Model predictive control to account for prolonged changes in insulin requirements following exercise in type 1 diabetes

Abstract

Automated insulin delivery systems are becoming increasingly available in the treatment of type 1 diabetes. They can improve glycemic outcomes while reducing patient burden, but good glycemic control during and after exercise remains challenging. Exercise causes substantial and prolonged changes in insulin sensitivity that consequently affect insulin requirements, and can lead to late-onset hypoglycemia if not accounted for.Here, we present a model predictive control algorithm that adjusts insulin delivery during the recovery period to improve glycemic outcomes after aerobic exercise. The algorithm continuously estimates insulin sensitivity from glucose measurements via an unscented Kalman filter. It integrates the estimate by continuously updating the target insulin input as well as the process model to account for changing insulin demands. The proposed approach is generic and transferable to other control formulations.We evaluate our new control strategy in-silico using a validated diabetes patient model with aerobic exercise. We consider a virtual patient population in full-day simulations including a wide variety of exercise scenarios covering moderate to high intensities and different timing and duration of the exercise. We demonstrate improved glycemic outcomes over night following a day with exercise for all scenarios and show robustness of our approach to common disturbances.