What the research found
UT Southwestern researchers paired emergency department patients from Stanford Medicine and Beth Israel Deaconess Medical Center who looked similarly ill on arrival, where one later deteriorated and one did not, and asked three open-source language models which to prioritise. The models picked the higher-risk patient well above chance, but adding details unrelated to the immediate complaint, such as frequent emergency visits or a history of psychiatric emergencies, lowered the priority of patients who went on to deteriorate, with nearly 10% wrongly downgraded in some scenarios. The models were tested on past records rather than used in live triage, and the effects varied by model and scenario.
Triage tools learn from charts that already carry clinicians' biases, so providers should test any AI triage aid on the patients most often labelled frequent attenders before trusting it at the front door.