background—all-free-download.

com
Insulin Estimation
and Prediction
a Review of the Estimation and Prediction of Subcutaneous
Insulin Pharmacokinetics in ­Closed-Loop Glucose Control

Jorge Bondia, Sergio Romero-Vivó,

Beatriz Ricarte, and José Luis Díez

P
atients affected by type 1 diabetes (T1D) suffer uous glucose monitoring (CGM) technology, a significant
from lack of insulin secretion by the pancreas, amount of research has been focused in the last decade
due to the autoimmune destruction of the insulin- on the development of an artificial pancreas (AP), that is,
producing beta cells. This translates into elevated a closed-loop glucose control system that automatically
plasma glucose concentration with deleterious dispenses insulin. The AP is now closer to reality [1]. An
effects in the cardiovascular system, leading to serious AP system was first approved by the U.S. Food and Drug
long-term complications. With the improvement of contin- Administration (FDA) in 2016 [2]. The system still requires
user intervention at mealtime, as currently done in open-
Digital Object Identifier 10.1109/MCS.2017.2766312 loop insulin pump therapy, to administer an insulin bolus
Date of publication: 19 January 2018 based on carbohydrate counting (a hybrid closed-loop system).

1066-033X/18©2018ieee FEBRUARY 2018  «  IEEE CONTROL SYSTEMS MAGAZINE  47

A first-generation AP is thus expected to be commercially delivery that may lead to late hypoglycemia, thereby risking
available soon. However, issues still need to be addressed the patient’s safety. Regardless of how these mechanisms
for the performance improvement of the AP, such as are incorporated into the control schemes, all of them rely
inter- and intrasubject physiological variability, efficient on pharmacokinetic models to predict either circulating
compensation of real-life disturbances (such as exercise, plasma insulin or a measure of insulin on board (IOB), such
alcohol, and diseases), pump faults, sensor accuracy, and as the insulin depot remaining at the subcutaneous tissue
large dynamic lags induced by the subcutaneous adminis- before entering circulation. In this article, methods to con-
tration of insulin. This last challenge constitutes the focus strain insulin delivery are reviewed as well as the subcuta-
of this article. neous insulin pharmacokinetic models and estimatorson
Insulin is a glucose-lowering hormone that promotes which they can rely.
glucose transport through the cell membrane for its A significant challenge in the prediction of physiologi-
consumption or storage. Glucose regulation is completed cal signals such as insulin concentration is the large intra-
in the human body through counterregulatory hor- subject variability that patients suffer. Indeed, in terms of
mones, such as glucagon that promotes glucose production control engineering, a patient is a highly time-varying
by the liver, resulting in an increase of blood glucose con- uncertain plant. The intraday and day-to-day patient’s
centration. However, glucagon secretion by the alpha cells behavior change due to circadian rhythms (24-h rhythmic
in the pancreas is also affected as the disease progresses physiological oscillations driven by the body clock, for
along time, increasing the risk of suffering hypoglycemia. instance, daily patterns in insulin sensitivity) and other mul-
Severe hypoglycemia can provoke coma and death. tiple sources of uncertainty arise in key physiological pro-
The unidirectional effect of insulin has major implica- cesses such as meal absorption and subcutaneous insulin
tions in the design of an AP. This effect is aggravated by the absorption. Despite this high variability and uncertainty,
significant lag introduced by subcutaneous insulin infu- the use of population models for the prediction of insulin
sion, which is currently the safest infusion route for a com- pharmacokinetics (that is, how the infused insulin appears
mercial AP, as opposed to the endogenous pancreas that in blood) is still common practice.
secretes insulin directly into portal circulation (the vessels The impact of variability on the model prediction and its
that connect the pancreas and other organs with the liver, implication in closed-loop performance is analyzed. Never-
which acts as a first blood filter before entering the main theless, large intrasubject variability suggests that real-
circulatory system). Even in approaches with concomitant time state and pharmacokinetic parameters estimation is
infusion of glucagon (the so-called dual-hormone AP), convenient, even when individualized models are consid-
mechanisms are necessary to avoid an excess of insulin ered. The availability of continuous glucose measurements
allows for addressing this problem, should an observable
glucose-insulin model be available. This article reviews
and discusses different proposed observer techniques. See
Summary “Summary” for an overview of the entire article.

A n AP is a closed-loop control system designed for auto-

matic insulin delivery in T1D patients who do not pro- The Subcutaneous Insulin Route
duce insulin. The lack of insulin secretion is replaced by The pancreas secretes insulin into the portal vein toward
exogenous insulin infusion driven by the controller. This ex- the liver, which acts as a first filter before insulin reaches
ogenous insulin is usually infused beneath the skin, which systemic circulation. In the liver, insulin promotes glucose
entails a significant lag bet­­ween its injection and peak storage in hepatic cells, decreasing glucose production. In
effect. This lag poses an important challenge to glucose the fat and muscle cells, insulin acts as a key that triggers
control because the controller cannot easily counteract the mobilization of glucose transporters to the cell mem-
the effect of injecting too much insulin. Thus, methods brane, promoting glucose uptake by the cell. Plasma glu-
are needed to limit insulin delivery based on some mea- cose concentration decreases as a result of both actions. The
sure of the insulin in the patient, as predicted by suitable dynamic lag of insulin action is estimated to be approxi-
pharmacokinetic models. Several methods of addressing mately 30 min [3].
this question using estimators are presented in the ar- An AP is a classic closed-loop glucose control system
ticle. The high variability present in key physiological pro- (see Figure 1) that automatically dispenses insulin to
cesses presents several challenges in obtaining accurate a patient (the process), and has three main components: the
estimates of the insulin level. The impact of this variabil- continuous glucose monitor (the sensor), the insulin infu-
ity on the model prediction and the design of the insulin sion pump (the actuator), and the control algorithm (the
observers is also analyzed in depth. Finally, several open controller). Contrary to the pancreas, an AP cannot have
challenges for limiting the insulin infusion in free-living con- direct access to the hepatic portal circulation for insulin
ditions are also reviewed. infusion. Implantable insulin pumps that infuse insulin
into the peritoneum (that is, the body cavity containing the

48  IEEE CONTROL SYSTEMS MAGAZINE  »  FEBRUARY 2018

abdominal organs) were proposed [4], in an attempt to transport between plasma and the interstitial fluid introduce
closely emulate the pancreas. However, external insulin a measurement lag estimated in approximately 10 min [3].
pumps that infuse insulin through the skin (the so-called The lag induced by the subcutaneous route sums to 90 min,
subcutaneous route) are the only ones that are currently since insulin is infused until its peak effect (50 min for insu-
considered feasible for a commercial AP, due to their lin absorption, 30 min for insulin action, and 10 min for glu-
minimal invasiveness. External insulin pumps were first cose measurement). This significant dynamic lag poses an
introduced in the 1980s and are now a well-established, important challenge to glucose control, especially because
open-loop insulin therapy [5]. Traditional pumps infuse once insulin is infused, its effect cannot be easily counter-
insulin through a small, flexible tube (a catheter) with a acted by the controller, unless the patient eats carbohydrates.
needle allocated into the fat layer under the dermis (the Only dual-hormone AP systems can counteract an insulin
subcutaneous adipose tissue). Patch pumps with no visible excess with glucagon infusion at the expense of much higher
catheter are also available [6]. system complexity. However, glucagon administration must
Insulin pumps use fast-acting insulin analogs (such as also be limited since its excess may produce side effects, such
insulin aspart [7] or insulin lispro [8]), which are synthetic as nausea and vomiting [11]. Thus, mechanisms are neces-
insulin molecules with some modification in the amino sary to avoid an excess of insulin delivery due to controller’s
acid chain (compared to human insulin) to speed up subcu- overactuation because it may lead to late hypoglycemia,
taneous absorption and obtain a faster insulin action. No thereby risking the patient’s safety. These mechanisms con-
significant differences in pharmacokinetics and metabolic stitute a critical component of any AP.
effects between insulin aspart and lispro are found [9]. Large intrasubject physiological variability is an addi-
Insulin molecules can be in the form of hexamers (groups tional important challenge for the AP and is responsible, at
of six molecules), dimers (groups of two molecules), or least in part, for the difficulties of achieving good glycemic
monomers (a single molecule). Once infused at the subcuta- control. Currently, patients still have an average exposure to
neous tissue, the generated insulin depot starts diffusing hypoglycemia (< 70 mg/dL) of over 1 h/day and to hyper-
from the infusion site while dissociation of hexamers into glycemia (> 180 mg/dL) of higher than 9 h/day [12]. The
dimers and monomers occurs. The molecule structure of sources of physiological variability are not fully under-
insulin analogs is designed to facilitate this dissociation stood, although variability of subcutaneous insulin absorp-
since only dimers and monomers are small enough to go tion [13], [14] and changes in insulin sensitivity (due to
through the capillary wall and thus be absorbed toward circadian rhythms [15] and premenstrual periods in women
circulation [10]. The lag introduced by subcutaneous absorp- [16]) seem to play a major role. Meal ingestion also has a
tion is estimated to be approximately 50 min [3]. highly variable effect on glucose homeostasis [17].
Subcutaneous insulin pumps are combined with sub- An intrasubject variability of 27% was reported for
cutaneous, needle-type, continuous glucose monitors that time-to-peak plasma insulin concentration [parameter t maxI
compute plasma glucose values from electrochemical in model (26)–(28)] in a pharmacokinetic study of insu-
measurements of interstitial glucose. Dynamics of glucose lin aspart in subjects with T1D [13]. Nearly 40% of this

Meals, Alcohol, Exercise, Disturbances

Artificial Pancreas
Stress, Diseases, …
Control System for Automatic Insulin Delivery
Patient

Target Deviation Control Insulin Insulin Glucose

Glucose from Algorithm Infusion Insulin Absorption Insulin Homeostasis
Level + Target Pump Infusion and Action Action
Reference Error Controller Actuator Actuator Lag Control Glucose
– Action Action Subprocess

Plasma Continuous Plasma to Interstitial

Glucose Glucose Interstitial Fluid Glucose Plasma
Estimation Monitor Glucose Transport Glucose
Estimated Sensor Measured Lag Controlled
Variable Value Variable Variable
Process

Figure 1  The artificial pancreas as a classic, closed-loop, glucose control system (medical vocabulary in bold and control engineering
vocabulary in italics).

FEBRUARY 2018  «  IEEE CONTROL SYSTEMS MAGAZINE  49

 frequent in T1D [18]. Compared to normal adipose tissue,
Why Is Insulin Estimation and Prediction lypohypertrophia substantially increased intrasubject
Needed in an Artificial Pancreas? variability both in pharmacokinetic and pharmacody-

T o keep glucose levels of patients affected by T1D in a namic parameters. The coefficient of variance (CV) of the
safe range, the AP can be used as a closed-loop control area-under-the-curve of plasma insulin after an insulin
system that automatically dispenses insulin. However, the injection in lypohypertrophic adipose tissue ranged from
use of insulin as control action has different limitations: 55 to 65% (compared to 11–20% for normal adipose tissue),
• Insulin has a unidirectional effect lowering glucose values. corresponding to a three- to five-fold increment. No effect
Any insulin excess cannot be compensated beyond pump of lypohypertrophia on the time-to-peak plasma insulin
switch off, and external actions are required, such as meal was found. Intrasubject variability of the insulin effect
consumption or glucagon administration. (measured as the area-under-the-curve of the exogenous
• Insulin is delivered subcutaneously, as opposed to the glucose infusion needed to keep plasma glucose constant
pancreas, which delivers insulin into blood. The lag in- after insulin injection) was generally higher, especially
troduced by subcutaneous insulin absorption amounts during the first hour when the CV reached 90% (compared
to 50 min. The control action peak is reached after to 66% for normal adipose tissue).
another 30 min, with a total dynamic lag of 80 min. An Although the above study was conducted with insulin
excessive insulin infusion might later provoke hypoglyce- injections and not an insulin pump, it highlights the challenge
mia. Severe hypoglycemia can lead to coma and death. that intrasubject variability poses for any insulin therapy with
Therefore, mechanisms are necessary to avoid an excess of subcutaneous administration, including an AP. Current insu-
insulin within the patient’s body. They rely upon the prediction lin infusion sets in pumps need to be replaced every two to
of plasma insulin or IOB (the injected insulin that didn’t have three days to avoid inflammation and infection. Differences
an effect yet) from subcutaneous insulin pharmacokinetic in insulin absorption of ­consecutive infusion sites due to the
models. Control action is then modulated accordingly. How- development of lypohypertrophia can jeopardize the AP per-
ever, high variability of insulin pharmacokinetics suggests formance, increasing the risk of hypoglycemia. Besides, cur-
that real-time estimation of pharmacokinetic parameters and rent systems need two skin punctures, one for the sensor and
signals are also needed for better and safer performance. one for the insulin pump, making the rotation of the infusion
This can be addressed with the design of insulin observers. site more difficult [19]. (See “Why Is Insulin Estimation and
Prediction Needed in an Artificial Pancreas?” for a summary
of the key facts described in this section that justify the use of
insulin estimators and predictors.
Meal
Glucose-insulin Models
Carbohydrate Models describing the glucose regulation by insulin (hence-
Digestion and forth referred to as glucose-insulin models) are needed for
Absorption
the design of controllers and observers. The reader is
Concentration

referred to [20] and [21] for a historical perspective on mod-

Subcutaneous
Glucose

Insulin Glucose eling the glucose-insulin system, including models to mea-

Insulin Insulin
Action Metabolism sure key physiological signals or parameters in glucose
Absorption
metabolism (the so-called minimal models), fine-grained
models that provide a detailed description of the physiolog-
Figure 2  Model components for the glucose-insulin system. ical processes for simulation purposes (maximal models),
and models for control with a simpler structure capturing
variability was attributed to variations in depth of cannula the most relevant dynamics.
insertion, insulin site age, and local tissue perfusion. In Independent of the model complexity, the description of
contrast, insulin clearance was highly reproducible in the glucose metabolism in the context of any subcutaneous
same study. A recent study with insulin lispro also showed insulin therapy such as the AP implies the following model
an important impact of lypohypertrophia (the accumula- components (see Figure 2):
tion of abnormal mass of fat under the skin) of the insulin »» A subcutaneous insulin pharmacokinetic model describ-
injection site on insulin absorption and insulin effect [18]. ing how insulin appears in blood after subcutaneous
Lypohypertrophia appears when the same injection site is infusion. Although modeling of subcutaneous insulin
used repeatedly and is characterized by hypertrophic adi- pharmacokinetics dates back to the 1980s, only works
pocytes, reduced vascularization, and lower capillary den- related to the fast-insulin analogs currently used in
sity. Although patients are advised to rotate the injection insulin pump therapy are relevant here (insulin lispro
site, a prevalence of lypohypertrophia ranging from 28 to and aspart) as well as new insulin formulations under
over 64% is reported (depending on the country) as more investigation for ultrafast insulin absorption.

