AI diagnosis system learns to order medical tests efficiently — first to model what evidence matters

Researchers built an AI system that decides what tests to order during diagnosis, rather than assuming all patient information is already available. The system learns which tests reduce uncertainty most efficiently, meaning doctors could get diagnoses with fewer tests and less guessing about what to check next.

Every diagnostic encounter involves deciding what to test. Until now, AI diagnosis systems either assumed perfect information upfront (unrealistic) or ordered tests randomly without learning from patterns. This work treats test-ordering itself as a learnable skill — the AI figures out which evidence actually matters. The practical implication is narrow: this is a benchmark paper with accuracy improvements on a single dataset (MIMIC-CDM), not a deployed system changing how doctors work. The work is technically sound but operates entirely in simulation.