Page 456 - Handbook of Biomechatronics

450 Graham Brooker

the integral component and use a PD controller in which the proportional
term is determined by the current glucose level, while the derivative term is
determined by its variation in the previous half hour or so. These control
algorithms are not very effective at providing a slow nocturnal response
as well as a more aggressive response after eating, and so also require the addi-
tion a feed forward component to accommodate regulation after meals
(Cobelli et al., 2009).
Model predictive control (MPC) provides the most effective approach to
glucose control to date. The main components of MPC are the model, the
cost function, and the constraints. The model is required to predict current
and future states as well as system outputs and variables. The actual algorithm
is not that important and can be linear, nonlinear, continuous, or discrete
time. The cost function, usually quadratic, measures the quality of the
closed-loop control and provides a penalty on future deviations from the
required glucose concentration out to a prediction horizon. Finally, there
may be constraints on the manipulated variables. For example, insulin flow
will always be greater than zero and less than some maximum. The principal
merit of MPC is that it reduces the control design problem to a sequence of
finite-horizon optimization problems that allow it to deal easily with
nonlinear dynamics. Fig. 29 shows schematically the operation of a MPC
control system (Cobelli et al., 2009).
In addition to the basic PID, PD, and MPC methods, a number of alter-
native control strategies have been investigated in the recent years. These
include nonlinear MPC and various neural network-based approaches
(Semizer et al., 2012) as well as fuzzy logic controllers (Mauseth et al.,
2013; DoyleIII et al., 2014).
Because the biggest risk of a closed-loop system is hypoglycemia caused
by an over delivery of insulin, intense exercise, or the consumption of alco-
hol, modern devices generally include safety systems. These can include a
low glucose prediction module or an insulin-on-board (IOB) calculator
(DoyleIII et al., 2014). A good example is the modular closed-loop control
structure described by Patek et al. (2012). The structure of the controller is
shown in Fig. 30. Three primary modules are defined of which the lowest
level one is the interface module (IM) that communicates with the glucose
monitor and the insulin pump as well as providing an external interface and a
data logging function. The real-time control module (RTCM) implements
one of the control algorithms discussed above. Finally, the continuous safety
module (CSM) monitors the patient’s state and authorizes insulin recom-
mendations that come to it form the RTCM. It can either use IOB estimates
or can involve the patient in the approval process.