50  IEEE CONTROL SYSTEMS MAGAZINE  »  FEBRUARY 2018

 »» An insulin action model describing how plasma insulin index of insulin action (the so-called insulin sensitivity) is
concentration exerts its effect on glucose metabolism. computed by means of a parsimonious (minimal) model.
This component mainly involves the transport lags from The Bergman model is a second-order, nonlinear model
plasma to the insulin action site. relating plasma insulin concentration I (t) (input) and plasma
»» A carbohydrate digestion and absorption model describ- glucose concentration G (t) (output). The insulin effect X (t)
ing how glucose enters blood after a meal. This model is modeled with a first-order system representing a lagged
has proven to be challenging due to the complex physi- action of insulin. The plasma glucose concentration G (t) is
ology of gastric emptying and intestinal absorption inhibited by the glucose itself and insulin effect. A constant
and the lack of glucose absorption rate measurements, hepatic glucose production p 1 G b is considered that drives
unless complex clinical studies are conducted. the system to equilibrium for a basal insulin concentration
»» A glucose metabolism model describing the compre- I b . Finally, the term U G (t) /VG describes the intravenous
hensive effect of insulin and meals on plasma glu- glucose infusion (or glucose rate of appearance from the
cose concentration. More complex models may also meal when a meal model is interconnected). Parameters
include other effects like glucagon and exercise. p 1, p 2, and p 3 are kinetic parameters. Parameter p 1 is called
Three widely used models in the AP community, from “glucose effectiveness” and represents glucose autoinhibi-
lower to higher complexity, are the Bergman model (mini- tion. Insulin sensitivity is computed as p 3 /p 2 . The Bergman
mal) [22], Hovorka model (intermediate complexity) [23], model was conceived as a model for measurement, which is
and Dalla Man model (maximal) [24]. The latter is the core why no subcutaneous insulin absorption or meal ingestion
of the UVA-Padova T1D simulator [25], a simulation tool for is considered.
the evaluation of glucose controllers accepted by the FDA The Hovorka model [23], [27] (see Table 1) is a seventh-
as a substitute for animal trials. Since models for simula- order, nonlinear model that includes all model components in
tion are not in the scope of this article, it will not be pre- Figure 2. Carbohydrate digestion and absorption is described
sented here. by the impulse response of a second-order linear model, with
Table 1 presents the equations for the Bergman and Hov- input D G the amount of carbohydrates in the meal and output
orka models (left and center columns). The Bergman model of the glucose rate of appearance in blood, U G (t). Parameter
originated in 1979 [26] to describe the intravenous glucose A G is the carbohydrate bioavailability (that is, the fraction of
tolerance test (IVGTT), a clinical test to analyze how the carbohydrates in the meal that will finally reach blood), and
body metabolizes glucose. The IVGTT consists of the infu- t maxG is the time constant. Subcutaneous insulin pharmacoki-
sion of glucose intravenously and the measurement of netics is a third-order linear model. Subcutaneous absorption
plasma glucose and insulin concentrations for a certain time shares the same model structure as the meal model, with u (t)
before and after the infusion. From these measurements, an the insulin infusion from the pump and states S 1 (t) and S 2 (t)

Table 1 Low and intermediate complexity models widely used in the artificial pancreas field.

Identifiable Virtual Patient

Bergman Model [22] Hovorka Model [23] Model [28]

D G A G te -t/t maxG D G A G te -t/t maxG

Carbohydrate absorption N/A U G (t) = R A (t) =
t 2maxG VG t 2maxG
S 1 (t) I SC (t) u SC (t)
Subcutaneous (SC) N/A So 1 (t) = u SC (t) - IoSC (t) =- +
t maxI x1 x 1 K cl
insulin pharmacokinetics
oS 2 (t) = S 1 (t) - S 2 (t) oI (t) =- I (t) + I SC (t)
t maxI t maxI x2 x2
oI (t) = S 2 (t) - k e I (t)
t maxI VI

Insulin action Xo (t) =-p 2 X (t) + p 3 (I (t) - I b) Xo 1 (t) =-k a1 X 1 (t) + k b1 I (t) IoE (t) =-p 2 I E (t) + p 2 S I I (t)
Xo 2 (t) =-k a2 X 2 (t) + k b2 I (t)
Xo 3 (t) =-k a3 X 3 (t) + k b3 I (t)

Glucose metabolism Go (t) =-( p 1 + X (t)) G (t) Qo 1 (t) =-X 1 Q 1 (t) + k 12 Q 2 (t) - F 01 c
(t) Go (t) =- (GEZI + I E (t)) G (t)
U G (t) - FR (t) + U G (t) + EGP0 (1 - X 3 (t)) + EGP + R A (t)
+ p1 Gb +
VG Qo 2 (t) = X 1 Q 1 (t) - k 12 Q 2 (t) - X 2 (t) Q 2 (t)
Q 1 (t)
G (t) =
VG

FEBRUARY 2018  «  IEEE CONTROL SYSTEMS MAGAZINE  51

representing the mass of insulin along subcutaneous tissue, presented in Table 1 (right column). Carbohydrate absorp-
with transport time constant t maxI . The insulin rate of appear- tion is described as in the Hovorka model, although the
ance in blood is S 2 (t) /t maxI . Once this insulin rate is converted output is divided by the glucose distribution volume VG for
to the concentration rate by dividing by the insulin distribu- R A (t) to represent a concentration rate, instead of a mass
tion volume VI, it enters the first-order kinetic equation rate (R A (t) = U G (t) /VG) . Insulin action and glucose metab-
defining plasma insulin concentration I (t). Parameter k e is olism are equivalent to the Bergman model, with the differ-
the fractional elimination rate. Plasma insulin exerts three ence that plasma insulin I (t) is considered, instead of its
actions, following a first-order dynamic (representing a deviation with respect to basal insulin I b . I E (t) and the glu-
“remote” action): X 1 (t) is the effect of insulin on glucose dis- cose effectiveness extrapolated to zero insulin (GEZI) are
tribution/transport, X 2 (t) is the effect on glucose disposal (it thus equivalent to X (t) and p 1, respectively, with the
promotes glucose entering muscle and adipose tissue cells), change in meaning due to the above difference. EGP is
and X 3 (t) is the effect endogenous glucose production equivalent to p 1 G b and describes the hepatic glucose pro-
(it inhibits glucose production by the liver). Constants duction. The subcutaneous insulin pharmacokinetic model
k ai, i = 1, 2, 3 are kinetic constants and k bi /k ai are the insulin consists of a second-order linear model with time constants
sensitivities for each effect. Finally, states Q 1 (t) and Q 2 (t) rep- x 1 and x 2 . Parameter K cl is the insulin clearance, that is,
resent the glucose mass in plasma and interstitial tissue, how much volume of plasma is cleared of insulin per unit
respectively. The transport of glucose between both compart- time. Applying the Laplace transform, the corresponding
ments happens at rates k 12 and X 1 (t), depending on the direc- transfer function is
tion. Besides these transport flows, glucose from the meal 1 1
U G (t) and hepatic glucose production EGP0 (1 - X 3 (t)) (with I (s) x1 x2
= . (1)
K cl c s + 1 mc s + 1 m
EGP0 the production extrapolated to zero insulin) are input u SC (s)
x1 x2
flows to the plasma compartment. Insulin-independent con-
c
sumption (mainly by brain) F 01 (t) and renal glucose eli­­ The impulse response for an insulin bolus u SC (t) = U 0 d (t) is
mination FR (t) are output flows. See [23] for the detailed
c
de­­­­­fi­­­nitions of F 01 (t) and FR (t) . Aside from the intercompart- I (t) = U 0 1 ^ e -t/x 2 - e -t/x 1 h . (2)
K cl (x 2 - x 1)
ment transport, the glucose uptake by muscle and adipose
tissue X 2 (t) Q 2 (t) is an output flow. Plasma glucose concentra- This model is often referred to as a biexponential model
tion (model output) is finally given by Q 1 (t) /VG, where VG is [29], [30]. An alternative state-space representation of (1) is
the glucose distribution volume. the compartmental model
Due to its simplicity and identifiability properties, the
Bergman model is widely used outside its original scope, in dq 1 (t)
combination with a subcutaneous insulin pharmacokinetic = u SC (t) - 1 q 1 (t), (3)
dt x1
and a carbohydrate absorption model. An example is the dq 2 (t)
so-called identifiable virtual patient (IVP) model [28], also = q 1 (t) - 1 q 2 (t), (4)
1
dt x1 x2
q 2 (t ) 1/x 2
I (t) = = q (t), (5)
VI K cl 2
6
Hovorka Model where q 1 (t) is the mass of insulin at the subcutaneous
5 Biexponential Model tissue, q 2 (t) is the plasma insulin mass, and VI is the insu-
Plasma Insulin (mU/L)

lin volume of distribution, which is the ratio between the

4 insulin clearance (K cl) and insulin elimination rate con-
stant (1/x 2) . It becomes apparent from (3)–(5) that the biex-
3 ponential model corresponds to the Hovorka subcutaneous
insulin pharmacokinetic model, with the subcutaneous
2
space simplified to a single compartment and the parame-
ter equivalences x 1 = t max I and 1/x 2 = k e . Figure 3 shows a
1
comparison of the response of both models to an insulin
0 bolus of 1 U, considering published nominal parameters.
0 2 4 6 8
t (h) Insulin Observers
Independent of the control algorithm used in an AP, either
Figure 3  A comparison of the unit impulse response of the Hovo- single-hormone (only insulin) or dual-hormone (insulin
rka insulin pharmacokinetic model [23] and the biexponential
model [30] with nominal parameters (Hovorka: t maxI = 55 min,
and glucagon) mechanisms are necessary to avoid an
k e = 0.138 min -1, VI = 0.12 L/kg, body weight 70 kg; biexponen- excess of insulin delivery for safety reasons. These mecha-
tial: x 1 = 55 min, x 2 = 70 min, K cl = 1 L / min) . nisms must rely on subcutaneous insulin pharmacokinetic

