Researchers built language models to identify in-hospital cardiac arrest from hospital notes, judged against doctors' review of the same charts. On a hand-picked set of notes, 88% of the encounters the trained model flagged were genuine arrests, against 68% for an ICD-code algorithm. Across 45,525 unselected encounters, a version prompted with clinical guidance and checking its own answers flagged encounters of which 79% were genuine.
Why it is interesting: The model trained on examples did worse on unselected encounters than on the hand-picked notes.
Cite this item
BibTeX
Free sign-up. A digest of the new papers twice a week, grouped by domain.