Duke-linked team's SAFE-Triage framework tests emergency triage AI; prompted LLM hits 57.5% ESI accuracy
The framework found rule-augmented prompting cut accuracy to 42.8% and raised undertriage to 29.8%, though it produced shorter, more readable outputs.
What happened
Researchers built SAFE-Triage, a way to test AI chatbots that tell patients how urgent their symptoms are. It checks more than whether the AI picks the right urgency level. It also checks how clearly the AI talks to patients and how useful its summary is for clinicians.
Why it matters to you
People may soon ask AI tools whether to go to the emergency room. A tool that rates a serious case as less urgent could put a patient at risk.
The takeaway
Judging a triage AI on accuracy alone misses safety and communication problems.
What it means for you
Evaluate triage tools on undertriage, communication, readability and summary quality, not on acuity accuracy alone.
Vendors of patient-facing triage tools can use a multidimensional evaluation to see where clinician oversight is needed.