Diagnosis of Insulin Pump Failures Based on Physical Informed Blood Glucose Prediction

Insulin pump system faults can lead to abnormal hyperglycemia in patients receiving intensive insulin therapy, potentially resulting in life-threatening complications such as diabetic ketoacidosis. However, current insulin pumps operate in an open-loop mode and lack the capability to dynamically monitor actual insulin infusion rates, making timely fault detection challenging. To address this problem, we propose a fault detection method for insulin pump systems based on the residuals between glucose prediction and monitoring. The innovations of this work are twofold: (1) To account for the high uncertainty in glucose fluctuations and the scarcity of personalized data, we develop a physics-informed neural network (PINN) as a fault-free glucose prediction model. (2)We design a neural network based fault diagnosis model that leverages the residuals between predicted and measured glucose values, along with meal intake information, to detect insulin pump faults. The effectiveness of the proposed method is validated using the UVa/Padova Type 1 Diabetes Simulator, certified by the U.S. FDA. This approach provides a promising direction for enhancing the safety and reliability of insulin pump therapy through non-invasive, data-driven fault detection.