1944-P: Clinical Cognition in Code: Evaluating a Generative Framework for Automated AGP Interpretation

Introduction and Objective: Ambulatory Glucose Profile (AGP) reports are central to diabetes management, yet their interpretation requires integration of glycemic patterns, pharmacologic effects, and behavioral factors. Most analytic tools focus on prediction rather than clinical reasoning. This study evaluated whether a generative framework could replicate clinician-level reasoning for AGP interpretation and therapeutic decision support. Methods: A generative framework integrating a language model with an encoded pharmacologic reasoning layer was developed to analyze AGP data. The system incorporated insulin and medication pharmacodynamics, timing effects, and guideline-based logic. Performance was evaluated using curated clinical scenarios, including nocturnal hypoglycemia, postprandial hyperglycemia, and rebound patterns. Outputs were compared with consensus interpretations from a blinded panel of board-certified endocrinologists. Primary outcomes were interpretive accuracy and time to actionable insight. Results: The framework showed strong agreement with expert interpretation, with a Cohen’s kappa of 0.81 for diagnostic classification. Medication-related versus behavior-related glucose variability was correctly distinguished in 92% of cases. Structured therapeutic recommendations were generated in under 10 seconds per case, compared with 4-6 minutes for expert review. Simulated insulin dose adjustments aligned with expert-recommended changes based on pharmacodynamic timing and glucose trends. Conclusion: A generative clinical reasoning framework can accurately interpret AGP data and deliver clinician-aligned therapeutic insights with substantial time efficiency. Embedding pharmacologic and physiologic logic into generative systems may support scalable decision support and medical education in diabetes care.