52  IEEE CONTROL SYSTEMS MAGAZINE  »  FEBRUARY 2018

models (as the ones described in the previous section) for in subjects with T1D is addressed in the context of insulin
the estimation of either circulating plasma insulin or IOB. observers for closed-loop control. The extra equation
IOB represents the injected but still unused insulin that
IG (t) = 1 ^G (t) - IG (t) h, (6)
.
will have an effect in the future. Nevertheless, even with
x
the use of individualized models, the real-time estimation
of those signals is convenient for effective closed-loop glu- where IG (t) represents interstitial glucose (what the CGM
cose control because large intrasubject variability can com- measures) and x is the time constant for plasma-to-intersti-
promise performance (see “Models and Variability” for tial space glucose transport, is added to the Hovorka model
details). This problem could be addressed with the design to describe the lagged measurement by the CGM. Process
of insulin observers, in case an observable glucose-insulin and observation noises are also added. In addition, state
model is available. See “Observability of Nonlinear Sys- observers are designed to estimate uncertain pharmacoki-
tems” for a description of observability tests applicable to netic parameters, in particular the fractional elimination
glucose-insulin models in Table 1. An observer is a real- rate, k e, and time-to-maximum insulin absorption, t maxI
time estimator with a feedback mechanism for recursively (see Table 1), by building new extended models in which
correcting its estimated state based on the actual outputs these parameters are considered as new states with no
measured from the real physical system. Observers can dynamics. The observability of these models is demon-
also be used for the real-time estimation of model parame- strated analytically through Lie derivatives.
ters via a state extension. The Luenberger observer and dif- Plasma insulin estimations are validated first in a simula-
ferent extensions of the Kalman filter to nonlinear systems tion study, with a total of five meals and 25-h duration. A clini-
(see “Kalman Filter Extensions to Nonlinear Systems”) are cal validation was then conducted, using real data from 12
proposed in the literature. Basic knowledge of the Luen- patients with T1D who underwent four mixed meal studies
berger observer and linear Kalman filter is assumed. during 5 h, after a glucose normalization phase [36]. Fig­­ure 4
Depending on the modeling framework, the term “ob­­­ shows an illustrative example of the observer performance
server” (deterministic) or “filter” (stochastic) is used. For against clinical data for a sample patient. As a compar­­ator,
instance, the Luenberger observer strictly applies to deter- the use of a population model for insulin prediction is
ministic systems, whereas the Kalman filter is formulated
in a stochastic framework that incorporates models for
measurement and system noise, and it produces an explicit
ke (min−1) Glucose (mg/dL) Insulin (mU/L)

estimate for the covariance of the system state. 100

Observers are integrated into some AP systems for sev- 50
eral purposes. In [31], a Kalman filter is used in a model
0
predictive control (MPC)-based system to update two 0 50 100 150 200 250 300
model parameters: 1) a glucose flux quantifying model 200
mismatch and 2) carbohydrate bioavailability. A stochastic- 150
based approach incorporating competing models differing 100
in the rate of subcutaneous insulin absorption and action, 50
as well as carbohydrate absorption, are used to account for 0 50 100 150 200 250 300
intrasubject variability. In [32], a Luenberger observer is 1
used to estimate the initial state of the prediction model in
0.5
a zone MPC system [33]. In [34], its performance is com-
pared to a moving horizon state estimator (MHSE), which 0
0 50 100 150 200 250 300
is based on a constraint optimization problem that calcu- 300
tmaxI (min)

lates the optimal sequence of the process and the measure- 200
ment noises to minimize a cost function within a fixed history 100
horizon. It is shown to have a better performance than the 0
0 50 100 150 200 250 300
Luenberger observer in simulation, rejecting a meal distur-
t (min)
bance quicker without inducing hypoglycemia.
Despite the risk to the patient’s safety from plasma insu- Figure 4  An example of performance of the insulin observer in
lin or IOB estimation errors due to large intrasubject vari- [35] for a sample patient. The dashed red line represents the case
ability, the literature addressing the design of insulin with nominal pharmacokinetic parameters in the Hovorka model;
observers and, most importantly, assessing its performance the solid blue line represents the case with real-time adaptation of
with clinical data, is scarce. k e and t maxI, with initial values set to nominal values; and the solid
green line represents the case with real-time adaptation of k e and
The Hovorka glucose-insulin model [23] and the extended t maxI, with initial values set to the average of the parameter values
Kalman filter (EKF) are used in [35], where real-time estimation estimated for the rest of the patients. Green diamonds represent
of plasma insulin concentration from CGM measurements measurements (adapted from [35]).

FEBRUARY 2018  «  IEEE CONTROL SYSTEMS MAGAZINE  53

 Models and Variability

T he large intrasubject variability in the patient’s response

represents a big challenge for modeling since it may jeop-
ardize any identification process, giving rise to individualized y0
patient models representing only an average behavior with lim-
x0 ∅(y0,t ∗)
ited prediction accuracy. From the mathematical perspective,
variability entails uncertainty in model parameters and initial
conditions. Simulation tools for models with uncertainty can ∅(x0,t ∗)
thus be helpful to understand the impact of intrasubject vari-
ability in the accuracy of model predictions as well as imple- ∅(x0,t ) ∅(y0,t )
menting more robust algorithms for the AP.
A natural way to describe uncertainty in model parameters
Figure S1  The relation order in the flow of a monotone system.
and initial conditions is through the representation of values as
intervals, giving rise to “interval models” [S1]. The output of an
interval model is an envelope enclosing all possible glucose From an analytical point of view (see [S4]), the rate of
trajectories (for each possible value in the interval), instead of a change of the state vector x = (x 1, f , x n) T of a monotone sys-
single trajectory, as is the case with a standard set of differen- tem can be described as
tial equations. Envelopes should be efficiently computed with
mathematical guarantee, that is, they must include all possible dx = f (x, p, u (t)),  (S2)
dt
responses with the minimum possible overestimation. Interval
methods [S2] generally lead to divergent envelopes, and Mon- where the vector function f = (f1, f, fn) T satisfies the property
te Carlo simulation [S3] is not applicable since the inclusion
condition is not guaranteed. Other tools as the differential in- 2fi · 2f j $ 0, for all i ! j, t ! R + . 
(S3)
clusions can be efficient under certain cooperative conditions 2x j 2x i
(see [S4]), but unfortunately, they are not useful for glucose-
For cooperative systems [S7], the relation order is induced by
insulin models. Currently, new techniques such as differential
the corresponding positive orthant, which is characterized by a
inequalities and monotone dynamical systems (see [S5] and
Metzler Jacobian matrix, that is,
[S6] and the references therein) have been implemented. In
fact, monotone systems theory has been successfully used for
2fi $ 0, for all i ! j, t $ 0.  (S4)
the prediction of an envelope containing all the possible glyce- 2x j
mic responses since, in some cases, it is possible to calculate
the exact envelope for all possible behaviors [S7]. Monotonicity can also be characterized by a species graph,
The interconnection of monotone systems may be studied in which a node is assigned to each state, parameter, input,
by considering a flow x (t) = z (x 0, t). A system is monotone if and output, and a spin assignment is conducted for each node
based on the sign of partial derivatives (see [S5]). Moreover,
x 0 # y 0 & z (x 0, t) # z (y 0, t),  (S1)
the monotonicity of the system with respect to the parameters
for all t $ 0, where ≤ is a given relation order [S4], as repre- of the model can be analyzed by considering the parameters
sented in Figure S1. as system states in an extended model [S8].

considered as a comparator, since it is still a common practice compared in Figure 4: 1) a population model is used as an
despite the expected variability in insulin pharmacokinetics insulin predictor considering nominal pharmacokinetic
(see the section “The Subcutaneous Insulin Route”). With parameters in the Hovorka model, which are clearly detuned
regard to variability, different scenarios are devised: natural for this patient (comparator); 2) pharmacokinetic parameters
physiological variability in insulin pharmacokinetics at a k e and t maxI are included in the Kalman filter for real-time esti-
given infusion site along the lifetime of the infusion set and mation, with initial conditions set to nominal parameters in
more abrupt changes in insulin pharmacokinetics due to the the Hovorka model, far from the actual values (representing a
use of a new infusion site after rotation, with different sub- new infusion site); and 3) the same case as 2), but the initial
cutaneous tissue properties (for instance, affected by lypo- conditions of the pharmacokinetic parameters are set to the
hypertrophia [18]). Due to the limited duration of the data average of the parameter values estimated for the rest of
(5 h), the above scenarios are characterized in [35] through the patients (following a cross-validation procedure), which
the observer initialization, since a change in infusion site is are closer to the actual values, as demonstrated by data.
expected to imply a larger mismatch between the observer From a practical perspective, only individual parameters
model and the actual behavior. Thus, three cases are are available for the initialization of pharmacokinetic

54  IEEE CONTROL SYSTEMS MAGAZINE  »  FEBRUARY 2018

 increase further in this range as well as uncertainty in initial
200 conditions of the model state. A maximum envelope width of
Glucose (mg/dL)

approximately 12 mU/L is produced for plasma insulin concen-

150
Plasma

tration, which corresponds to more than twice its basal value.

To conclude, interval models embed variability into the pa-
100
tient’s model [S10] and can be useful in worst-case approach-
50 es for an increased robustness. In any case, robustness to
0 50 100 150 prediction errors and model adaptation are essential features
t (min)
of an AP.
(a)
30 References
Insulin (mU/L)

25 [S1] J. Bondia and J. Vehí, “Physiology-based interval models: A frame-

Plasma

20 work for glucose prediction under intra-patient variability,” in Predic-

tion Methods for Blood Glucose Concentration, H. Kirchsteiger, J. B.
15 Jørgensen, E. Renard, and L. del Re, Eds. New York: Springer Interna-
10 tional, 2015, pp. 159–181.
5 [S2] N. S. Nedialkov and K. R. Jackson, “An interval hermite-obresch-
0 50 100 150 koff method for computing rigorous bounds on the solution of an initial
t (min) value problem for an ordinary differential equation,” Reliable Comput-
(b) ing, vol. 5, no. 3, pp. 289–310, 1999.
[S3] R. Calm, M. García-Jaramillo, J. Bondia, M. Sainz, and J. Vehí,
“Comparison of interval and Monte Carlo simulation for the prediction
Figure S2  (a) Plasma glucose and (b) plasma insulin con- of postprandial glucose under uncertainty in type 1 diabetes mellitus,”
centration envelopes of the Hovorka model for an uncertainty Comput. Methods Programs Biomed., vol. 104, no. 3, pp. 325–332,
of ! 30% in parameter t maxI, as computed by the simulator in 2011.
[S9]. Monte Carlo simulations are also presented in grey for [S4] M. Kieffer and E. Walter, “Guaranteed nonlinear state estimator
for cooperative systems,” Numerical Algorithms, vol. 37, no. 1, pp.
comparison.
187–198, 2004.
[S5] E. Sontag, “Monotone and near-monotone biochemical networks,”
In [S9], tight glucose envelopes were obtained for mod- Syst. Synthetic Biol., vol. 1, no. 2, pp. 59–87, 2007.
els of certain complexity like the Hovorka model [23], under [S6] K. Bondar, “Some scalar difference inequalities,” Appl. Math. Sci.,
vol. 5, no. 60, pp. 2951–2956, 2011.
uncertain input, initial state, and model parameters with little [S7] H. Smith, Monotone Dynamical Systems: An Introduction to the
computational complexity. For example, Figure S2 shows the Theory of Competitive and Cooperative Systems. Providence, RI: AMS
envelopes for plasma glucose and plasma insulin concentra- Bookstore, 2008.
[S8] V. Puig, A. Stancu, and J. Quevedo, “Simulation of uncertain dy-
tions for the interval Hovorka simulator developed in [S9], con- namic systems described by interval models: A survey,” in Proc. 16th
sidering an uncertainty of ! 30% in time-to-peak plasma insu- Int. Federation Automatic Control World Congr., 2005, pp. 1–12.
lin concentration, t maxI, while the rest of the model parameters [S9] D. de Pereda, S. Romero-Vivó, B. Ricarte, and J. Bondia, “On the
prediction of glucose concentration under intra-patient variability in
are kept at their nominal value. A prediction horizon of 150 min type 1 diabetes: A monotone systems approach,” Comput. Methods
is considered. Reported variability of this single pharmacoki- Programs Biomed., vol. 108, no. 3, pp. 993–1001, 2012.
netic parameter yields to a glucose concentration after 150 [S10] A. J. Laguna, P. Rossetti, F. J. Ampudia-Blasco, J. Vehí, and
J. Bondia, “Experimental blood glucose interval identification of
min ranging from 84 to 154 mg/dL, that is, almost a twofold dif- pa­­t ients with type 1 diabetes,” J. Process Control, vol. 24, no. 1,
ference. Variability in insulin sensitivity and meal intake would pp. 171–181, 2014.

parameters. Figure 4 shows how real-time estimation of improves estimation accuracy, with a reduction of RMSE of
pharmacokinetic parameters leads to the detection of sig- 73%, with respect to a population model approach.
nificant deviations in plasma insulin prediction with the pop- In the simulation study presented in [37], the unscented
ulation model, reaching more accurate estimations after a Kalman filter (UKF) is used to calculate the bolus insulin size
transient time (the solid blue line). The impact of this transient based on a state estimate (which includes plasma insulin) and
time on closed-loop performance needs to be investiga­­ted. Once the meal size announced by the patient. The rationale is that
adapted, insulin estimation is expected to remain accurate aside from the meal size announcement, information pro-
enough (the solid green line). The root mean square error vided by the early postprandial period can help to reduce the
(RMSE) and mean absolute relative deviation (MARD) are impact of errors in carbohydrate counting by the patient,
computed for the assessment of the observer performance in allowing for the computation of more accurate insulin bolus
the cohort of patients. The obtained RMSE values are 24.5 ± doses even at the expense of a delayed administration. The
16.5 mU/L for case 1), 14.9 ± 7.7 mU/L for case 2), and 6.6 ± IVP glucose-insulin model [28] described in the section “Glu-
3.9 mU/L for case 3). Consideration of the pharmacokinetic cose-Insulin Models” and Table 1 is used in this work. Insulin
parameters k e and t maxI as extended states significantly sensitivity and the meal compartment are also estimated via a

