A Data-Driven Behavior Modeling and Analysis Framework for Diabetic Patients on Insulin Pumps

About 30%-40% of Type 1 Diabetes (T1D) patients in the United States use insulin pumps. Current insulin infusion systems require users to manually input meal carb count and approve or modify the system-suggested meal insulin dose. Users can give correction insulin boluses at any time. Since meal carbohydrates and insulin are the two main driving forces of the glucose physiology, the user-specific eating and pump-using behavior has a great impact on the quality of glycemic control. In this paper, we propose an "Eat, Trust, and Correct" (ETC) framework to model the T1D insulin pump users' behavior. We use machine learning techniques to analyze the user behavior from a clinical dataset that we collected on 55 T1D patients who use insulin pumps. We demonstrate the usefulness of the ETC behavior modeling framework by performing in silico experiments. To this end, we integrate the user behavior model with an individually parameterized glucose physiological model, and perform probabilistic model checking on the user-in-the-loop system. The experimental results show that switching behavior types can significantly improve a patient's glycemic control outcomes. These analysis results can boost the effectiveness of T1D patient education and peer support.