Back

Zero-Shot Prompting is the Most Accurate and Scalable Strategy for Abstracting the Mayo Endoscopic Subscore from Colonoscopy Reports Using GPT-4

Yim, R. P.; Rudrapatna, V. A.

2024-03-24 health informatics
10.1101/2024.03.22.24304745 medRxiv
Show abstract

Structured AbstractO_ST_ABSIntroductionC_ST_ABSLarge-language models can help extract information from clinical notes, making them potentially useful for research in ulcerative colitis. However, it remains unclear if these models will scale well in practice. MethodsWe analyzed the performance and cost of programmatically using GPT-4 to abstract Mayo endoscopic subscores (MES) from 499 colonoscopy reports using different prompting strategies. ResultsZero-shot prompting, where GPT-4 is instructed without examples, was most accurate (83.55%) and cost-effective ($0.097/note). DiscussionUsing GPT-4 to automatically curate the MES and other variables is a practical strategy for quantifying UC activity and measuring improvements to clinical care.

Matching journals

The top 3 journals account for 50% of the predicted probability mass.

50% of probability mass above

"Similar papers" are the closest papers from that journal in the model's embedding space. They show what the match is built on, but the ranking comes mostly from a classifier over the whole training set, not from these examples alone.