FEBRUARY 2018  «  IEEE CONTROL SYSTEMS MAGAZINE  55

 Observability of Nonlinear Systems

O bservability measures how well internal states of a system The Gramian observability matrix can be computed from
can be deduced by knowledge of its external outputs. A experimental or simulation data within a region where the
dynamical system is called observable on a finite time interval process is to be operated (see [43], [S12], and [S13] for de-
[t 0, t f ] if any initial state x (t 0) = x 0 is uniquely determined by the tails). In these cases, it is called the empirical observability
corresponding output y(t) for t ! [t 0, t f ]. Gramian. For nonlinear systems, each state is normalized
All physiological glucose-insulin models belong to the such that its energy transfer to the output is equal to one. Its
category of nonlinear models. The observability of a non- singular values must then be determined. The smallest sin-
linear system gular value m quantifies the energy transfer from the least
observable state to the outputs. A system is considered un-
xo (t) = f (x (t), u (t)),  (S5)
observable if m = 0, and practically observable if m 2 0.1.
y (t) = h (x (t)),  (S6)
Moreover, m 1 0.1 for the normalized system indicates that
can be analyzed by means of analytical methods, through the the signal-to-noise ratio might be too small for a reliable
Lie derivatives [S11], or by numerical methods, through the em- state estimation.
pirical observability Gramian [S12]. Both of these models have It is worth noting that even when the system is observ-
been used in T1D research. In particular, Lie derivatives are able, difficulties in the reliable estimation of pharmacokinetic
used in [35] (Hovorka model) and [42] (Bergman model) and parameters may arise due to nonidentifiability issues. The a
the empirical observability Gramian in [43] (Bergman model). priori or structural identifiability property of a model indicates
if the parameters of the model can be determined, assuming
Lie Derivatives that all observable variables are error free. The lack of struc-
The Lie derivative of h (x) with respect to f (x, u) is defined as tural identifiability arises due to the model structure only and
L f (x, u) h (x) =4 h (x) f (x, u) and is calculated recursively as follows: is independent of the amount and quality of the given experi-
mental data. However, a parameter that is structurally iden-
L if (x, u) h (x) = L f (x, u) ` L f (x, u) h (x) j,
(i -1)
if i = 1, 2, ...
0
tifiable may still be practically nonidentifiable. This can arise
L f (x, u) h (x) = h (x) .
due to insufficient amounts and quality of experimental data
The system (S5)–(S6) is considered observable if the following or the chosen measurement time points. Common problems
Jacobian matrix in the identification of glucose-insulin models are the lack of
R V excitability of basal insulin, which is rarely variable enough,
S h (x) W
2 SL f (x, u) h (x)WW 
S
(S7)
especially during the night, and the administration of insulin
2x S g W boluses at mealtime, making it difficult to separate the meal
SL nf (x-,1u) h (x)W
T X and insulin effects on glucose after meal intake.

has rank n, where n is the system order.

References
[S11] R. Hermann and A. J. Krener, “Nonlinear controllability and ob-
Empirical Observability Gramian
servability,” IEEE Trans. Autom. Control, vol. 22, no. 5, pp. 728–740,
The Gramian observability matrix W 0 is also determined for 1977.
observability analysis. This matrix quantifies generalized en- [S12] D. Geffen, D. Findeisen, M. Schliemann, F. Allgower, and M.
Guay, “Observability based parameter identifiability for biochemi-
ergy transfer E 0 from initial state x 0 to the output within an
cal reaction networks,” in Proc. American Control Conf., 2008, pp.
infinite time horizon 2130–2135.
[S13] S. Lall, J. Marsden, and S. Glavaski, “A subspace approach to
balanced truncation for model reduction of nonlinear control systems,”
E0 = # y T (x) y (x) dx = x T0 W 0 x 0 .  (S8) Int. J. Robust Nonlinear Control, vol. 12, pp. 519–535, Feb. 2002.

state extension. Different bolus administration strategies are is exactly known. Due to the problem posed in this work,
investigated (at mealtime, 15 min after, and 30 min after the administration before the meal is not considered. However,
beginning of the meal) for ten patients for three-day simula- clinical trials that assessed the best timing for meal bolus
tions with three meals per day, checking the median time times generally conclude that the administration of the in­­
spent in hyperglycemia, within target, and in hypoglycemia. sulin bolus between 10 and 20 min before the meal leads to
No specific results on the observer performance are reported. reduced glycemic excursions, as compared to mealtime or
They conclude that administering the meal bolus 15 min after after the meal [38], [39].
mealtime both reduces the risk of hypoglycemia when meal Other works address the design of insulin observers, but
size is overestimated by 50% and the time spent in hyperglyce- either an intravenous route is considered, not used in current
mia for a meal size underestimation by 50%. However, admin- AP systems, or endogenous insulin secretion is included,
istration of insulin at mealtime is optimal when the meal size which does not apply to T1D. These works are reviewed below,

56  IEEE CONTROL SYSTEMS MAGAZINE  »  FEBRUARY 2018

Kalman Filter Extensions to Nonlinear Systems

P hysiological variability can be represented as noise in the

system dynamics. The use of sensors also introduces Actual (Sampling) Linearized (EKF) UT
noise into the measurements. If stochastic properties of these Mean
noise sources are available, state estimation may be performed
more effectively than simply using sensor signals as noise-free Covariance Sigma Points
signals and estimating the state based on noise-free state tran-
sition models. Rudolph Kalman investigated this problem and y = f (x ) γ = f (x)
developed the widely used Kalman filter [S14]. Originally, he y = f (x ) Weighted
Py = ∇f Px∇f T Sample Mean
addressed the general problem of estimating the state x ! R n and Covariance
of a discrete-time controlled process that is governed by the
linear stochastic difference equations True Mean f (x ) UT Mean
Transformed
Sigma
x (k + 1) = Ax (k) + Bu (k) + w (k), w (k) ~ N (0, Q),  (S9) ∇f Px∇f T Points
True
y (k) = Cx (k) + v (k), v (k) ~ N (0, R),  (S10) Covariance UT Covariance

Figure S3  An example of the unscented transformation

where y ! R r are the measurements and the random variables (UT) for mean and covariance propagation, compared to the
w (k) and v (k) represent the process and measurement noise, extended Kalman filter (adapted from [S19]).
which are assumed to be independent, white, and with normal
probability distributions of variances Q and R, respectively.
The Kalman filter provides an efficient computational method approximation, and a weighted mean and covariance is found.
to estimate the state of a process in a way that minimizes the The sigma points are chosen so that their sample mean and
mean of the squared error in a two-step process: the prediction sample covariance are in agreement with the mean xr and
step (the current state estimate is projected ahead in time) and covariance P of the n-dimensional random variable x. The
the correction step (the projected estimate is adjusted by an underlying idea is to approximate the probability distribution
actual measurement). instead of the function (see Figure S3).
Since the time of its introduction, the Kalman filter has Moreover, Kalman filter extensions can also be used in
been the subject of extensive research and application. The continuous-time systems that are represented by differen-
Kalman filter theory has been extended to nonlinear process- tial equations
es, given by
xo (t) = f (x (t), u (t)) + w (t), w (t) ~ N (0, Q (t)),  (S13)
x (k) = f (x (k - 1), u (k - 1)) + w (k - 1), w (k - 1) ~ N (0, Q), (S11) y (t) = h (x (t)) + v (t), v (t) ~ N (0, R (t)),  (S14)
y (k) = h (x (k)) + v (k), v (k) ~ N (0, R),  (S12)
after a discretization procedure. The glucose-insulin models
where f (x (k - 1), u (k - 1)) defines the system dynamics and found in the literature belong to these kinds of systems.
h (x (k)) denotes the measurement function. For that, these
kinds of systems are transformed through a linearization pro- References
cedure known as the EKF [S15], [S16] or by an unscented [S14] R. Kalman, “A new approach to linear filtering and prediction prob-
transformation (UT) called the UKF [S17]–[S19]. The use of the lems,” Trans. ASME J. Basic Eng., vol. 82, ser. D, pp. 35–45, 1960.
[S15] P. Maybeck, Stochastic Models, Estimation, and Control (Math-
EKF has two well-known drawbacks in practice. First, linear- ematics in Science and Eengineering v. 141a). New York: Academic,
ization can produce highly unstable filters if the assumption 1979.
of local linearity is violated. In addition, the derivation of the [S16] G. Smith, S. Schmidt, and L. McGee, “Application of statistical
filter theory to the optimal estimation of position and velocity on board
Jacobian matrices are nontrivial in most applications and often a circumlunar vehicle,” Tech. Rep., National Aeronautics and Space
lead to significant implementation difficulties. These disadvan- Administration, NASA TR R-135, 1962.
tages may be overcome with the use of the UKF. The UKF is [S17] S. Julier and J. K. Uhlmann, “New extension of the Kalman
filter to nonlinear systems,” in Proc. AeroSense’97 Conf., 1997, pp.
based on the UT, which is a deterministic sampling approach 182–193.
to capture mean and covariance estimates with a minimal set [S18] S. Julier and J. K. Uhlmann, “Unscented filtering and nonlinear
of 2n + 1 state sample points, called sigma points, based on estimation,” Proc. IEEE, vol. 92, no. 3, pp. 401–422. 2004.
[S19] E. Wan and R. van der Merwe, “The unscented Kalman Filter for
a square-root decomposition of the prior covariance. These nonlinear estimation,” in Proc. IEEE Adaptive Systems Signal Process-
sigma points are propagated through the nonlinearity, without ing Communications Control Symp, 2000, pp. 153–158.

FEBRUARY 2018  «  IEEE CONTROL SYSTEMS MAGAZINE  57

since they address glucose-insulin models and may be rele- simulation. Results show that all three filters successfully trace
vant to future developments in the AP context. glucose and insulin levels from noisy blood glucose measure-
A reduced-order Luenberger observer [40] is designed ments with similar performances (the RMSE plots overlap), but
in [41] for the three-state minimal Bergman model [22], the UFK computational load is lower.
where endogenous insulin secretion in the original model In [45]–[47], nonlinear, discrete-delay, differential equa-
is substituted with an intravenous insulin infusion to rep- tion (DDE) models are used in the context of closed-loop
resent subjects with T1D. Disturbances are included in the glucose control. The authors state that DDE models better
form of an exogenous glucose infusion rate. The observer represent the pancreatic insulin delivery rate, allowing for
aims to reconstruct the remote insulin compartment and the extension of the methodologies to type 2 diabetes. More-
plasma insulin deviation from plasma glucose deviation over, the model has satisfactory properties, such as positiv-
measurements. The obtained results are tested on nonlin- ity and boundedness of solutions or local attractivity of a
ear, closed-loop simulations using the disturbance rejec- single positive equilibrium. The model is also statistically
tion linear-quadratic method. A standard meal disturbance robust since its parameters are identifiable with very good
with a 6-h duration is considered as the test case. It is shown precision by means of standard perturbation experiments,
that the observer is faster than the system itself and can such as IVGTT. The authors consider the problem of track-
provide very good state recovery performance. ing a desired plasma glucose trajectory from initial hyper-
In [42], the Bergman model is used in combination with dif- glycemia, making use of only glucose measurements. To
ferent filters (symmetric UKF, EKF, simplex UKF, and particle this aim, they use the nonlinear observer for time-delay
filter) for the estimation of plasma insulin concentration. systems to estimate the plasma insulin concentration
Endogenous insulin secretion is considered. Furthermore, R t V R V
S dG (t) W S t t Tgh W
linear interpolation is used to obtain glucose measurements - K xgi G (t) I (t) +
S dt W S VG W
in a time grid of 1 min. After evaluating observability by =
S dIt (t) W S u (t) W 
t TiG max c t mW
the Lie derivatives, the designed observers are validated SS W S-K xi I (t) + f G (t - x g) +
dt W S VI VI W
with data from an IVGTT (see the section “Glucose-Insulin T X T -1 t t (t)), X
+ Q (G (t), It (t)) W (G (t) - G (7)
Models”) in nondiabetic subjects by computing the RMSE.
The symmetric UKF showed better results with an RMSE where Gt (t) and It (t) denote the glucose and insulin es­­
value of 10.277 mU/L, followed by EKF with 13.533 mU/L. timates, respectively, and Q -1 is the inverse matrix of
This work is extended later in [43] to incorporate parameter the function
estimation via state extensions characterizing first- and
Q (x 1, x 2) = ; E . (8)
1 0
second-phase insulin responses as well as compartments for
-K xgi x 2 -K xgi x 1
glucose and subcutaneous insulin inputs and subcutaneous
glucose measurements. Both the observability of states and The observer gain matrix W ! R 2 # 1 is designed to en­­
external inputs and the identifiability of model parameters sure that
are analyzed by the empirical observability Gramian from
data. For the purpose of model validation, four scenarios are t = ;0 1E - W 61 0@
H
0 0
simulated: an IVGTT for nondiabetic subjects, an oral glucose
tolerant test (OGTT) for nondiabetic subjects, an IVGTT for is Hurwitz with prescribed eigenvalues in the left half of
diabetic subjects, and an IVGTT with an insulin bolus after the complex plane.
30 min for diabetic subjects. Similar to IVGTT, the OGTT is a Numerical results from simulations show that this ap­­
clinical test to analyze glucose metabolism. However, glucose proach is robust with respect to the uncertainties of the
is administered orally instead of intravenously. Simulated model parameters as well as the glucose measurement
scenarios are compared with measured data from nondia- errors and insulin pump malfunctions. Table 2 summa-
betic and diabetic pigs. These data are used for parameter rizes the above results.
identification and model adaptation. The results are graphi-
cally evaluated, concluding that a real-time estimation of Insulin Infusion Limitation
states such as plasma insulin, and parameters such as second- in the Artificial Pancreas
phase insulin response gain, is possible. It may improve real- As previously stated, the main justification for the predic-
time state prediction and the personalized model. tion (or estimation from data) of plasma insulin or IOB in an
In [44], the Bergman model is used with the UKF, the cuba- AP is the avoidance of the controller’s overactuation (for
ture quadrature Kalman filter (CQKF), and the Gauss–Her- instance, after a meal disturbance), inducing an excessive
mite filter (GHF) to track plasma glucose, plasma insulin, and insulin infusion with the consequent risk of hypoglycemia.
interstitial insulin levels. The model is based on the work pre- MPC, proportional-integral-derivative (PID), and fuzzy
sented in [42]. Thus, endogenous insulin secretion is again con- logic (FL) are the major control strategies used in the AP
sidered. The above filters are compared based on estimation framework and most extensively evaluated through clinical
accuracy, in terms of RMSE, and computation efficiency by trials [48]. According to [49], 18 AP systems are currently

58  IEEE CONTROL SYSTEMS MAGAZINE  »  FEBRUARY 2018

 Table 2 Summary of publications addressing the design of observers in the context of diabetes. MHSE: moving horizon state
estimator; EKF: extended Kalman filter; RMSE: root mean square error; MARD: mean absolute relative deviation; UKF: unscented
Kalman filter; CQKF: cubature quadrature Kalman filter; GHF: Gauss–Hermite filter; DDE: discrete-delay, differential equation.

