Zara Baloch, Observer at HCA Houston Healthcare, shared a post on X:
“AI hallucinations in oncology are not just technical errors, they can become patient safety risks.
At ASCOBT26, Abstract #19 evaluated disease-specific safety risks of LLMs across solid tumor subtypes using 186 tumor-board vignettes.
The key issue: a model may appear reliable in broad oncology prompts but still produce unsafe or guideline-discordant responses in specific disease contexts.
The study assessed guideline concordance, hallucinations, safety risk, readability, and reasoning highlighting why oncology AI needs disease-level validation before clinical use.”
Title: Disease-specific safety risks in oncology large language models: A multi-axis evaluation across solid tumor subtypes
Authors: Yan Leyfman, Connor Yost, Helena S. Coloma, Muskan Joshi, Taha Kassim Kassim Dohadwala, Soumiya Nadar, Harashita Vallabhaneni, Diksha Sanjana Pasnoor, Chandler H. Park, Arturo Loaiza-Bonilla
Other articles featuring Zara Baloch on OncoDaily.