AI Chatbots Show Promise in Identifying Unnecessary Medical Tests
Translated & summarized from Ynet by baba
A study by Clalit and Ben-Gurion University found that AI chatbots can identify unnecessary medical tests in most cases, making mistakes in 5.4% of situations where no intervention was needed. The AI models, tested on hundreds of scenarios, performed well in areas of medical consensus but struggled with topics debated by professionals. Researchers noted that while AI is improving, it still lacks human nuance and can be influenced by user anxiety and language.
The story in 6 lines · by baba
- AI chatbots can identify unnecessary medical tests, according to a Clalit and Ben-Gurion University study.
- AI models recommended unnecessary interventions in 5.4% of cases where none were needed.
- Errors were concentrated in areas with professional disagreement, like thyroid ultrasounds and EKGs.
- The AI's ability to decline unnecessary tests surprised researchers.
- Factors like patient anxiety and query language can influence AI recommendations.
- Ongoing evaluation of AI safety in medicine is crucial, researchers stated.
A joint study by Clalit Health Services and Ben-Gurion University of the Negev has found that large language models, the technology behind AI chatbots, can effectively identify when medical tests or treatments are unnecessary. Researchers tested six different AI models, including those from Google, Anthropic, and OpenAI, using hundreds of medical scenarios. The study, published in the Annals of Internal Medicine, revealed that in most cases, the AI models correctly distinguished between situations requiring intervention and those that did not.
According to the findings, the AI models mistakenly recommended avoiding necessary medical intervention in only 0.6% of cases. However, they recommended unnecessary interventions in 5.4% of cases where no action was needed. Many of these errors occurred in areas where medical professionals themselves have differing opinions, such as the necessity of ultrasounds for small thyroid nodules or routine EKGs for low-risk individuals. The study's lead researcher, Dr. Dean Ariel, a family doctor and AI researcher, expressed surprise at the AI's ability to decline unnecessary tests, noting that the models learned from medical texts, including less optimal practices.
The research was inspired by the "Choosing Wisely" initiative, which aims to reduce superfluous medical procedures. Dr. Ariel explained that the study focused on both directions: ensuring AI doesn't recommend unnecessary actions and doesn't withhold necessary ones. The scenarios were carefully crafted to be neutral, avoiding emotional language that could influence the AI's response, though the researchers acknowledge that factors like patient anxiety and the language of the query can affect chatbot recommendations in real-world use.
While the AI models demonstrated a significant ability to identify unnecessary procedures, the study also highlighted variations between different models, with some making far more unnecessary recommendations than others. Dr. Guy Lavan, a senior faculty member at Ben-Gurion University, noted that AI responses can differ significantly on topics where human experts disagree. The researchers emphasize the need for ongoing evaluation of AI systems as they are updated, suggesting that while AI can be a useful consultation tool, it still lacks the nuanced human understanding required for complex medical decision-making.