Reference Model Observer Validation Evaluation

Wilinska et al., 2012 [31] Hovorka Kalman N/A N/A
Gondhalekar et al., 2013 [32] Linear Luenberger N/A N/A
Lee et al., 2014 [34] Linear Luenberger, MHSE Simulation SSE
Pereda et al., 2015 [35] Hovorka EKF Simulation and RMSE and MARD
data from type 1
diabetes subjects
Boiroux et al., 2015 [37] MVP UKF Simulation Median time in hyperglycemia,
within target or hypoglycemia
Kovács et al., 2007 [41] Bergman Luenberger Simulation Graphs
Eberle and Ament, 2011 [42] Bergman EKF, UKF, PF Simulation RMSE
Eberle and Ament, 2012 [43] Bergman UKF Simulation and Graphs
data from pigs
Biswas et al., 2016 [44] Bergman UKF, CQKF, GHF Simulation RMSE
Palumbo et al., 2012–2015 [45]–[47] DDE Nonlinear Simulation Amount of hypoglycemic events
and plasma glycemia

U max (k) =)
under development at different stages, of which eight are I CHO (k) + I G (k)-IOB (k) if I CHO (k) +I G (k) 2 IOB (k )
,
based on MPC, six on a PID-type control algorithm, three I CHO (k) otherwise
on FL, and one in a bio-inspired approach. All control strat- (11)
egies must deal with the limitation of insulin infusion for with
the sake of safety. I CHO (k) = M (k) ·IC, (12)
(G (k) - G ss (k)) · CF if G (k) - G ss (k) 2 0
Model Predictive Control I G (k) = ' , (13)
0 otherwise
One of the advantages of MPC is that it can easily incorporate
constraints into the optimization process of the defined cost where I CHO (k) is the insulin needed at time k to compen-
function. An example of such a procedure is given in [50], sate a meal, M (k) is the amount of carbohydrates from the
where a dynamic safety constraint on the maximum rate of meal absorbed at time k, and IC [U/g] is the insulin-to-
insulin delivery based on empirical clinical knowledge is inte- carbohydrate ratio used currently by clinicians (that is, how
grated into an MPC scheme, to avoid overdelivery of insulin. much insulin per gram of carbohydrate must be adminis-
The following constrained cost function is considered tered to compensate a meal). Moreover, I G (k) represents
the insulin needed to compensate for a deviation from the
min ^~ y yt T yt + ~ u u
t T ut + ~ Du Du T Du h, (9) target (considered here to be the steady-state glucose
u!R
concentration) G ss (k), as given by the correction factor
s.t. - 0.5ut ss # ut # U max (k) 61, 1, f, 1@T, (10) CF 6U (mg/dL) -1@ . CF is used by clinicians in current in­­
­s ulin therapy to determine how much insulin must be
where yt ! R p # 1 is the vector of predicted plasma glucose administered to compensate for an excess of one glucose
deviations with respect to the glucose setpoint, with p the concentration unit with respect to the glucose target.
output prediction horizon; ut ! R m # 1 is the vector of future Finally, IOB(k) is the predicted IOB at time k, which is
calculated deviations of the insulin infusion rate with obtained from the convolution of the insulin administered
respect to the basal rate ut ss (that is, the insulin rate required in the last 8 h and a family of “active insulin curves,” as
at steady state to obtain the desired steady-state glycemia), given by insulin pump manufacturers ranging in different
with m the control horizon; and Du ! R m # 1 is the vector durations of insulin action from 2 to 8 h. These curves are
of future insulin infusion increments. The constraint, based on insulin action plots and express the percentage of
expressed in terms of deviations of the insulin infusion rate remaining active insulin along time. Curvilinear and linear
with respect to the basal rate ut ss, indicates that the insulin versions are available, depending on the manufacturer [51].
infusion rate is limited between half its basal value and a Although linear versions are easier to use by patients,
time-dependent maximum deviation from its basal value curvilinear versions represent better subcutaneous insulin
U max (k), given by pharmacokinetics and are preferred. The use of dynamic

FEBRUARY 2018  «  IEEE CONTROL SYSTEMS MAGAZINE  59

curves to account for daily variations in insulin sensitivity glucose is decreasing rapidly; and 4) capping the insulin
is proposed in [52]. infusion to the preprogrammed basal rate if a pump occlu-
In a simulation study reproducing a 24-h clinical proto- sion is inferred by the MPC. Unfortunately, no more techni-
col with three meals [50], limitations of IOB according to cal details were published other than the ones presented
(9)–(10) resulted in a reduction of the number of simula- here, to the best of our knowledge.
tions yielding a hypoglycemic event, compared to the same
MPC controller without the IOB constraint, from 50% PID Control
(18.6% of the overall simulation time) to 10% (0.75% of the Two approaches with clinical validation are found for insu-
overall simulation time). This was achieved at the expense lin limitation in PID-based AP systems: insulin feedback
of mild hyperglycemia but for very short periods (3.5% of (IFB) and sliding mode reference conditioning (SMRC).
overall simulation time). Although applied to PID control, these approaches may be
In [53], the insulin limitation problem is cast into an aug- used with other control structures.
mented single-input, multiple-output MPC formulation,
where the linear combination Insulin Feedback
In a healthy person, the beta cells automatically avoid insu-
U (t) = s I I (t) + s q q (t) (14) lin excess by autoinhibiting the endogenous insulin produc-
tion, depending on the insulin plasmatic concentration. In
of plasma insulin I (t) and a measure of the pending effect T1D this phenomenon is lost, due to the destruction of the
of IOB q (t) are added as new system output. A generalized autoregulating beta cells. However, an AP can emulate this
predictive control (GPC) with extended output y (t) = process using the IFB algorithm [55], where a prediction of
[G (t) U (t)] T (where G (t) is plasma glucose) is then formu- plasma insulin is fed on the forward delivery of insulin of
lated. GPC optimizes the multistage quadratic cost func- the PID algorithm. All PID-based studies have used this
tion, which, after signals discretization, is written as strategy since then [48], with one exception that is reviewed
in the next section [56].
Nm Nu
The classic PID control within the AP framework usu-
/ + / m k ^Du t +k h2, (15)
2
J GPC = d k C (rt +k - y t +k)
k = Nd k =0 ally includes the following equation:

where the first term penalizes the error with respect to the u AP = u PID + u basal + u bolus, (18)
setpoint r along the prediction horizon [N d, N m], and the
second term penalizes the control action along the control where u basal is the basal insulin needed by the patient to
horizon with bound N u . Parameters d k and m k are weights. maintain a proper glycemia level between meals and at night
In this case, the following biexponential insulin phar- (basal metabolic conditions), u PID is the additional control
macokinetic model is used: action (insulin delivered) proposed by the PID algorithm to

I (t) = KU 0 ^e -a1 t - e -a2 t h, (16)

deal mainly with disturbances, and u bqlus is the extra insulin

dose needed at meal time (considering meal announcement,
which corresponds to (2) with K = 1/(K cl (x 2 - x 1)), a 1 = 1/x 2, where the patient informs the control algorithm about the meal
and a 2 = 1/x 1 . The pending effect of IOB at time t is com- time and estimated amount of carbohydrates).
puted as A possible implementation of the IFB algorithm (see [30]
or [29]) is
q (t) = #t 3 I (x) dx = aKU 0
^a 2 e - a1 t
- a 1 e -a2 th . (17)
1 a2 u IFB = u AP - cI, (19)

Factors s Ip and s q in (14) scale the corresponding units of I (t) where the original control action u AP is limited by an action
[U/dL] and q (t) [Umin/dL] into [mg/dL] for an homogeneous proportional to the predicted plasma insulin concentration
cost index. It is noted that q (t) is an area-under-the-curve of I. For this purpose, the biexponential model (2) is used,
plasma insulin concentration, not the mass of insulin. with the transfer function given by
Finally, another relevant MPC-based AP system is pre-
sented in [23], which was already tested in free living con- 1 1
I (s) x1 x2
ditions without supervision over three months [54]. The G PK IFB (s) = = , (20)
K cl c s + 1 mc s + 1 m
u IFB (s)
system includes the following safety checks to avoid excess x1 x2
of insulin delivery [31]: 1) insulin infusion rate is limited to
two to five times the preprogrammed basal rate, depend- which corresponds to (1) with input signal u IFB . The c
ing on the current glucose measurement, the time since the parameter multiplying I must be set by the control engi-
previous meal(s), and carbohydrate content of meal(s); 2) neer. The IFB approach described is, from a classical control
shutting off insulin delivery when the glucose measure- point of view, mathematically equivalent to a cascade con-
ment is below 77 mg/dL; 3) reducing insulin delivery when trol system [30], as shown in Figure 5. For IFB to maintain

60  IEEE CONTROL SYSTEMS MAGAZINE  »  FEBRUARY 2018

 Gr + uPID + + uAP + uIFB Patient
Controller PID (s )
G (s )
– + + –
ubolus ubasal
Insulin
Pharmacokinetic
Model

I
γ

Figure 5  A block diagram for an artificial pancreas controller with insulin feedback (IFB).

the original insulin infusion u AP at steady state, the original Sliding Mode Reference Conditioning Control
PID parameters must be retuned, or a reduction in insulin SMRC is proposed in [56] to build an external feedback loop
infusion would lead to an increase in glucose concentration called the safety auxiliary feedback element (SAFE), to keep
[30]. To avoid this, the term u AP in (19) must be multiplied by IOB (represented by the size of the subcutaneous insulin depot)
f = ^1 + c lim s " 0 G PK IFB (s) h, which is equivalent to retuning inside predefined limits. This measure differs from the active
the controller’s proportional gain. Integral and derivative insulin curves currently used in insulin pumps that include
terms may also require some adjustments, depending on insulin pharmacodynamics [59]. Upper and lower limits for
the tuning of c. IOB can be defined if desired. SMRC uses the modulation of
The equation glucose targets as a new degree of freedom for glucose regula-
tion (see Figure 6). A prediction of IOB, IOB(t), drives a switch-
u IFB = u AP - c ^I - I B h, (21) ing function v SM (t), triggering a discontinuous signal ~ (t)
that, after a filtering step, sums to the standard glucose target.
presents an equivalent IFB implementation as a limiting The rationale behind SMRC is that it provides an upper limit
action proportional to I - I B, where I B is the predicted of IOB by increasing the desired glucose target. Any controller
basal plasma insulin In this case, the additional term is will react in the sought direction of diminishing insulin infu-
zero in basal (stationary) conditions and PID gain retuning sion. The opposite applies when a lower limit for IOB is desired.
is not necessary. SMRC originates on concepts of invariance control and
Different studies show the good performance of the IFB sliding regimes as a transitional mode of operation. In con-
approach. In [30], simulation results using the Hovorka trast to conventional sliding mode control, the aim here is
model [23] are presented for 24 h with three meals and a not evolving toward the equilibrium point. Only when the
decrement of the peak postprandial excursion. At the same system (by itself) reaches a given sliding surface separating
time, a reduction of late-postprandial hypoglycemia is the space into feasible and unfeasible regions (in this case,
reported. In [55], the ability of IFB to improve the breakfast characterized by the constraint on IOB) is the sliding regime
meal profile in a 30-h closed-loop, in-clinic study with stabilized by conditioning the reference until the system
eight adult subjects with T1D was assessed with a substan- returns (by itself again) to the feasible region. In this sense,
tial improvement over prior studies, but supplemental car- there is no reaching mode because no control effort is done
bohydrates after breakfast were needed for three subjects. to drive the system to the sliding surface [60].
The clinical study also showed feasibility of overnight con- Given a (possibly time-varying) upper limit on IOB
trol. On the other hand, in [57], four subjects were studied ^IOB (t) h and system state x (t), the set
for 24 h and the IFB, as compared to a standard PID con-
R = " x (t)|IOB (t) # IOB (t) , (22)
trol, reduced the occurrence of hypoglycemia without
increasing meal-related glucose excursions. Finally, [58] is invariant (that is, if the initial state fulfills the constraint
presents simulation results of the IFB strategy used in a IOB (t) # IOB (t), then it will be fulfilled for all t 2 0 ) for a
fully implantable AP using intraperitoneal insulin deliv- discontinuous signal ~ (t) of the form
ery and glucose sensing (that is, the peritoneal cavity,
~ (t) = '
~
+
if v SM (t) 2 0
which is the area in the body that contains the abdominal , (23)
0 otherwise
organs, is used for rapid sensing and actuation, better
emulating the healthy pancreatic activity). In this case, the with ~ + 2 0 large enough (see [56] for technical details) and
IFB alone was not enough to attenuate postprandial under-
v SM (t) = IOB (t) - IOB (t) + / x i ^IOB (t) (i) - IOB (t) (i)h, (24)
l -1
shoot but in combination with antiwindup provided very
good results. i =1

