Large language models seldom over-recommend imaging, but thyroid ultrasound guidance shows substantial discrepancies.
Key Details
- 1In a test of 4,050 LLM responses to 225 clinical scenarios from 15 Choosing Wisely recommendations, unnecessary care was recommended in only 5.4% of cases.
- 2Thyroid ultrasound for incidental small nodules led to overuse recommendations in 42.8% of cases—reflecting controversy between ACR and ATA guidelines.
- 3Eight out of fifteen recommendations saw no overuse recommendations from the LLMs.
- 4Routine ECG in low-risk adults was another area of overuse (28.3% of responses).
- 5False restraint (withholding needed care) was rare, highest for PSA screening at 3.3%.
- 6Researchers emphasize targeted AI improvement and call for studies using real-world cases.
Why It Matters
This marks the first published study explicitly quantifying LLM-induced imaging overuse, highlighting a nuanced risk: while LLMs are generally safe, domain-specific guideline controversies (e.g., thyroid imaging) may propagate unnecessary tests. Ensuring proper AI alignment with radiology guidance is critical to avoid both overuse and underuse in clinical decision support.
Source
AuntMinnie
Related News
•Radiology Business
UC Irvine Implements AI to Find Lung Cancer Screening Candidates
UC Irvine's radiology department developed an NLP tool to identify patients eligible for lung cancer screening from EHR notes.
•Radiology Business
Webinar to Showcase Successful AI Automation in Radiology Workflows
A webinar will provide insights from imaging leaders on expanding and optimizing AI-driven automation in radiology clinical workflows.
•Radiology Business
A Practical Guide to Evaluating Radiology AI Solutions
Experts propose a three-step approach for radiologists to assess and select AI imaging solutions.