Multi-Reader Study Finds Radiologists Miss Up to 11% of Clinically Significant Errors in LLM-Generated Patient Reports

Radiologists missed up to 11% of clinically significant errors in LLM patient communications; dual-reader review cut miss rate from 12.4% to 0.6%.

• Retrospective multi-reader study of 3 radiologists evaluating 1,200 LLM-generated answers from 187 MIMIC-IV radiology reports.
• For clinically significant errors, single-reader miss rate was 5.6–11.1%; dual-reader reduced overall deficiency miss rate to 0.6%, triple-reader eliminated misses.
• Limitations: retrospective, single-center, simulated patient questions, only 3 readers; results may not generalize to real-world settings.

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