4 Can AI enhance understanding and communication in clinical research and patient care?

Objectives/Goals: The need for accessible medical information has never been greater. Existing research and clinical documents are often too complex for the public to understand. This study aimed to evaluate the use of AI algorithms to convert research and clinical materials into language that is more easily understood while maintaining accuracy and informativeness. Methods/Study Population: We abstracted 21 half-page texts from informed consents, clinical documents, abstracts, and health-related websites. Each original text was then rewritten using 3 AI algorithms (resulting in 63 new, rewritten texts). Participants were given a set of 3 randomly selected surveys. Each survey contained the original text and an AI-rewritten text. Participants were then asked to answer 5 questions using a Likert scale 1-5 about the readability, clarity, and content of the two texts, their overall score of the new text, and which text they preferred. The study was IRB-approved. Participants were given a 10 gift card. Results/Anticipated Results: 68 participants (69. 6% non-white, 65. 2% female, mean age 58. 3) completed the study task. 3 AI Algorithms were tested: * ConBART Cripwell 2023: context-aware method for text simplification * UL+Decoder Flores 2023: Flesh-Kincaid readability-driven method for text simplification * UBclear House-made: Iteratively improves readability with self-correction and retrieval-augmented generation. ConBART and UL+Decoder scored similarly across readability, clarity, and detail-fidelity metrics, but were preferred over original texts less than 50% of the time. UBclear scored higher in readability, clarity, detail, and overall score. Importantly, UBclear text was preferred over the original text 80% of the time, showing stronger alignment with human judgments of clarity, readability, and fidelity. Discussion/Significance of Impact: Our pilot study shows that AI can simplify existing research and medical care documents, potentially contributing to inclusivity and transparency in research and clinical care. “UBclear” was superior to the other 2 AI methodologies in readability, accessibility, and clarity while also being preferred over original texts and competing AI versions.