FEBRUARY 2018  «  IEEE CONTROL SYSTEMS MAGAZINE  61

 Gr IOB

ω Insulin
IOB
First-Order Filter σSM Pharmacokinetic
Model
Grf

+
uPID + + uAP
Controller PID(s ) Patient G (s )
– + +
ubolus ubasal

Figure 6  A block diagram for an artificial pancreas controller including the sliding mode reference conditioning control structure.

where l is the relative degree between the output IOB (t) insulin pharmacokinetic model [23] was finally chosen in
and the input ~ (t), superscript (i) denotes the i th deriva- the implementation for clinical evaluation due to parsi-
tive, and x i are constant gains. mony criteria, that is,
The first-order filter
dS 1 (t)
dG RF (t) = u sc (t) - 1 S 1 (t), (26)
dt t maxI
=-aG RF (t) + a (G R (t) + ~ (t)) (25)
dt dS 2 (t) 1
= (S (t) - S 2 (t)) . (27)
dt t maxI 1
keeps all signals in the control loop smooth, where G RF (t) is
the modulated (conditioned) glucose target fed to the con- IOB was then defined as
troller, G R (t) is the standard glucose target (usually con-
IOB (t) := S 1 (t) + S 2 (t) . (28)
stant), and a defines the filter cut-off frequency. Stability of
the system is guaranteed, since the SMRC loop acts only on Figure 7 shows the typical time profile of IOB as defined by
the setpoint, which is always bounded. (26)–(28), expressed as a percentage of the bolus size, yield-
Relative degree l is determined by the relative degree of ing a curvilinear subcutaneous insulin pharmacokinetics,
the filter (25) and the relative degree of the IOB predictor, also characteristic of active insulin curves provided by
since the controller has a proportional action. Although the insulin pump manufacturers [51].
subcutaneous insulin pharmacokinetic model in [61] was Equations (26)–(28) have a relative degree of one, giving rise
considered in the initial work [56] (which is the model in to a total relative degree of l = 2. The switching function (24) is
the UVA/Padova simulator), the Hovorka subcutaneous
v SM (t) = IOB (t) - IOB (t) + x c m . (29)
dIOB (t) dIOB (t)
-
dt dt

For an interpretation of the above switching function, con-

100 sider, for simplicity, a constant upper IOB limit IOB. This is
Bolus Remaining at s.c. Depot (%)

90 used, for instance, in a postprandial period or during the

80 night. The switching function v SM (t) will be positive when
70 IOB (t) + xd (IOB (t))/(dt) exceeds the limit IOB, that is, when
60 the predicted IOB is about to violate the constraint. How close
50 IOB is allowed to approach (from below) its limit will depend
40 on the IOB trend, weighted by the tuning parameter x.
30
This defines a threshold on IOB, which corresponds to
IOB - xd (IOB (t))/(dt) . A rapidly changing IOB will lower
20
this threshold, compared to an IOB slowly approaching the
10
limit. A similar interpretation holds for the variable upper-IOB
0
0 1 2 3 4 5 6 7 8 limit, IOB (t), but in this case, the threshold on IOB depends
t (h) on the difference in the trends for IOB and its limit. When
IOB goes beyond this threshold, an increment of the glucose
Figure 7  A time profile of subcutaneous insulin depot as a per-
setpoint is triggered, provoked by the discontinuous signal
~ (t) and amounting to ~ , after a transient given by the
+
centage of the bolus size administered at t = 0, according to the
model in (26)–(28) (t max I = 55 min) . filter (25). As a reaction to the new (generally much higher)

62  IEEE CONTROL SYSTEMS MAGAZINE  »  FEBRUARY 2018

glucose setpoint, the controller will reduce the insulin deliv- Z1 t # t1
]
ery, impeding the violation of the IOB limit for high enough ] P2 t1 1 t # t2
~ . The glucose setpoint will return to its original value after
+ fI/G [%] = [ P3 (t - t 3) , (30)
] t3 - t2 t2 1 t # t3
the transient imposed by the filter, when IOB returns below ]0
\ t3 1 t
the threshold, in which case ~ (t) = 0.
Similar formulas are obtained for lower-limiting IOB,
changing the inequality in (23) and the sign for ~ (t) (glu- and subtracted from the insulin to be delivered (see supple-
cose target must be decreased). mental data in [65]). Unfortunately, no further technical
SMRC was evaluated, in combination with a PID con- details are provided.
troller, in an in-clinic clinical study where 20 patients with Another FL approach is the dose safety system presented
T1D underwent a standardized 60 g-carbohydrate mixed in [66] and [67]. The fuzzy controller has two components: 1)
meal on four occasions (two in open loop and two in closed the FL dosing component, which computes via a rule matrix
loop). SMRC showed an improvement of postprandial con- an insulin dose based on the glucose value, rate, and accel-
trol, with a significant decrement of postprandial peak and eration, to keep glucose in the range of 80–120 mg/dL; and 2)
increment of time-in-range (70–180 mg/dL) without a cli­ the dosing personalization component, which computes a
nically meaningful increased risk of hypoglycemia [62]. personalization scale factor for the computed insulin dose.
Simulation validation of SMRC in combination with other In [66], it is stated that the system does not track IOB, although
controllers can be found in [63], demonstrating the benefit the function will be included in subsequent versions. In [67],
of the SAFE loop. an extension of the system with a hypoglycemia prevention
module is presented, which incorporates a hypoglycemia
Fuzzy Logic predictor based on pattern-matching techniques from glu-
The two control strategies presented previously (MPC and cose and IOB signals. However, to the best of our knowledge,
PID) are based on the use of mathematical models of glu- no further technical details have been disclosed.
cose-insulin dynamics. However, the biological system to
be controlled is nonlinear, complex, and subject to lags and Open Challenges
uncertainties. Therefore, it is very difficult to capture the The use of the AP in free living conditions poses extra chal-
physiological behavior of a patient. FL allows the develop- lenges to the performance of the controller and, with it, the
ment of a fuzzy controller without any patient model by efficiency of its mechanisms for insulin infusion limitation. An
including fuzzy rules (IF input is A THEN output is B), example is physical activity, which has very different effects
where A and B are linguistic variables. In this way, a con- on plasma glucose concentration, depending on its type and
troller can be built from medical expertise [64]. In addition, intensity. Mild and moderate-intensity physical activity re­­
multiple inputs and multiple outputs can be included in a quires a reduction of insulin infusion, while more intense activ-
natural way. Different groups have proposed a fuzzy-con- ities necessitate a rise in insulin, at least in early recovery [68].
troller-based AP [65], [66], and insulin limitation based on This can affect algorithms limiting insulin infusion under
predictions of IOB are also applied in FL systems. physical activity scenarios, which may need different tuning.
The MD-logic system presented in [65] is based on 1) a However, it implies that the system must be informed in some
control-to-range layer consisting of fuzzy rules, giving rise to way about physical activity, such as patient’s announcement or
an insulin recommendation expressed as percentage of measurements from a physical activity monitor.
patient’s basal insulin; and 2) a control-to-target layer that Moreover, the dynamic lags imposed by the subcutaneous
takes into account, among other things, special glucose route might be a barrier for efficient compensation of physical
dynamics requiring specific corrections and safety mea- activity, which may provoke rapid fluctuations in glucose.
sures. The control-to-range layer considers four inputs: past Subcutaneous insulin pharmacokinetics can also be affected
and future glucose trend and current and future glucose by physical activity. In [69], an increase of circulating insulin
level. The latter is predicted through an autoregressive is observed after a pump basal insulin rate reduction before a
model. The values of the trend input variables are defined as moderate-intensity aerobic exercise, which is hypothesized
steep descent, descent, stable, ascent, and steep ascent, to be due to the increased cardiac output and subcutaneous
defined over the range of ±5 mg/dL/min. Glucose level blood flow during exercise, provoking an acceleration of
values are defined as very low, low, normal, normal high, insulin pharmacokinetics. It is thus expected that physical
high, and very high, defined over the physiological range for activity may affect the performance of insulin observers. To
glucose. Two output variables are considered: percent change date, there are no accurate exercise models. Physical activity
in patient’s standard basal rate and standard bolus. In the can also compromise the accuracy of continuous glucose
control-to-target layer, the final insulin rate to be delivered is measurements, with an overall overestimation of glucose
computed. A prediction of IOB based on the following piece- values [70]. Possible sources of this loss of performance are
wise functional approximation of insulin aspart pharmaco- changes in subcutaneous blood circulation and increases in
kinetics is considered [7]: body and skin temperature during physical activity, which

FEBRUARY 2018  «  IEEE CONTROL SYSTEMS MAGAZINE  63

affect the dynamics of glucose transport between plasma and insulin curves, and the size of the subcutaneous insulin
the interstitial space where the sensor is located. depot are examples of such measures. Methods for insulin
Another important challenge is postprandial control. limitation differ depending on the control framework. MPC
There is growing evidence that insulin-dosing algorithms incorporates such measures as constraints or in the cost
should consider not only the carbohydrate content of a meal index for optimization. PID control makes use of additional
but also fat and protein [71]–[73]. An AP study investigating structures such as cascade control or reference condition-
the performance of a low-fat versus high-fat dinner with iden- ing. FL incorporates insulin measures as components of the
tical carbohydrate content showed elevated plasma glucose rule base. All these approaches are based on insulin predic-
levels during the high-fat dinner, despite the administration tors (or estimators). Due to the large variability, another
of 42% more insulin [74]. Model fitting from data produced in critical component of an AP is the real-time estimation and
this study revealed not only the expected lag in gastric empty- adaptation of model parameters. The availability of continu-
ing due to an increased fat content but also a lower insulin ous glucose measurements addresses this problem through
sensitivity in the high-fat meal [72]. In [73], a controlled study the design of observers. Different observer techniques have
was conducted comparing a low, fat-low-protein meal with a been used for this purpose, although validation with clini-
high-fat, high-protein meal with identical carbohydrate con- cal data is scarce. Results published on the development of
tent, covered with identical insulin doses, showing a twofold insulin observers were reviewed. Some open challenges
glucose incremental area under the curve in the latter case. In were also noted. This article aims to introduce the challenge
the same study, an adaptive, model-predictive, insulin bolus of automatic subcutaneous insulin delivery and the oppor-
strategy required 65% more insulin, with a 30/70% split over tunities that it brings to control engineering, to alleviate the
2.4 h, to achieve the glucose target in the high-fat, high-pro- current burden of self-control for patients with T1D.
tein case, compared to low fat, low protein. Further investiga-
tion is needed to understand how these findings affect the Acknowledgments
performance of insulin limitation algorithms in an AP. Meal This work was supported by the Spanish Ministry of Econ-
composition announcement is not desirable, due to the extra omy and Competitiveness (MINECO) through grant DPI2013-
burden for patients. Besides, current meal models only con- 46982-C2-1-R and the EU through FEDER funds.
sider carbohydrates, which is an important limitation.
The need to incorporate psychological stress into the AP Author information
is also under investigation. In [75], a significant relation- Jorge Bondia ([email protected]) received the M.Sc. degree
ship between daily stress (rated by the patient in a five- in computer science in 1994 and the Ph.D. degree in con-
point scale) and glucose variability was found as well as an trol engineering in 2002 from the Universitat Politècnica de
increased percentage of time in hypoglycemia and reduced València, Spain. He is currently a full professor and has been
carbohydrate consumption. However, mean glucose was teaching since 1995 in the Systems Engineering and Control
not affected. More studies are needed to reveal the clinical Department at the Universitat Politècnica de València, spe-
significance of these findings and justify its consideration cializing in the fields of automation, control engineering,
in an AP for individuals more reactive to stress. and biomedical engineering. His research activity began
As more data is available from long-term, clinical, closed- as a member of the University Institute of Control Systems
loop studies in free living conditions, improvements address- and Industrial Computing (Institute ai2), where he began a
ing all of these aspects may be expected. research program on diabetes technology and the artificial
pancreas in 2004. His main research interests are modeling
Conclusion of type 1 diabetes pathophysiology, intrapatient variability
Glucose control in T1D is a complex problem from the control and glucose prediction under uncertainty, tools for insulin
engineering point of view. The plant (that is, the patient) is a therapy optimization, new calibration algorithms for accu-
highly variant, nonlinear system due to the nature of physi­ racy improvement of CGM, and control algorithms for an
ology and its intrinsic variability. Significant lags are in­­ efficient and safe artificial pancreas. He has published more
troduced by the use of the subcutaneous route, due to its than 40 articles on the artificial pancreas and contributed
minimal invasiveness. The fact that insulin has a unidirec- to more than 100 conferences. He is currently the head of
tional action poses additional challenges to the efficacy and the MEDERI Living Lab at Institute ai2, dedicated to new
safety of control algorithms. These aspects were introduced technologies for chronic diseases, rehabilitation, and health
in this article, focusing both on physiological knowledge for education. He can be contacted at the Dep. Ing. de Sistemas
a thorough understanding of the problem at hand and differ- y Automtica, Universitat Politècnica de València, C/Camino
ent control engineering approaches proposed in the AP field de Vera, s/n, 46022 Valencia, Spain.
to address these challenges. Sergio Romero-Vivó received the M.Sc. degree in math-
It becomes apparent that a critical component of any AP ematics in 1996 from the Universidad de Valencia, Spain, and
is the limitation of insulin delivery based on some measure the Ph.D. degree in 2001 from the Universidad Politècnica de
of insulin in the body. Plasma insulin concentration, active València, Spain, where he is currently an associate professor.

