Analysis and Development of a Meal Detection Algorithm for the Artificial Pancreas
Authors:Eyal Dassau, Rebecca Harvey, Sensika Niyathapala
Mentor:Frank Doyle, Professor, Mellichamp Chair in Process Control , University of California Santa Barbara
Type 1 diabetes is a disease where no insulin is produced in the body, making it necessary to inject insulin to regulate blood glucose (BG) and avoid complications from high or low BG. The artificial pancreas is a device that will automatically control BG, consisting of a glucose sensor, a controller that calculates the amount of insulin to deliver, and an insulin pump. A drawback of the artificial pancreas is the delay in sensing BG (10 min) and in insulin absorption and action (45 min). By identifying meals with sensor data, the controller can address the changes in BG that occur after a meal to maintain desired blood glucose levels. A basic meal detection algorithm to identify meals when glucose is rising was designed and tested. Data from clinical trials were used in silico to evaluate the algorithm. One parameter was the number of subsequent detections required (SDR), which mandated that one to three positive detections occur in a row in order to flag a meal. After applying the algorithm to the clinical data, the percentage of meals detected using SDRs of one, two and three was 80%, 70%, and 27%, respectively. Although we predicted the algorithm with an SDR of one to flag more often when no meal was given there were none detected before the meal. The average time between the meal and the algorithm detection with an SDR of one, two and three was 34, 39, and 48 minutes, respectively. The one detection algorithm is better at detecting meals earlier and is the most effective; however with a larger dataset it can be more sensitive to detecting nonexistent meals. The algorithm overall is able to detect a meal and with further fine tuning, the gap between meal consumption and detection can diminish resulting in safer BG levels.