An Unannounced Meal Detection Module for Artificial Pancreas Control Systems

Abstract

The emergence of real-time glucose sensors has prompted the development of closed-loop insulin delivery systems for type 1 diabetes patients, termed the artificial pancreas. The existing closed-loop systems rely on the user's input to provide meal insulin boluses. However, patients, particularly adolescents, sometimes forget to announce consumed meals to the system. The performance of closed-loop systems after an unannounced meal may be improved with the addition of a meal detection module to the closed-loop system. We have developed a novel meal detection algorithm that detects unannounced meals using glucose measurements and insulin data. The model-based detection algorithm continually estimates an internal patient state using a linear Kalman filter. A generalized likelihood ratio test (GLRT) statistic is computed to evaluate the consistency of the Kalman filter under the null hypothesis that all consumed meals have been announced. A threshold criterion is applied on the GLRT to distinguish if the observed glucose increase is due to an unannounced meal. Simulation results, based on nonlinear time-varying virtual patients and noisy glucose measurements, show a sensitivity of 93.23% and a false positive rate of 4.17%. Moreover, 108 hours (4 patients x 3 visits x 9 hours) of clinical data is used to demonstrate the safety and feasibility of the meal detection module. Four patients underwent a nine-hour three-way inpatient experiment where the lunch meal was not announced to the system. The algorithm successfully detected all unannounced meals within 35 [30]-[40] minutes, without any false positives.