Study in Israel: AI recommended unnecessary ultrasounds in 42.8% of responses — Jerusalem Post

Researchers from Israel’s Ben-Gurion University of the Negev in Beersheba and Clalit Health Services tested six large language models from Google, Anthropic and OpenAI using clinical scenarios. As the Jerusalem Post reports, the models generally identified when tests or treatment were not needed, but made mistakes more often in situations where the medical community has no unified position.

Scenarios for testing

The peer-reviewed study was published in the journal Annals of Internal Medicine. The authors created scenarios based on 15 recommendations from the international Choosing Wisely initiative, aimed at reducing low-value medical interventions.

For each recommendation, they prepared 15 situations: ten cases in which intervention should be avoided and five in which it was medically justified. The scientists analyzed thousands of model responses. For example, imaging is not needed for a young man with acute lower back pain after moving furniture, but it may be necessary for a woman with a history of breast cancer and new nighttime back pain.

More current news is available on the UA.News Telegram channel Telegram.

Where the models made mistakes most often

When an intervention was indicated, the models wrongly recommended refraining from care in 0.6% of cases, the authors said. If a test or treatment was clearly unnecessary, the systems recommended it in 5.4% of responses. In eight of the 15 recommendations examined, the models did not recommend any unnecessary procedure at all.

The greatest number of errors concerned two ambiguous areas. In 28.3% of responses, the models proposed an unnecessary electrocardiogram, mostly when patients reported palpitations. In 42.8% of responses, they also supported unnecessary ultrasound examinations of small incidentally detected nodules. The researchers linked this to differences in the approaches of the American Thyroid Association and the American College of Radiology regarding the need to image such nodules.

The models were tested only in English. The researchers stressed that further clinical studies and safeguards are needed before such systems are used in a doctor’s office, and that final responsibility for a medical decision must remain with the physician.