64  IEEE CONTROL SYSTEMS MAGAZINE  »  FEBRUARY 2018

He is an active member of the University Institute of Multi- [9] C. Homko, A. Deluzio, C. Jimenez, and J. Kolaczynski, “Comparison of insu-
lin aspart and lispro pharmacokinetic and metabolic effects,” Diabetes Care, vol.
disciplinary Mathematics, and his research interests special-
26, pp. 2027–2031, July 2003.
ize in matrix analysis, structural properties of multidimen- [10] J. Brange and A. Vølund, “Insulin analogs with improved pharmacokinetic
sional positive systems, and uncertain dynamical systems profiles,” Adv. Drug Delivery Rev., vol. 35, no. 2-3, pp. 307–335, 1999.
[11] L. Ranganath, F. Schaper, R. Gama, and L. Morgan, “Mechanism of gluca-
and their applications to biomedical control models.
gon-induced nausea,” Clin. Endocrinol., vol. 51, no. 2, pp. 260–261, 1999.
Beatriz Ricarte received the M.Sc. degree in agricultur- [12] A. J. Kowalski, “Can we really close the loop and how soon? Accelerating
al engineering in 1997 and the Ph.D. in mathematics in 2003 the availability of an artificial pancreas: A roadmap to better diabetes out-
comes,” Diabetes Technol. Ther., vol. 11, no. 1, pp. S113–S119, 2009.
from the Universitat Politècnica de València, Spain. Since
[13] A. Haidar, D. Elleri, K. Kumareswaran, L. Leelarathna, J. M. Allen, K.
2001, she has been with the Department of Applied Math- Caldwell, H. R. Murphy, M. E. Wilinska, C. L. Acerini, M. L. Evans, D. B. Dung-
ematics, Universitat Politècnica de València, where she is er, M. Nodale, and R. Hovorka, “Pharmacokinetics of insulin aspart in pump-
treated subjects with type 1 diabetes: Reproducibility and effect of age, weight,
an associate professor. She is a member of the University
and duration of diabetes,” Diabetes Care, vol. 36, no. 10, pp. e173–e174, 2013.
Institute of Multidisciplinary Mathematics, where her re- [14] L. Heinemann, “Variability of insulin absorption and insulin action,” Dia-
search includes matrix analysis and its application to the betes Technol. Ther., vol. 4, no. 5, pp. 673–682, 2002.
[15] G. Scheiner and B. Boyer, “Characteristics of basal insulin requirements by
mathematical modeling of biological systems and control
age and gender in type-1 diabetes patients using insulin pump therapy,” Diabe-
theory. She is a member of the International Linear Algebra tes Res. Clin. Pract., vol. 69, no. 1, pp. 14–21, 2005.
Society and the Spanish Society of Applied Mathematics. [16] M. C. Jolly, R. Hovorka, I. Godsland, R. Amin, N. Lawrence, V. Anyaoku,
D. Johnston, and S. Robinson, “Relation between insulin kinetics and insulin
She is also an editorial board member of the Open Journal
sensitivity in pregnancy,” Eur. J. Clin. Investigat., vol. 33, no. 8, pp. 698–703, 2003.
of Modeling and Simulation. She is currently the secretary in [17] O. U. Petring and H. Flachs, “Inter- and intrasubject variability of gastric
the School of Agricultural Engineering and Environment emptying in healthy volunteers measured by scintigraphy and paracetamol ab-
sorption,” Br. J. Clin. Pharmacol., vol. 29, no. 6, pp. 703–708, 2012.
from the Universitat Politècnica de València, where she is
[18] S. Famulla, U. Hövelmann, A. Fischer, H.-V. Coester, L. Hermanski, M. Kal-
the public authenticator of events and agreements and as- theuner, L. Kaltheuner, L. Heinemann, T. Heise, and L. Hirsch, “Insulin injec-
sists the head of organization and administration tasks. tion into lipohypertrophic tissue: Blunted and more variable insulin absorption
and action, and impaired postprandial glucose control,” Diabetes Care, vol. 39,
José Luis Díez received the M.Sc. degree in industrial
no. 9, pp. 1486–1492. 2016.
engineering in 1995 and the Ph.D. degree in control engi- [19] L. Heinemann, L. Hirsch, and R. Hovorka, “Lipohypertrophy and the artifi-
neering in 2003 from the Universitat Politècnica de Valèn- cial pancreas: Is this an issue?” J. Diabetes Sci. Technol., vol. 8, no. 5, pp. 915–917, 2014.
[20] C. Cobelli, C. Dalla Man, G. Sparacino, L. Magni, G. De, and B. P. Kovatchev,
cia, Spain. He is currently an associate professor and has
“Diabetes: Models, signals, and control,” IEEE Trans. Biomed. Eng., vol. 2, pp.
been teaching since 1995 in the Systems Engineering and 54–96, Jan. 2009.
Control Department, Universitat Politècnica de València, in [21] C. Cobelli, C. Dalla Man, M. G. Pedersen, A. Bertoldo, and G. Toffolo, “Ad-
vancing our understanding of the glucose system via modeling: A perspec-
a wide range of subjects in areas such as automation, linear
tive,” IEEE Trans. Biomed. Eng., vol. 61, no. 5, pp. 1577–1592, Apr. 2014.
systems control theory, digital signal processing, and intel- [22] R. N. Bergman, L. S. Phillips, and C. Cobelli, “Physiologic evaluation of fac-
ligent systems. His research interests include complex sys- tors controlling glucose tolerance in man: Measurement of insulin sensitivity
and beta-cell glucose sensitivity from the response to intravenous glucose,” J.
tems modeling and identification (biomedical, biological,
Clin. Invest., vol. 68, no. 6, pp. 1456–1467, 1981.
energy, and social systems), data mining, clustering tech- [23] R. Hovorka, V. Canonico, L. J. Chassin, U. Haueter, M. Massi-Benedetti, M.
niques, intelligent control, and control education. O. Federici, T. R. Pieber, H. C. Schaller, L. Schaupp, T. Vering, and M. Wilinska,
“Nonlinear model predictive control of glucose concentration in subjects with
type 1 diabetes,” Physiol. Meas., vol. 25, no. 4, pp. 905–920, 2004.
References [24] C. Dalla Man, R. Rizza, and C. Cobelli, “Meal simulation model of the glu-
[1] B. Kovatchev, W. V. Tamborlane, W. T. Cefalu, and C. Cobelli, “The artificial cose–insulin system,” IEEE Trans. Biomed. Eng., vol. 54, no. 10, pp. 1740–1749, 2007.
pancreas in 2016: A digital treatment ecosystem for diabetes,” Diabetes Care, vol. [25] C. Dalla Man, F. Micheletto, D. Lv, M. Breton, B. Kovatchev, and C. Cobe-
39, pp. 1123–1126, July 2016. lli, “The UVA/Padova type 1 diabetes simulator: New features,” J. Diabetes Sci.
[2] Food and Drug Administration. (2016). The artificial pancreas device system. Technol., vol. 8, no. 1, pp. 26–34, Jan. 2014.
[Online]. Available: http://www.fda.gov/medicaldevices/productsandmedical [26] R. N. Bergman, Y. Z. Ider, C. R. Bowden, and C. Cobelli, “Quantitative esti-
procedures/homehealthandconsumer/consumerproducts/artificialpancreas/ mation of insulin sensitivity,” Am. J. Physiol. Endocrinol. Metabol., vol. 236, no. 6,
default.htm pp. E667–E677, June 1979.
[3] R. Hovorka, “Continuous glucose monitoring and closed-loop systems,” Dia- [27] M. E. Wilinska, L. Chassin, C. Acerini, J. M. Allen, D. B. Dunger, and R.
betic Med., vol. 23, no. 1, pp. 1–12, 2006. Hovorka, “Simulation environment to evaluate closed-loop insulin delivery
[4] E. Renard, J. Place, M. Cantwell, H. Chevassus, and C. C. Palerm, “Closed- systems in type diabetes,” J. Diabetes Sci. Technol., vol. 4, no. 1, pp. 132–144, 2010.
loop insulin delivery using a subcutaneous glucose sensor and intraperitoneal [28] S. S. Kanderian, S. A. Weinzimer, and G. M. Steil, “The identifiable virtual
insulin delivery: Feasibility study testing a new model for the artificial pan- patient model: Comparison of simulation and clinical closed-loop study re-
creas,” Diabetes Care, vol. 33, no. 1, pp. 121–127, 2009. sults,” J. Diabetes Sci. Technol., vol. 6, no. 2, pp. 371–379, 2012.
[5] J. Pickup and H. Keen, “Continuous subcutaneous insulin infusion at 25 [29] G. M. Steil, K. Rebrin, C. Darwin, F. Hariri, and M. F. Saad, “Feasibility of
years: Evidence base for the expanding use of insulin pump therapy in type 1 automating insulin delivery for the treatment of type 1 diabetes,” Diabetes, vol.
diabetes,” Diabetes Care, vol. 25, no. 3, pp. 593–598, 2002. 55, no. 12, pp. 3344–3350, 2006.
[6] L. Heinemann and L. Krinelke, “Insulin infusion set: The Achilles heel of [30] C. C. Palerm, “Physiologic insulin delivery with insulin feedback: A con-
continuous subcutaneous insulin infusion,” J. Diabetes Sci. Technol., vol. 6, no. trol systems perspective,” Comput. Methods Program Biomed., vol. 102, no. 2, pp.
4, pp. 954–964, 2012. 130–137, 2011.
[7] S. R. Mudaliar, F. A. Lindberg, M. Joyce, P. Beerdsen, P. Strange, A. Lin, and R. [31] M. E. Wilinska, E. S. Budiman, G. A. Hayter, M. B. Taub, and R. Hovorka,
R. Henry, “Insulin aspart (B28 asp-insulin): A fast-acting analog of human insu- “Method of overnight closed-loop insulin delivery with model predictive con-
lin: Absorption kinetics and action profile compared with regular human insulin trol and glucose measurement error model,” U.S. Patent 201 113 240 855, Jan. 2012.
in healthy nondiabetic subjects,” Diabetes Care, vol. 22, no. 9, pp. 1501–1506, 1999. [32] R. Gondhalekar, E. Dassau, H. C. Zisser, and F. J. Doyle, III, “Periodic-zone
[8] F. Holleman and J. Hoekstra, “Insulin lispro,” New Eng. J. Med., vol. 337, no. model predictive control for diurnal closed-loop operation of an artificial pan-
3, pp. 176–183, 1997. creas,” J. Diabetes Sci. Technol., vol. 7, no. 6, pp. 1446–1460, 2013.

