clinical trials, and regulatory approval before deployment.
With medical diagnostic errors affecting millions annually and contributing to significant mortality, MAI-DxO represents a step forward in AI-assisted healthcare diagnostics.