Page 457 - Handbook of Biomechatronics

The Artificial Pancreas 451

Fig. 29 The MPC prediction scheme relies on past inputs and outputs with future
outputs predicted as a function of future inputs within the control horizon. (Based on
Cobelli, C., DallaMan, C., Sparacino, G., Magni, L., DeNicolao, G., Kovatchev, B., 2009.
Diabetes: models, signals and control. IEEE Rev. Biomed. Eng. 2: 54–96.)

5.3 The Future of Automated Insulin Delivery
As discussed earlier, managing diabetes without hypoglycemia is complicated
by the wide fluctuation of insulin requirements between people and even for
the same person from day to day and under different circumstances. These
fluctuations are driven by a host of factors including meal size, activity level,
illness, sleep deprivation, emotional stress, and menstrual cycle. Any success in
effectively controlling levels is driven primarily by the accuracy and reliability
of CGM systems and effective modeling of individual responses.
CGM accuracy has improved over the years due to improved filtering and
denoising, a reduction in biofouling and enzyme degradation as well as
numerous proprietary techniques used by various manufacturers.
A consensus guideline relating to the best way to assess CGM accuracy is avail-
able but it has not been agreed upon by researchers. However, mean absolute
relative difference (MARD) is quite common, and in conjunction with the
percentage of readings exceeding 20% error is a good measure of CGM per-
formance. A threshold below a 10% MARD and below a 12% probability of
readings exceeding 20% error has is considered to be the sweet spot beyond