ASSESSING THE READABILITY OF COLON CANCER INFORMATION GENERATED BY LARGE LANGUAGE MODELS Max Shteiman, The University of Queensland–Ochsner Clinical School, New Orleans, LA.
Introduction: Colon cancer is the second deadliest and third most common cancer in the United States. Therefore, it is important for the public to understand certain useful information regarding colon cancer, including risk factors, screening methods, and prevention strategies. Large language models (LLMs) are promising tools that may improve patient autonomy through customized education. However, the readability of the responses from these programs in the context of colon cancer patient education has not yet been thoroughly assessed. The National Institute of Health (NIH) recommends that medical information for the general public is written at a sixth-grade level. This study investigated the readability of responses to questions specific to colon cancer using three of the most popular LLMs: ChatGPT, Google Gemini, and Meta AI. Methods: Ten questions that an average adult in the United States may inquire about colon cancer were used as input in each LLM. Subsequently, each interface was tasked with revising the output to match a sixth-grade reading level. Each output set was analyzed and assigned a Flesch-Kincaid
readability score. Statistical analysis of the data was performed using GraphPad Prism 10.2.3. Results: The average Flesch-Kincaid readability scores for initial output from ChatGPT, Meta AI, and Google Gemini were 49.69, 54.77, and 62.02, respectively. The average scores for the sixth- grade level outputs from these LLMs were 77.83, 74.34, and 76.04, respectively. All three LLMs demonstrated statistically significant improvement in readability when tasked with adapting output to a sixth-grade level with the following respective p-values: <0.0001, 0.0007, and 0.0011. Discussion: ChatGPT, Google Gemini, and Meta AI each demonstrated statistically significant increases in Flesch-Kincaid scores between both sets of output, indicating an improvement in readability. The results of this study suggest that LLMs are potential tools that may empower and improve patient autonomy through customized education. Future studies could expand on this work by utilizing a larger question set as input and including other LLMs to further investigate this topic..
HUNGRY FOR HELP: THE NEED FOR FOOD SECURITY SCREENING IN NORTH LOUISIANA’S OPIOID-USING POPULATION Megan Gremillion, Claire Crosby, Morgan Uebinger, Marie Morgan Vazquez, Ammar Husan; Louisiana State University School of Medicine, Shreveport, LA.
Introduction: Numerous studies have explored the association between opioid use and food insecurity (FI). FI has been shown in the literature to induce stress, anxiety, and depression–all of which are risk factors for opioid abuse. The authors of this study hypothesize that, while FI screening in opioid-using patients is beneficial, it is an underutilized preventative health tool. Methods/Results: Frequency of FI screening was assessed via the validated Hunger Vital Sign (HVS) questionnaire in opioid-using adult patients as defined by ICD-10 code F11 in Shreveport and Monroe, Louisiana, Epic’s SlicerDicer tool was utilized to extract numerical data across three years
(2020–2023). A Chi-squared test was used to assess the association between FI and opioid use in the study’s population, and the percentage of opioid- using patients who were screened for FI was also calculated. Analysis revealed that opioid-using, adult-aged patients who were screened were 1.97 times more likely to experience FI (n=2,610, 95% CI ± 0.83, P<0.001) than patients who were screened and did not use opioids (n=32,075), yet 82.61% of opioid-using patients were not screened for FI. Discussion: While opioid use and FI appear to be strongly associated, opioid-using patients were under-screened and thus underserved. Such findings are of note to healthcare providers in North Louisiana, 73
Made with FlippingBook Digital Publishing Software