The LibreAssist GenAI system was noninferior to expert dieticians in predicting meal glycemic impact, achieving a 62.9% vs 58.6% agreement rate on images (difference 4.3%; CI lower bound 2.5%).
Does a GenAI algorithm provide glycemic impact estimates of meals comparable to expert dieticians?
A generative AI system can predict the glycemic impact of meals from images and text with an agreement rate non-inferior to expert dieticians, suggesting utility for scalable, real-time meal feedback in diabetes care.
Estimación del efecto: Difference 4.3% (95% CI lower bound 2.5%)
Tasa de eventos absoluta: 62.9% vs 58.6%
Introduction and Objective: There are about 40.1 million people living with diabetes in the United States and approximately 19,200 CDCES, or roughly 1 CDCES for every 2000 individuals with diabetes. LibreAssist is a generative AI (GenAI) feature in the Libre app that predicts the glycemic impact of meals as minor, moderate, or major based on user-input images or text descriptions, offering potential to enhance diabetes care by providing timely, personalized meal guidance that complements CDCES-delivered education at scale. This study evaluates whether the GenAI provides glycemic impact estimates comparable to dieticians’ assessments. Methods: The validation dataset includes 988 meal images and 200 text descriptions collected from a pilot study of people with diabetes using an early system prototype, supplemented with additional images to ensure the dataset represents a typical American diet. Each meal was annotated by 5 experts with CDCES and Registered Dietician certifications from a panel of 34, who were blinded to the algorithm predictions and expert evaluations. Performance was assessed using a Leave-One-Out (LOO) Agreement Rate, calculated by removing one annotator and comparing both the algorithm and the held-out annotator to the remaining four annotators. This process is repeated for all five annotators, producing an average agreement rate for the algorithm and annotators. Statistical non-inferiority was tested using a 10% margin for images and 20% for text. Results: On images (N=988), the algorithm achieved a 62.9% LOO Agreement Rate versus 58.6% for the annotators (difference: 4.3%; CI lower bound: 2.5%), demonstrating non-inferiority. On text descriptions (N=200), the algorithm achieved 68.1% versus 63.9% for annotators (difference 4.2%; CI lower bound 0.5%), also meeting the non-inferiority margin. Conclusion: These findings demonstrate that LibreAssist provides glycemic impact predictions comparable to expert dieticians, supporting its role in enhancing CDCES care by providing real-time individualized meal feedback at scale. Disclosure S. Vaughan: Consultant; Current; Abbott Diabetes. S. Gupta: Employee; Current; Abbott Diabetes. J.C. Nishida-Boucher: Employee; Current; Abbott Diabetes. B. Olson: Employee; Current; Abbott Diabetes.
VAUGHAN et al. (Fri,) conducted a other in Diabetes (n=1,188). LibreAssist GenAI glycemic impact prediction system vs. Expert dieticians was evaluated on Leave-One-Out (LOO) Agreement Rate on meal images (Difference 4.3%, 95% CI lower bound 2.5%). The LibreAssist GenAI system was noninferior to expert dieticians in predicting meal glycemic impact, achieving a 62.9% vs 58.6% agreement rate on images (difference 4.3%; CI lower bound 2.5%).