Model Confidence and Reader Expertise Shape LLM Collaboration in Chest Imaging

For chest imaging diagnosis with LLM assistance, model confidence (OR 3.82) and radiologist expertise (OR 2.06) independently predict appropriate acceptance/rejection of advice, while rationale quality increases overreliance on incorrect suggestions (OR 1.71).

• The effect of model confidence on adequate interaction was weaker among expert readers (interaction OR 0.79, P=.008).
• Higher rationale quality reduced rejection of correct advice (OR 0.79) but raised acceptance of incorrect advice (OR 1.71).
• Study included 10 readers interpreting 100 chest cases (radiography, CT, MRI, PET) with LLMs of varying accuracy.

Automated summary

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