- Radiologists receiving chain-of-thought explanations achieved the highest diagnostic accuracy, outperforming the control group by 12.2 percentage points.
- Standard-output and differential-diagnosis formats did not significantly improve physician performance compared with no large language model support in the primary analysis.
- Differential-diagnosis outputs were associated with higher adherence to incorrect model recommendations, suggesting a potential risk of automation bias.
- GPT-4 achieved 80% diagnostic accuracy with chain-of-thought prompting compared with 75% using standard prompting and 65% top-1 accuracy using differential-diagnosis prompting.
- The study was conducted in a controlled radiology vignette setting and did not evaluate patient outcomes, real-world workflow integration, or long-term clinical use.
Daily News
Stay up to date with the latest clinical headlines and other information tailored to your specialty.
