A novel meal detection algorithm for an artificial pancreas

Background: Many studies have reported that a high number of missed meal boluses occur, especially in adolescents during insulin pump therapy. It is predicted that this behavior will carry over to artificial pancreas therapy and therefore, a means to reduce poor outcomes due to unannounced meals must be implemented. Objective: To implement an algorithm to detect meals using data from a continuous glucose monitor (CGM) and the insulin delivered to the subject. Methodology: An Unscented Kalman Filter is employed to predict the states of a composite Bergman-Hovorka model altered to include an auxiliary disturbance parameter. Then, an algorithm checks the cross-correlation between the disturbance parameter and continuous glucose monitor levels and employs a threshold to detect an abnormal event. At this time point, a positive disturbance parameter value indicates a rise in glucose due to a meal. This methodology was tested using meals simulated in silico with 10 adult patients over a period of ten days (30 meals per subject). Results: Carbohydrate amounts tested were 85 ± 17 g. The true positive rate (sensitivity) and false positive rate were 90 ± 12.7% and 5 ± 1.8%, respectively. The accuracy and specificity were 94 ± 1.5%, and 95 ± 1.8%, respectively. The change in glucose experienced at detection was 19 ± 5.2 mg/dl and the detection time was 28 ± 2.8 min. Conclusion: This algorithm provides meal detection with minimal change in blood glu