Large language models successfully avoid recommending unnecessary care about 95% of the time when tested against Choosing Wisely guidelines, though they struggle with thyroid ultrasounds for small nodules where overuse was recommended in over 40% of responses.
- LLMs recommended unnecessary care in only 5.4% of cases when tested against 15 Choosing Wisely recommendations with 225 scenarios
- Thyroid ultrasound for small incidental nodules showed the highest overuse rate at 42.8% of LLM responses, driven by conflicting society guidelines
- Routine electrocardiography in low-risk adults was the second-highest area of concern with 28.3% overuse recommendations
- Eight of 15 Choosing Wisely recommendations had zero overuse recommendations from LLMs
- False restraint (failing to recommend necessary care) occurred in only 0.6% of cases, indicating LLMs rarely undertreat
Large language models (LLMs) rarely recommend unnecessary care in Choosing Wisely tests, though thyroid ultrasound is more susceptible, suggest findings published October 19 in the Annals of Internal Medicine.
LLMs in general aligned with evidence-based medical guidance based on recommendations from the Choosing Wisely initiative, recommending potentially unnecessary care in about 5% of responses, wrote a team led by Dean Ariel, MD, from Clalit Health Services in Tel Aviv, Israel. However, LLM-recommended overuse occurred in over 40% of thyroid ultrasound exams for small incidental nodules.
This test “is also the one on which society guidance is openly contested,” the Ariel team wrote.
Despite one decade of the Choosing Wisely campaign naming commonly overused low-value services, including those used in medical imaging, overuse persists. Clinicians meanwhile are increasingly using LLMs for assistance on medical advice. However, the team noted that few studies have evaluated whether these models can distinguish appropriate restraint in low-value care scenarios from similar scenarios where additional testing or treatment is appropriate.
Ariel and colleagues used 15 Choosing Wisely recommendations to develop 225 scenarios. They tested the performance of six commercial LLMs that generated 4,050 responses to the scenarios, using a prompt from the clinician perspective.
LLM-recommended overuse occurred in 5.4% of cases. And LLMs did not recommend overuse in eight of 15 recommendations.
Two recommendations accounted for most LLM-recommended overuse: thyroid ultrasound for an incidental sub-centimeter nodule (overuse recommended in 42.8% of LLM responses) and routine electrocardiography in low-risk adults (overuse recommended in 28.3% of LLM responses).
The LLMs recommended overuse in 0.6% to 3.3% of responses for the remaining five recommendations. False restraint occurred in 0.6% of cases, was absent for 11 of 15 recommendations, and was highest for prostate-specific antigen screening at 3.3%.
The researchers highlighted that thyroid ultrasound for small incidental nodules is the only recommendation in the study with active disagreement between society guidelines.
“The American Thyroid Association advises dedicated ultrasound of incidental subcentimeter nodules, whereas the American College of Radiology advises no further workup below 1.0 to 1.5 cm depending on age,” they wrote. “What divides them, and what the models confronted, is precisely whether to image.”
The authors called for targeted improvements in areas where overuse is more common.
Studies that use real-life cases are needed to measure AI’s clinical value, according to an accompanying editorial written by Deborah Korenstein, MD, from the Icahn School of Medicine at Mount Sinai, and Daniel Morgan, MD, from the VA Maryland Health Care System and the University of Maryland in Baltimore.
“This study [by Ariel et al] represents an important step toward measuring the impact of AI on value in terms of both overuse and underuse,” they wrote. “To our knowledge, it is the first published study to explicitly address overuse.”
Read the full study here.