FEBRUARY 2018  «  IEEE CONTROL SYSTEMS MAGAZINE  65

[33] R. A. Harvey, E. Dassau, W. Bevier, D. E. Seborg, L. Jovanovic, F. J. Doyle, tium, “Home use of an artificial beta cell in type 1 diabetes,” New Eng. J. Med.,
and H. C. Zisser, “Clinical evaluation of an automated artificial pancreas using vol. 373, pp. 2129–2140, Nov. 2015.
zone-model predictive control and health monitoring system,” Diabetes Technol. [55] G. M. Steil, C. C. Palerm, N. Kurtz, G. Voskanyan, A. Roy, S. Paz, and F. R.
Ther., vol. 16, no. 6, pp. 348–357, 2014. Kandeel, “The effect of insulin feedback on closed loop glucose control,” Endo-
[34] J. J. Lee, R. Gondhalekar, and F. J. Doyle, “Design of an artificial pancreas crine Res., vol. 96, no. 5, pp. 1402–1408, 2011.
using zone model predictive control with a moving horizon state estimator,” in [56] A. Revert, F. Garelli, J. Picó, H. De Battista, P. Rossetti, J. Vehí, and J. Bondia,
Proc. IEEE 53rd Annu. Conf. Decision Control, 2014, pp. 6975–6980. “Safety auxiliary feedback element for the artificial pancreas in type 1 diabe-
[35] D. de Pereda, S. Romero-Vivó, B. Ricarte, P. Rossetti, F. J. Ampudia-Blasco, tes,” IEEE Trans. Biomed. Eng., vol. 60, no. 8, pp. 2113–2122, 2013.
and J. Bondia, “Real-time estimation of plasma insulin concentration from con- [57] J. L. Ruiz, J. L. Sherr, E. Cengiz, L. Carria, A. Roy, G. Voskanyan, W. V. Tam-
tinuous glucose monitor measurements,” Comput. Methods Biomech. Biomed. borlane, and S. A. Weinzimer, “Effect of insulin feedback on closed loop glucose
Eng., vol. 19, no. 9, pp. 934–942, 2016. control: A crossover study,” J. Diabetes Sci. Technol., vol. 6, no. 5, pp. 1123–1130, 2012.
[36] P. Rossetti, F. J. Ampudia-Blasco, A. Laguna, A. Revert, J. Vehí, J. F. Ascaso, [58] L. M. Huyett, E. Dassau, H. C. Zisser, and F. J. Doyle, III, “Design and evalu-
and J. Bondia, “Evaluation of a novel continuous glucose monitoring-based ation of a robust PID controller for a fully implantable artificial pancreas,” Ind.
method for mealtime insulin dosing—the iBolus—in subjects with type 1 dia- Eng. Chem. Res., vol. 54, no. 42, pp. 10,311–10,321, 2015.
betes using continuous subcutaneous insulin infusion therapy: A randomized [59] B. W. Bequette, “Glucose clamp algorithms and insulin time-action pro-
controlled trial,” Diabetes Technol. Ther., vol. 14, no. 11, pp. 1043–1052, 2012. files,” J. Diabetes Sci. Technol., vol. 3, no. 5, pp. 1005–1013, 2009.
[37] D. Boiroux, T. Aradóttir, M. Hagdrup, N. Poulse, H. Madsen, and J. [60] R. Mantz, H. De Battista, and F. Bianchis, “Sliding mode conditioning for
Jørgensen, “A bolus calculator based on continuous-discrete unscented constrained processes,” Ind. Eng. Chem. Res, vol. 43, no. 26, pp. 8251–8256, 2004.
Kalman filtering for type 1 diabetics,” IFAC Papers OnLine, vol. 48–20, pp. [61] L. Magni, D. Raimondo, L. Bossi, C. Dalla Man, G. D. Nicolao, B. Kovatchev,
159–164, 2015. and C. Cobelli, “Model predictive control of type 1 diabetes: An in silico trial,”
[38] E. Cobry, K. McFann, L. Messer, V. Gage, B. VanderWel, L. Horton, and H. P. J. Diabetes Sci. Technol., vol. 1, no. 6, pp. 804–812, 2007.
Chase, “Timing of meal insulin boluses to achieve optimal postprandial glyce- [62] P. Rossetti, C. Quirós, V. Moscardó, A. Comas, M. Giménez, F. J. Ampudia-
mic control in patients with type 1 diabetes,” Diabetes Technol. Ther., vol. 12, no. Blasco, F. León, E. Montaser, I. Conget, J. Bondia, and J. Vehí. (2016, July). Bet-
3, pp. 173–177, 2010. ter postprandial glucose control with a new closed-loop system as compared
[39] R. Ziegler, G. Freckmann, and L. Heinemann, “Boluses in insulin therapy: with open-loop treatment in patients with type 1 diabetes, Proc. 76th Scientific
A commentary,” J. Diabetes Sci. Technol., vol. 11, no. 1, pp. 1–7, Jan. 2017. Sessions of the American Diabetes Association, New Orleans. [Online]. Available:
[40] D. Luenberger, “Observers for multivariable systems,” IEEE Trans. Automat. http://www.abstractsonline.com/pp8/#!/4008/presentation/39723
Contr., vol. 11, no. 2, pp. 190–197, 1966. [63] F. León-Vargas, F. Garelli, H. De Battista, and J. Vehí, “Postprandial re-
[41] L. Kovács, B. Paláncz, and B. Zoltán, “Design of Luenberger observer for sponse improvement via safety layer in closed-loop blood glucose ­controllers,”
glucose–insulin control via Mathematica,” in Proc. 29th Annu. Int. Conf. IEEE Biomed. Signal Process. Control, vol. 16, pp. 80–87, Feb. 2015.
Engineering Medicine Biology Society, 2007, pp. 624–627. [64] L.-X. Wang, A Course in Fuzzy Systems and Control. Englewood Cliffs, NJ:
[42] C. Eberle and C. Ament, “The unscented Kalman Filter estimates the plasma Prentice-Hall, 1997.
insulin from glucose measurement,” BioSystems, vol. 103, no. 1, pp. 67–72, 2011. [65] E. Atlas, R. Nimri, S. Miller, E. A. Grunberg, and M. Phillip, “MD-Logic ar-
[43] C. Eberle and C. Ament, “Real-time state estimation and long-term model tificial pancreas system: A pilot study in adults with type 1 diabetes mellitus,”
adaptation: A two-sided approach toward personalized diagnosis of glucose Diabetes Care, vol. 33, no. 5, pp. 1072–1076, 2010.
and insulin levels,” J. Diabetes Sci. Technol., vol. 6, no. 5, pp. 1148–1158, 2012. [66] R. Mauseth, I. B. Hirsch, J. Bollyky, R. Kircher, D. Matheson, S. Sanda, and
[44] P. Biswas, S. Bhaumik, and I. Patiyat, “Estimation of glucose and insulin C. Greenbaum, “Use of a ‘fuzzy logic’ controller in a closed-loop artificial pan-
concentration using nonlinear Gaussian filters,” in Proc. IEEE 1st Int. Conf. Con- creas,” Diabetes Technol. Ther., vol. 15, no. 8, pp. 628–633, July 2013.
trol Measurement Instrumentation, 2016, pp. 16–20. [67] R. Kircher, J. Lee, D. Matheson, and R. Mauseth, “Efficacy and computa-
[45] P. Palumbo, P. Pepe, S. Panunzi, and A. D. Gaetano, “Observer-based glu- tional efficiency of the dose safety hypoglycemia prevention module (HPM),”
cose control via subcutaneous insulin administration,” in Proc. 8th Int. Federa- Diabetes Technol. Ther., vol. 17, pp. A97–A97, Feb. 2015.
tion Accountants Symp. Biological Medical Systems, 2012, pp. 107–112. [68] M. C. Riddell, D. P. Zaharieva, L. Yavelberg, A. Cinar, and V. K. Jamnik,
[46] P. Palumbo, P. Pepe, S. Panunzi, and A. D. Gaetano, “Time-delay model- “Exercise and the development of the artificial pancreas: One of the more dif-
based control of the glucose-insulin system, by means of a state observer,” Eur. ficult series of hurdles,” J. Diabetes Sci. Technol., vol. 9, no. 6, pp. 1217–1226, 2015.
J. Control, vol. 18, no. 6, pp. 591–606, 2012. [69] S. A. McAuley, J. C. Horsburgh, G. M. Ward, A. La Gerche, J. L. Gooley, A. J.
[47] P. Palumbo, P. Pepe, S. Panunzi, and A. D. Gaetano, “Recent results on glu- Jenkins, R. J. MacIsaac, and D. N. O’Neal, “Insulin pump basal adjustment for
cose-insulin predictions by means of a state observer for time delay systems,” exercise in type 1 diabetes: A randomised crossover study,” Diabetologia, vol. 59,
in Prediction Methods for Blood Glucose Concentration, H. Kirchsteiger, J. B. Jør- no. 8, pp. 1636–1644, 2016.
gensen, E. Renard, and L. del Re, Eds. New York: Springer, 2015, pp. 227–241. [70] N. Taleb, A. Emami, C. Suppere, V. Messier, L. Legault, J.-L. Chiasson, R.
[48] F. J. Doyle, III, L. M. Huyett, J. B. Lee, H. C. Zisser, and E. Dassau, “Closed- Rabasa-Lhoret, and A. Haidar, “Comparison of two continuous glucose moni-
loop artificial pancreas systems: Engineering the algorithms,” Diabetes Care, toring systems, dexcom Gsss4 platinum and Medtronic paradigm veo enlite sys-
vol. 37, no. 5, pp. 1191–1197, 2014. tem, at rest and during exercise,” Diabetes Technol. Ther., vol. 18, no. 9, pp. 1–7, 2016.
[49] S. Trevitt, S. Simpson, and A. Wood, “Artificial pancreas device systems for [71] K. J. Bell, C. E. Smart, G. M. Steil, J. C. Brand-Miller, B. King, and H. A. Wolp-
the closed-loop control of type 1 diabetes: What systems are in development?” ert, “Impact of fat, protein, and glycemic index on postprandial glucose control
J. Diabetes Sci. Technol., vol. 10, no. 3, pp. 714–723, 2016. in type 1 diabetes: Implications for intensive diabetes management in the con-
[50] C. Ellingsen, E. Dassau, H. Zisser, B. Grosman, M. W. Percival, L. Jovanovic, tinuous glucose monitoring era,” Diabetes Care, vol. 38, no. 6, pp. 1008–1015, 2015.
and F. J. Doyle, III, “Safety constraints in an artificial pancreatic b cell: An imple- [72] S. Laxminarayan, J. Reifman, S. S. Edwards, H. Wolpert, and G. M. Steil,
mentation of model predictive control with insulin on board,” J. Diabetes Sci. “Bolus estimation: Rethinking the effect of meal fat content,” Diabetes Technol.
Technol., vol. 3, no. 2, pp. 536–544, 2009. Ther., vol. 17, no. 11, pp. 1–7, 2015.
[51] H. Zisser, L. Robinson, W. Bevier, E. Dassau, C. Ellingsen, F. Doyle, III, and [73] K. J. Bell, E. Toschi, G. M. Steil, and H. A. Wolpert, “Optimized mealtime in-
L. Jovanovic, “Bolus calculator: A review of four ‘smart’ insulin pumps,” Diabe- sulin dosing for fat and protein in type 1 diabetes: Application of a model-based
tes Technol. Ther., vol. 10, no. 6, pp. 441–444, 2008. approach to derive insulin doses for open-loop diabetes management,” Diabetes
[52] C. Toffanin, H. Zisser, F. J. Doyle, III, and E. Dassau, “Dynamic insulin on Care, vol. 39, no. 9, pp. 1–4, July 2016.
board: Incorporation of circadian insulin sensitivity variation,” J. Diabetes Sci. [74] H. A. Wolpert, A. Atakov-Castillo, S. A. Smith, and G. M. Steil, “Dietary fat
Technol., vol. 7, no. 4, pp. 928–940, 2013. acutely increases glucose concentrations and insulin requirements in patients
[53] F. H. El-Khatib, S. Russell, D. Nathan, R. G. Sutherlin, and E. R. Damiano, “A with type 1 diabetes: Implications for carbohydrate-based bolus dose calculation
bihormonal closed-loop artificial pancreas for type 1 diabetes,” Sci. Transl. Med., and intensive diabetes management,” Diabetes Care, vol. 36, no. 4, pp. 810–816, 2013.
vol. 2, no. 27, pp. 27ra27, 2010. [75] L. A. Gonder-Frederick, J. H. Grabman, B. Kovatchev, S. A. Brown, S. Patek, A.
[54] H. Thabit, M. Tauschmann, J. M. Allen, L. Leelarathna, S. Hartnell, M. E. Basu, J. E. Pinsker, Y. C. Kudva, C. A. Wakeman, E. Dassau, C. Cobelli, H. C. Zisser,
Wilinska, C. L. Acerini, S. Dellweg, C. Benesch, L. Heinemann, J. K. Mader, M. and F. J. Doyle, “Is psychological stress a factor for incorporation into future closed-
Holzer, H. Kojzar, J. Exall, J. Yong, J. Pichierri, K. D. Barnard, C. Kollman, P. loop systems?,” J. Diabetes Sci. Technol., vol. 10, no. 3, pp. 640–646, May 2016, 2016.
Cheng, P. C. Hindmarsh, F. M. Campbell, S. Arnolds, T. R. Pieber, M. L. Evans,
D. B. Dunger, and R. Hovorka, APCam Consortium, and AP@home Consor- 

66  IEEE CONTROL SYSTEMS MAGAZINE  »  FEBRUARY 2018