Page 78 - Control Theory in Biomedical Engineering

Adaptive control of artificial pancreas systems for treatment of type 1 diabetes 65

information with estimators designed based on glucose-insulin dynamic
models (de Pereda et al., 2016; Eberle and Ament, 2011; Neatpisarnvanit
and Boston, 2002; Hajizadeh et al., 2017a, 2018c). In our previous work,
the design of adaptive-personalized PIC estimators that directly take into
account the intersubject and intrasubject variabilities in glucose-insulin
dynamics is investigated using different estimation techniques (Hajizadeh
et al., 2017a). Clinical experimental data from subjects with T1D were used
to analyze the accuracy and reliability of the PIC estimates (Hajizadeh et al.,
2017a, 2018c).
The control algorithm computes the optimum amount of insulin by con-
sidering safety constraints that avoid insulin overdosing. Model predictive
control (MPC) formulations are efficient techniques for AP systems as they
can handle multivariable complex systems with constraints (Hajizadeh et al.,
2019b, c; Hovorka et al., 2004; Clarke et al., 2009; Laguna Sanz et al., 2017;
Gondhalekar et al., 2016; Chakrabarty et al., 2018; Cameron et al., 2011;
Toffanin et al., 2013; Boiroux et al., 2018; El Fathi et al., 2018; Messori
et al., 2018). However, MPC performance is affected by different factors
such as accuracy of the model, formulation of the objective function, and
system constraints. For AP systems, these factors need to be defined appro-
priately for effective glycemic control. In this work, accurate time-varying
glycemic models are identified recursively for BGC predictions. The key
controller parameters including the controller set-point, the objective func-
tion weights, and the system constraints are defined in an adaptive way to
accommodate various situations faced daily by people with diabetes. Using
machine-learning techniques and patients’ historical data, these parameters
are also modified in advance for the anticipated periods of disturbance effects
such as meals and exercise (Hajizadeh et al., 2019a, d).
Motivated by the previous considerations, a personalized multivariable,
multimodule artificial pancreas (PMM-AP) system is proposed to effectively
control the BGC without manual user announcements for meals and exer-
cise. The proposed PMM-AP uses physiological signals from a wearable
device and estimates of unannounced meal effects (from recent glucose
and insulin data) and PIC in addition to glucose measurements. A general
flowchart of the proposed method is presented in Fig. 1. An adaptive-
personalized PIC estimator summarized in Section 2.1 generates estimates
of the PIC. To identify time-varying glycemic models, a recursive system
identification technique summarized in Section 2.2 is used to characterize
the time-varying glucose-insulin dynamics. Then, the identified models
are employed for the design of an adaptive-learning model predictive