Why Chatty AI Doctors Order Too Many Tests

Here's a strange thing about large language models: they can recite basically the entire medical curriculum, yet when you ask them to actually diagnose a patient step by step, they panic and order every test in the building. Turns out knowing things and reasoning under constraints are very different skills. Encyclopedic ≠ economical.

A new framework called GraphDx tries to fix this by giving the AI something like a map instead of just a library. Researchers built "Medical Diagnosis Knowledge Graphs" — structured networks where symptoms, tests, and diagnoses are connected not just by relevance, but by cost. Some connections say "this test is very telling." Others whisper "this test is expensive, use sparingly."

Then comes the clever part: instead of one AI doing everything, GraphDx splits the job among three specialized agents. One perceives the case (reads and understands the patient info), one reasons (navigates the knowledge graph with cold, deterministic logic), and one decides what to say next. It's basically the difference between asking one overwhelmed intern to do triage, workup, and paperwork simultaneously, versus having a team where everyone has a lane.

The results are the kind of numbers that make you sit up. Tested on medical exam questions and real ICU records, across several different AI models, GraphDx pushed diagnostic success from a shaky 50–68% up to a confident 79–93% — while cutting the number of tests ordered by 20–54%.

Same brain, better plan. Sometimes the fix for an AI that knows too much isn't more knowledge — it's a map, some discipline, and a little help staying in its lane.

Distilled from arXiv cs.AI