A study published in medRxiv highlights the issue of AI hallucinations in foundation models used in medicine. These hallucinations, defined as instances where models generate misleading medical content, pose a significant risk to clinical decisions and patient safety. The study focused on understanding the characteristics, causes, and implications of these hallucinations, using a taxonomy, benchmarking models with a medical hallucination dataset, and analyzing physician-annotated responses to real medical cases. The research revealed that techniques like chain-of-thought and search augmented generation can reduce hallucination rates. However, despite these improvements, non-trivial levels of hallucinations persist. A multi-national clinician survey emphasized the need for robust detection and mitigation strategies, as well as clearer ethical and regulatory guidelines. In related news, Medicomp Systems’ CEO reported that 8% to 10% of AI-captured information from complex medical encounters may be correct, with their tool designed to flag these issues for review. Additionally, the American Cancer Society and Layer Health are collaborating to use LLMs to expedite cancer research, analyzing data from 300,000 participants in the Cancer Prevention Study-3.
Source: www.mobihealthnews.com















