Predicting Postprandial Glucose Excursions Using Gaussian Process Regression

Abstract

In recent years, continuous glucose measurement (CGM) devices have increased the quantity of data available to a patient and care team by an order of magnitude. We believe that a thorough evaluation of the control strategy of a patient or artificial pancreas can only occur when blood glucose measurements can be placed in the context of the patient's behaviors. Exciting results from the Juvenile Diabetes Research Foundation CGM study demonstrate a reduction in hemoglobin A1c when using CGM in intensive treatment in adults,1 but we believe that the potential value of this technology has yet to be reached. Augmenting CGM data with lifestyle data can also help address the data overload problem.2 In this letter, we present initial results using Gaussian process regression to predict 2-hour postprandial blood glucose using patient behavior data. The intelligent diabetes assistant (IDA) is a telemedicine diabetes management system composed of three parts: data collection equipment used by the patient, a database, and a Web interface for the care team. Because the system is designed to capture lifestyle data along with glucose measurements, it is applicable for assisting with the management of type 1 and type 2 diabetes. The IDA system was designed to measure a patient's behavior and learn models that link actions to some outcome. A custom data collection application that runs on a mobile phone is used to collect blood glucose measurements, images of meals, time and dose of medications, and messages. Figure 1 displays screen shots of the glucose, meal image, and insulin injection collection applications. Figure 2 displays a screen shot of the meal image nutrition analysis Web interface. IDA also collects continuous estimates of energy expenditure using the SenseWear Pro® armband from BodyMedia. Figure 1. Screen shots of glucose, meal image, and insulin injection collection applications. Figure 2. A screen shot of the meal image nutrition analysis Web interface. The primary focus for diabetes modeling in recent years has been for model predictive control in an artificial pancreas system.3,4 IDA can aid in the development of an artificial pancreas through controller performance monitoring using its telemedicine capabilities; in this case, learning a model to predict potential controller problems using behavior data.