AI to Create Accessible & Reliable Patient Education Materials

Overview

This study investigates the use of Generative AI (GAI) to support primary care practices in delivering accurate, accessible patient education. With the rise of health misinformation, increasingly complex patient needs, and a strained healthcare workforce, primary care must find new ways to communicate trusted health information effectively. Leveraging the Canadian Primary Care Information Network (CPIN), this study will generate patient education messages on key health topics using both GAI and human content experts. Diverse review panels of patients and providers will assess the messages on quality of information, adaptability, and relevance and usefulness, with special attention to socioeconomic factors that may impact message accessibility. CPIN will recruit a diverse sample of participants to evaluate both GAI- and human-generated messages. Review panels will provide structured feedback via surveys, aiming to identify differences in content quality and effectiveness. The study's goal is to determine whether GAI can produce high-quality health information that meets primary care standards. Results will reveal how GAI tools can support primary care in reducing misinformation and administrative burdens, fostering patient-provider relationships, and improving health equity. Findings will inform best practices for integrating GAI in primary care to ensure accessible, timely patient education across Canada.

Background. The increasing prevalence of health misinformation, complex patient needs, and a strained healthcare workforce necessitate innovative approaches to patient education in primary care. Generative AI (GAI) offers the potential to deliver accurate, accessible health information while reducing administrative burdens. This study explores the use of GAI to support primary care practices in producing trusted, high-quality patient education materials. Objective. The investigators propose to leverage advances in GAI and our experience with CPIN to provide timely and accurate health information for primary care practices across Canada. Our goal is to determine whether GAI can produce education material for primary care that is non-inferior compared to experts in primary care and public health. Methods. The content team for this study will consist of experts specializing in primary care, public health, or health communication. Team members will create digital health messages in two formats: a short, text-messaging format (850 characters or less), and a one-page handout for patients. On the other hand, a generative AI system will also generate messages. Topics and prompts for message writing will be provided to both the content team and the GAI. Messages in English and French will be available. To evaluate the generated content, two review panels (a review panel of 25 providers and one of 25 patients) will assess messages created by both human experts and generative AI over the course of 12 months. Each month, using an evaluation grid provided to assess the quality of information, adaptability, and relevance and usefulness of the message, panelists will review a total of 16 messages (four topics x 4 messages). Panelists will be blinded to the message generation source (AI or human). Short messages will be shown first to minimize potential bias from the detailed information in longer messages and ensure their clarity and completeness are effectively assessed. Both providers and patients on the review panels will complete assessments via REDCap surveys. The evaluation grid will be the same for providers and patients and will use a Likert scale from 1 to 4 (1: Strongly disagree; 4: Strongly agree). Specifically, there will be statements on Adaptability (subcategories: Clarity and understandability, Appropriate emotional appeal, Appropriate rational appeal, Tone, and Inclusivity) and Relevance and Usefulness. Statements on Quality of Information (subcategories: Accuracy, Reliability, and Completeness) will only be asked to providers. Patients, in contrast, will be asked at the end whether they noticed any inaccuracies in the message.