•  
  •  
 

Document Type

Conference Proceedings

Abstract

Objective: This study aims to evaluate the accuracy, cultural appropriateness, and readability of AI-generated discharge education materials for diabetic ketoacidosis (DKA) in the Emergency Department (ED). Methods: Ten representative DKA case scenarios reflecting diverse cultural, linguistic, and socioeconomic contexts were developed. ChatGPT generated customized discharge instructions for each case based on a rubric derived from American Diabetes Association (ADA) guidelines, incorporating tailored recommendations such as culturally relevant meal plans, cost-conscious insulin access options, and follow-up resources. The same rubric was used by three human reviewers and an independent AI model (Google Gemini) to grade the materials for accuracy, personalization, and health literacy. Qualitative differences were analyzed to assess efficacy and identify areas for improvement in AI-generated patient education. Results: Mean composite human rating across all domains was 3.7 ± 1.0 (out of 5), while Gemini’s mean was 4.7 ± 0.4. Highest inter-rater consistency occurred in language appropriateness (4.0 ± 0.9) and culturally relevant meal plans (4.3 ± 0.7). Presentation of ADA guidelines at 7th-grade reading level (3.3 ± 1.1) and affordable food resources (3.1 ± 1.2) showed the lowest averages. Inter-rater agreement among human graders was substantial (κ ≈ 0.75), and AI-to-human correlation was strong (r ≈ 0.82). Overall, 80% of ratings met “Good” or “Excellent” criteria. Conclusion: AI-generated discharge instructions demonstrated strong alignment with ADA standards and effective cultural and socioeconomic personalization, supporting their potential to improve efficiency and equity in ED discharge education.

Download

Share

COinS