OpenEvidence outperforms ChatGPT-3.5 and matches ChatGPT-5 for ACR neuroimaging appropriateness selections

OpenEvidence had top accuracy for selecting neuroimaging per ACR Appropriateness Criteria: 79.5% on select-all and 86.3% on open-ended prompts, versus ChatGPT-5 (71.2%, 80.6%) and ChatGPT-3.5 (44.4%, 69.2%). Acuity did not affect performance across 78 prompts.

• On select-all-that-apply prompts, accuracy was OpenEvidence 79.5%, ChatGPT-5 71.2%, ChatGPT-3.5 44.4%.
• On open-ended prompts, OpenEvidence (86.3%) was significantly more accurate than ChatGPT-3.5 (69.2%) but not significantly different from ChatGPT-5 (80.6%).
• Accuracy did not differ by case acuity across 78 prompts from 11 ACR priority clinical areas.

Automated summary

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