Can AI tell when patients don't need medical tests? Israeli study puts chatbots to the test

Israeli researchers have evaluated six leading language models, including those developed by Google, Anthropic and OpenAI, in determining when medical tests are unnecessary. The study, published in Annals of Internal Medicine, found that these AI systems successfully identified situations where testing or treatment is redundant, though they occasionally recommended such measures in cases where they were not warranted.

The researchers examined 15 Choosing Wisely recommendations, creating 15 scenarios for each. For every recommendation, they developed five scenarios where action was necessary and ten where restraint was appropriate. They then assessed how the models handled these cases.

In most instances, the AI systems correctly identified when a test or treatment was not required. However, they occasionally recommended unnecessary procedures, particularly in cases involving palpitations and unimportant nodules discovered during routine examinations. For example, the AI systems recommended electrocardiograms (EKGs) in 28.3% of cases where they were not medically necessary, and ultrasounds for incidentally found nodules in 42.8% of instances.

The study highlights the challenge of balancing overtesting and underdiagnosis, as AI models often reflect the hesitations and uncertainties found in human clinicians. The researchers suggest that further research is needed to determine whether AI assistants can maintain appropriate clinical restraint when interacting with anxious or insistent patients.