Model of Sensor Error from Multiple Simultaneous Continuous Glucose Monitors in the Same Subject

Objective: A model of the sensor error is crucial to design reliable simulation scenarios for testing algorithms relying on continuous glucose monitoring (CGM), e.g., artificial pancreas control algorithms. The aim of this study is to develop such a model by using multiple CGM recordings simultaneously measured on the same subject. Method: The database consists of 36 data sets collected in 19 adults with type 1 diabetes. Subjects have been recruited at the Oregon Health and Science University (Portland, OR) and admitted for two sessions of 9 h duration each. Glucose concentration was measured simultaneously by four different Dexcom SEVEN PLUS sensors at 5 min sampling. HemoCue Glucose 201 Analyzer plasma samples were collected in parallel every 15 min as reference. Plasma samples were fitted against each of the four CGM traces, exploiting individualized models of plasma-to-interstitium kinetics, calibration, and sensor drift. The four residuals profiles were compared in order to identify the different components of the sensor error. Result: The first identified component is the measurement error. In addition, a model error component was identified, which reflects CGM errors due to both plasma-to-interstitium physiological model and technology issues, e.g., calibration/drift. Both components are satisfactorily described by a low-order autoregressive model. Use in simulation of this sensor error model generates realistic profiles. Conclusion: The availability of multiple sensor data measured simultaneously in the same subject is key to develop a sensor error model. Our results show that, in addition to measurement error, an additional component is needed to account for physiological and technological uncertainties. This sensor error model will be incorporated in the Food and Drug Administration-accepted University of Virginia/Padova type 1 diabetes simulator.