Book Description

Type 1 Diabetes Mellitus or Juvenile Onset Diabetes is currently a permanent, incurable disease that removes the ability of the patient's body to control blood glucose levels. This loss of automatic control greatly increases the patient's exposure to clinical risks of high and low blood glucose levels. These risks can be mitigated through tight, regulation of blood glucose levels using insulin injections, but only at the price of paying frequent attention to the blood glucose levels and manually providing accurate dosing decisions. This can be very trying for all patients, especially teenagers and children. Recent technological advances enable automatic external regulation of patient's blood glucose levels. Pumps can infuse insulin into the subcutaneous tissue to lower blood glucose levels. Continuous glucose monitors can sense subcutaneous glucose levels, specifically the rises caused by meals and the drops caused by insulin. This has caused a flurry of control and modeling research, in the hopes of mitigating the clinical risk without the price of constant human attention. The most common approach, and the one taken here, is to use model predictive control, where the predictions from a model of glucose dynamics are optimized against a cost function using the future insulin injections. We directly minimize the asymmetric clinical risk instead, and recognize that our control authority (the potential effects of injecting insulin) is largely limited to reducing the blood glucose level. We further consider likely future blood glucose measurements, since we both respond better to positive disturbances than negative ones, and because negative disturbances are more risky. Also, we explicitly estimate the uncertainty of predictions, since glucose dynamics incorporate uncertainty from the complex biology, stochastic patient behaviour, and extrapolation. More uncertainty should mean more cautious insulin injection. Lastly, since meals occur faster than insulin acts and can raise the blood glucose by 2 to 4 times the width of the acceptable range, this work develops a novel Bayesian framework for detecting meals and estimating their effects. This work improves prediction root mean squared error by 20% relative to predictions excluding meals for prediction horizons from 1 to 4 hours and improves robustness to meals. These prediction improvements alone reduce the avoidable clinical risk by 38% relative to predictions excluding meals. When the improvements to the predictions are combined with minimizing clinical risk under uncertainty and measurement anticipation the avoidable clinical risk is reduced by 30% relative to a published MPC controller that has privileged information and tunes independently for each patient.