Microsoft's Diagnostic Orchestrator just hit 83.2% accuracy on NEJM's hardest diagnostic cases vs 36.1% for human docs with internet access. The architecture matters: it runs a multi-agent loop where the AI asks an average of 25.5 questions before ordering tests, compared to 5.6 for physicians. That question-first approach mirrors classic expert systems from the 70s but now runs on LLMs with real hospital pricing constraints.
On routine ER cases it scored 80.3% vs 46.8% human baseline. When budget-constrained, it maintained ~80% accuracy while cutting test costs significantly. The system treats diagnosis as a sequential decision problem: gather history → narrow differential → order minimum viable tests → iterate.
This isn't just pattern matching on symptoms. It's reasoning under uncertainty with economic constraints, which is exactly what expert systems tried to do in 1968 but lacked the compute and training data to pull off. The gap between asking 5 questions and 25 questions is the gap between heuristic shortcuts and actual Bayesian updating.
Full study dropped in Science, October 2026. We're watching the expert system dream from the 70s finally work at scale.
On routine ER cases it scored 80.3% vs 46.8% human baseline. When budget-constrained, it maintained ~80% accuracy while cutting test costs significantly. The system treats diagnosis as a sequential decision problem: gather history → narrow differential → order minimum viable tests → iterate.
This isn't just pattern matching on symptoms. It's reasoning under uncertainty with economic constraints, which is exactly what expert systems tried to do in 1968 but lacked the compute and training data to pull off. The gap between asking 5 questions and 25 questions is the gap between heuristic shortcuts and actual Bayesian updating.
Full study dropped in Science, October 2026. We're watching the expert system dream from the 70s finally work at scale.