2009-P: Expert Preference for AI-Generated Diabetes Treatment Strategies: A Blinded Comparison of Simulation-Based Recommendations and Real-World Prescribing

Introduction and Objective: The expanding therapeutic landscape for T2D, characterized by diverse efficacy, safety, and cost profiles, has complicated treatment selection during time-constrained clinical encounters. We evaluated an AI-driven clinical decision support tool (CDS) that leverages electronic health record (EHR) data and microsimulation to generate automated, individualized treatment recommendations. Methods: Study included 100 randomly selected NIH All of Us participants with T2D who had a 2023 index encounter. Using baseline EHR data, the AI algorithm assessed treatment intensification needs across glycemic, blood pressure, lipid, and weight domains and selected therapies from 18 pharmacotherapeutic classes to maximize projected 5-year Quality-Adjusted Life Years. AI-driven recommendations were evaluated against real-world clinical decisions documented in the EHR following the index encounter. Three independent experts an endocrinologist, a clinical pharmacist, and a health economist, blinded to the source of each therapeutic option, adjudicated 100 paired cases and selected their preferred treatment. Results: The study cohort had a mean age of 67.1 years, 40.0% female, mean BMI of 33.0, and 41.6% had established cardiovascular disease. Agreement between model-recommended and EHR-prescribed drug classes was low (Cohen’s κ = −0.01; 2%), suggesting that the model often proposed alternative treatment options. Agreement was defined as an exact match in drug class between model and EHR recommendations. In blinded expert review, the AI-recommended treatment was preferred in 71% of cases (p 0.001). Conclusion: Although agreement with real-world prescribing was low, expert reviewers more frequently preferred AI-recommended therapies, indicating that the model often proposed clinically reasonable alternatives to routine care. These findings suggest that AI-driven simulation can serve as a powerful catalyst for advancing precision medicine. Disclosure Z. Li: None. P. Li: None. Y. Shao: None. F.J. Pasquel: Research Support; Current; Dexcom, Inc., Insulet Corporation, Ideal Medical Technologies. Research Support; Ended; Novo Nordisk, Tandem Diabetes Care, Inc. Consultant; Ended; Insulet Corporation. M.K. Ali: Consultant; Ended; Eli Lilly and Company, Siemens, Novo Nordisk. K. Narayan: None. Q. Xue: None. V. Fonseca: Consultant; Current; Abbott Diabetes. Stock/Shareholder; Current; BRAVO4HEALTH, LLC. Consultant; Ended; Bayer AG. Consultant; Current; Eli Lilly and Company, Corcept Therapeutics, Regeneron Pharmaceuticals Inc., Boehringer Ingelheim International GmbH. Stock/Shareholder; Current; Vertex Pharmaceuticals Incorporated. L. Shi: None. H. Shao: None.