{07/09} · Trust & safety
All workFrom Physician to Prompt: GPT for Emergency Department Notes
With Dr Ekanath Rangan, Stanford School of Medicine
Can a language model at the bedside collect the history of presenting illness?
A prospective study at AIMS with bedside tablets and the GPT-3.5 API to collect the history of presenting illness.
- 208patients enrolled, 71 analysed
- 0.998PubMedBERT similarity to physician notes
- 90.1%of cases had a section above the hallucination risk threshold
What the record supports
Claim
Generated notes were very similar to physician notes.
Evidence
PubMedBERT similarity to physician notes was 0.998.
Limits, stated plainly
90.1 percent of cases had at least one section above the hallucination risk threshold.