Expertise and Model Confidence Key to Safe Radiologist-LLM Interaction

In a chest imaging study, reader expertise (odds ratio 2.06) and high model confidence (OR 3.82) were independently associated with correct radiologist-large language model collaboration. However, expert readers were less influenced by model confidence, and high rationale qualit…

• The study found that higher rationale quality from LLMs reduced radiologists' rejection of correct suggestions (OR 0.79) but also increased their acceptance of incorrect suggestions (OR 1.71).
• Higher reader expertise (OR 0.54) and the reader's own confidence (OR 0.80) were protective factors that reduced acceptance of incorrect AI advice.
• The retrospective study involved 10 readers interpreting chest imaging cases from 100 patients, comparing interpretations made with a high-accuracy (76%) and a low-accuracy (27%) large language model.

Automated summary

RadPigeon summaries are original and for information only. They are not clinical advice.