Prevalent knowledge omissions: An unseen risk of LLMs in patient-friendly adaptation of medical literature

Abstract

Large language models (LLMs) can produce patient-friendly adaptations, such as simplifying specialized medical literature for readers with a middle school education. However, any distortion of knowledge during this process may have serious consequences. Current evaluations overlook such risks. In this study, we proposes a knowledge triplet comparison (KTC) method to assess knowledge distortion by systematically quantifying the retention and alteration of medical knowledge before and after text adaptation. Using the Chemical-Disease Relations dataset, this study evaluates the performance of six LLMs across eight adaptation scenarios (e.g., adjusting reading level, modifying text length) and draws the following key conclusions: First, from a quantitative perspective, knowledge distortion is prevalent in all adaptation scenarios, with the models retaining only approximately half of the original medical knowledge. Additionally, about 10% more knowledge is added. Second, the distribution of knowledge distortion follows a clear categorical distribution pattern, with etiology and pathological mechanisms, as well as complications and risk factors, being the most frequently distortion types. Third, targeted corrections through prompt engineering have no significant impact on the quantity and categorical distribution patterns of knowledge distortion, indicating that knowledge distortion is persistent and regular. This study shows that, in addition to the more obvious risk of hallucinations caused by a small amount of added knowledge, patient-friendly rewrites by LLMs also pose a significant risk of subtle yet potentially serious knowledge omissions. Furthermore, this research reveals the regularity of knowledge distortion, laying the groundwork for further investigation into the patterns of knowledge distortion and the development of solutions to this problem.