Observer based plasma glucose prediction in type I diabetes

Recent years´ progress in the development of Continuous Glucose Monitors (CGM) has made rich well-sampled glucose data readily available. Reliable, frequent measurements are of outmost importance for the emerging closed-loop control of diabetic plasma glucose. However, these sensors do not measure the variable of primary interest - plasma glucose, but a delayed signal - the interstitial glucose. To overcome this difficulty this paper presents a novel model, merging a black-box model of the glucose dynamics together with a CGM sensor model. Using an observer the plasma glucose level is estimated and predicted. The outlined scheme was evaluated on one patient, with a significant sensor delay, from a clinical trial of the DIAdvisor European FP7-project. Using the raw signal from the CGM device together with meal and insulin infusion data predictions for 20, 40 and 60 min were produced for a breakfast meal. Results: RMSE of the prediction error was smaller than 26 mg/dl for validation data even for the longest prediction horizon and no points in the C/D/E zones in the pCGA evaluation. The model clearly outperformed the CGMS and the results indicate that the method could be used successfully.