Generating Patient-Friendly Summaries of Clinical Notes With Large Language Models: A Pilot Evaluation in an Oncology Setting.

Complex medical jargon describing procedures, medications, and care notes have long posed a barrier for patients when it comes to understanding their personal medical information, impacting critical outcomes such as adherence to post-discharge care directives or overall patient literacy. With this issue in mind, there is a motivation to make clinical notes more accessible and readable for patients, while maintaining accuracy of the contained medical information. Thus, large language models (LLM) have been identified as a potential solution thanks to their ability to generate text based on given prompt at high speed based on a corpus of training data, outperforming any human summarizer. But what does it mean to make clinical notes “patient friendly” and is establishing a universal definition possible?  This research uses LLMs to explore a potential definition for “patient friendly” medical summaries, using a MIMIC-IV-Ext-BHC dataset based on 270,033 free-text clinical notes from the Beth Israel Deaconess Medical Center from 2008—2019. Several LLMs were trained on this dataset and tasked with generating a “patient-friendly” summary that prioritized the following elements: 1) explanation of the patient’s medical condition 2) actionability of text for medication or post-discharge care  and 3) prevention information . Afterwards, LLM-generated summaries were preference ranked against each other and physician-generated summaries in a double-blind study, with participants ranking their preferred summaries on a scale of 1—10. Our findings suggest that LLMs could play a pivotal role in improving patient literacy on a broader scale, without necessitating upfront financial and labor costs that would otherwise be burdensome for overworked, understaffed healthcare facilities.