WorksEngineering
Safe Data-Driven Insulin Dosage Decision Based on Physiology-Constrained Gaussian Process Learning
Abstract
Multiple daily injections (MDI) therapy is widely adopted by insulin-treated subjects with diabetes mellitus. In this work, we study how to ensure that data-driven model learning is aligned with the underlying physiological principles, and introduce a safe data-driven learning and decision framework for the MDI therapy. To provide credible glucose predictions for the decision, a physiology-constrained Gaussian process learning method is developed for glycemic dynamics modeling by encoding the physiological priors as constraints in the model training. Considering the natural day-to-day cycle of insulin decisions, a prediction-based optimization formulation with an objective function that has a penalty term associated with the previous dosage is employed for the insulin therapy, together with the asymmetric penalties on hyper- and hypoglycemia for safety. The effectiveness of the proposed method is verified using the 10-adult cohort from the FDA-accepted UVA/Padova T1DM simulator, and an advisory mode analysis is also conducted based on the clinical data from 40 type 1 diabetes and 55 type 2 diabetes subjects to indicate the safe and physiologically plausible decisions in the real-life applications.