1882-P: Attention Is All You Need: Temporal Transformer-Based Personalization of Insulin Delivery Parameters from CGM and Pump Data

Introduction and Objective: While Transformer architectures have revolutionized language processing, their application to physiological time-series remains underexplored. This pilot study investigates whether a Temporal Fusion Transformer (TFT) can autonomously extract personalized, invariant insulin sensitivity factors (ISF) and insulin-to-carbohydrate ratios (ICR) from Type 1 Diabetes (T1D) data without explicit physiological constraints or hand-tuned heuristics. Methods: We trained a TFT model on 2-4 years of retrospective Nightscout data (continuous glucose monitor, insulin, and carbohydrate records) from a pilot T1D subject. To interrogate the physiological relationships learned by the model, we implemented a synthetic validation pipeline, injecting synthetic insulin and carbohydrate impulses onto stable baselines at varying noise levels to measure predicted glucose excursions and reverse-engineer physiological parameters. We then evaluated this approach in a cohort of 5 additional T1D subjects (3 pediatric, 2 adult) to assess generalization. Results: For the pilot subject, the model converged on robust invariant values (ISF 92.8 ± 2.1 mg/dL/U; ICR 6.41 ± 0.15 g/U) once training noise exceeded a critical threshold (≥6.0). Low-noise baselines systematically underestimated sensitivity. In the 5-subject validation cohort, convergence was achieved for all individuals, with high parameter stability (mean variance 5.8% ± 3.9%). Extracted parameters differed substantially between subjects (e.g., ISF range 38.6-122.2 mg/dL/U), indicating subject-specific physiological learning rather than population averaging. Conclusion: Transformer-based attention mechanisms can learn individualized glucose-insulin dynamics directly from routine CGM and pump data. Synthetic probing of trained models enables stable extraction of personalized insulin delivery parameters, suggesting a pathway toward fully automated, data-driven pump setting personalization beyond static heuristic approaches. Disclosure S. Glassenberg: Employee; Current; Relevate Health.