Page 290 - Glucose Monitoring Devices

Closed-loop glycemic control algorithms 297

difference between the reference (or desired) glucose concentration and the
measured value of the glucose concentration [9,49,50]. The error between the refer-
ence and measured glucose is assessed as a proportional, integral, and derivative
term. The proportional term considers the current value of the error, the integral
term considers the sum of the errors over a past time window, and the derivative
term considers the rate of change in the current error from the previous error. The
three terms are multiplied by coefficients that adjust the contributions of the individ-
ual terms to the overall amount of insulin dosage to be infused. The three coefficients
are the adjustable parameters of the controller that can be tuned by practitioners to
render the PID controller more aggressive or more conservative. Although the
simplicity and ease of implementation of the PID controller make it an appealing
choice for a control algorithm, the unsophisticated structure limits its ability to effec-
tively control systems with multitudes of disturbances, time-varying delays, and
temporal dynamics. The PID control algorithm is employed in early versions of
AP systems.

Fuzzy logic control
Fuzzy logic and knowledge-based expert systems are developed for closed-loop
insulin delivery by manually encoding the experience and knowledge of practi-
tioners as a set of rules or by using learning mechanisms [15,24]. The most common
form of the fuzzy logic controller involves constructing the set of logic rules to be
evaluated with all the available information at each sampling time. The fuzzy logic
system computes the input through the fuzzy rules and the (input and output) mem-
bership functions developed on expert knowledge or through the observation of the
control actions taken by the practitioner. Challenges in developing and maintaining
the set of fuzzy logic rules are limitations, as a survey of experts may lead to a
diverse array of possible rules and uncertainties about fuzzy set membership func-
tions. Moreover, the set of rules may become overly expansive, rendering personal-
ization of the controller to individual subjects or adaption of the rules over time a
challenge. A large set of rules may also cause conflicting actions to arise among
the rules, which will require conflict resolution schemes to draw insulin-dosing
decisions.

Model predictive control
MPC is widely adopted in controlled drug-delivery applications as an effective
approach to deal with large multivariable constrained control problems. The prin-
cipal function of MPC is to choose the optimal control actions (i.e., insulin infusion
quantitates) by repeatedly solving a constrained optimization problem online that
minimizes a performance index (for instance, predicted glucose tracking error
from target glycemic set-point) over a finite prediction horizon with predictions
obtained using a dynamic glucose-insulin system model [16,35,39,51e53]. There-
fore the three main components of MPC are as follows: (i) a dynamic model of