Prediction of Dynamic Glycemic Trends Using Optimal State Estimation

Efficacious therapeutic regimens to treat type 1 diabetes mellitus require devices capable of continuous feedback control; recent advances in medical technology mean that such devices are now available. Any closed-loop controller would require a predictive aspect to avoid sluggish control related to delays in insulin action, or hypoglycemia from an overdose of insulin. Using clinical data and an adaptive version of the simple Bergman minimal model (Bergman et al., 1979), glycemic prediction was performed. Model parameters were estimated using clinical data. An augmented state Kalman filter was then used to estimate parameters dynamically. Predictive accuracy varied from subject to subject, with median R values for the best validation days of 80% for 30 minute predictions. Such techniques would be useful in a closed-loop control framework for adapting a glycemic controller to subject-based variations in insulin sensitivity.