996
6, JUNE 2006
Run-to-Run Control of Blood Glucose Concentrations
for People With Type 1 Diabetes Mellitus
Camelia Owens, Howard Zisser, Lois Jovanovic, Bala Srinivasan, Dominique Bonvin, and Francis J. Doyle, III*
Abstract—Run-to-run control has been applied to several tradi-
tional batch processes in the chemical industry. The 24-h cycle of
eating meals, measuring blood glucose concentrations, and deliv-
ering the correct insulin bolus, with the goal of achieving the optimal
blood glucose profile, can be viewed in the same spirit as traditional
batch processes such as emulsion polymerization. In this paper, we
aim to exploit the “repetitive” nature of the insulin therapy of people
with Type 1 diabetes. A run-to-run algorithm is used on a virtual di-
abetic patient model to control blood glucose concentrations. The
insulin input is parameterized into the timing and amount of the
dose while the glucose output is parameterized into the maximum
and minimum glucose concentrations. Robustness of the algorithm
to variations in the meal amount, meal timing, and insulin sensi-
tivity parameter is addressed. In general, the algorithm is able to
converge when the meal timing is varied within 40 min. If the meal
size is underestimated by approximately 10 grams (g), the algorithm
is able to converge within a reasonable time frame for breakfast,
lunch, and dinner. If the meal size is overestimated by 20–25 g, the
algorithm is able to converge. When random variations in the meal
timing and the meal amount are introduced, the variation on the
output variables, and , scales according to the amount
of variation allowed. Along with this, the insulin sensitivity of the
virtual patient model is varied. The algorithm is robust for differ-
ences in insulin sensitivity less than 50% of the nominal value.
Index Terms—Artificial pancreas, control, diabetes, glucose,
run-to-run.
I. INTRODUCTION
D IABETES mellitus affects over 100 million individuals
worldwide, and this number is expected to double by 2010
[1]. In the US, the estimated healthcare costs of the 12 million
affected is estimated to be 136 billion dollars annually [2]. Di-
abetes is a disorder of the metabolism characterized by the in-
ability of the pancreas to secrete sufficient insulin. When glu-
cose levels remain high for extended periods of time (hyper-
glycemia) the patient is at risk for neuropathy, nephropathy, and
other long-term vascular complications. According to the Di-
abetes Control and Complications Trial (DCCT), intensive in-
sulin therapy can help reduce the risks of developing complica-
tions [3]. Intensive insulin therapy requires three or more insulin
Manuscript received March 3, 2005; revised October 1, 2005. This work
was supported in part by Roche Pharmaceuticals, Indianapolis, IN, and in part
by the National Institutes of Health (NIH), Bethesda, MD, under Grant R01-
DK068706-01/R01-DK068663-01 and Grant R21-DK69833. Asterisk indicates
corresponding author.
C. Owens is with the Department of Chemical Engineering, University of
Delaware, Newark, DE 19716 USA.
H. Zisser and L. Jovanovic are with the Sansum Diabetes Research Institute,
Santa Barbara, CA 93105 USA.
B. Srinivasan and D. Bonvin are with the Laboratoire d’Automatique, EPFL,
CH-1015 Lausanne, Switzerland.
F. J. Doyle, III is with the Department of Chemical Engineering, University
of California, Santa Barbara, CA 93106 USA (doyle@engineering.ucsb.edu).
Digital Object Identifier 10.1109/TBME.2006.872818
0018-9294/$20.00 © 2006 IEEE
IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING, VOL. 53, NO.
injections per day or the use of an external insulin infusion pump
in order to minimize the risk for complications.
The development of external insulin infusion pumps, along
with the introduction of rapid acting insulin analogs has greatly
aided in making intensive insulin therapy feasible. The efficacy
of the insulin therapy is quantified by measurement of the per-
centage of glycosylated hemoglobin in the bloodstream [Akaike
information criterion(AIC)]. Values less than 6% are represen-
tative of a nondiabetic patient; whereas, higher percentages are
indicative of hyperglycemia [4].
Several algorithms have been developed to optimize the in-
sulin therapy of people with diabetes, and those related to this
study are highlighted in Table I. Jovanovic et al. proposed sev-
eral insulin delivery logic rules for an Neutral Protamine Hage-
dorn/Regular insulin system, a Lente insulin system and a con-
stant subcutaneous insulin infusion system [5]. Heuristics were
determined to account for food intake, weight loss, exercise,
childhood, adolescence, and pregnancy. Chanoch et al.. eval-
uated the use of a pocket computer to aid in determining the
proper insulin dose for people with diabetes [6]. Patient specific
parameters such as weight, gender, and physical activity along
with carbohydrate content of meals and blood glucose measure-
ments were used by the algorithm to optimize the insulin dose.
With use of the pocket computer, values decreased from
7.2% to 5.8%, with a resulting mean blood glucose of 130 mg/dl
versus 160 mg/dl before the study. Schiffren et al. also explored
the use of computer algorithms for insulin dose adjustment [7].
Seven patients with type 1 diabetes were recruited to use the
computer algorithm for an eight-week period. During the con-
trol period, the mean blood glucose concentration for all patients
was 178, 187, 208, and 207 mg/dl for breakfast, lunch, dinner,
and bedtime snack premeal measurements. Upon completion of
the algorithm phase of the study, these values decreased to 116,
110, 148, and 135 mg/dl, respectively, for the same premeal
measurements. Chiarelli et al. investigated the use of a computer
algorithm in children [8]. Their results showed fewer episodes
of hypoglycemia in the computer-assisted group. Other studies
have employed the use of computer algorithms to maintain nor-
moglycemia in patients with hyperglycemia, insulin-resistance,
and type 2 diabetes [9]. Albisser highlighted several algorithms
that have been developed to aid in the progress of self-manage-
ment of diabetes for better blood glucose control [10].
The repetitive nature of an intensive glucose control therapy
regimen (i.e, taking blood glucose measurements, eating meals,
and delivering the correct bolus of insulin) is analogous to
industrial batch processes, such as semiconductor manufac-
turing [11]. On a day-to-day basis, the patient is performing the
same tasks to adjust their insulin therapy in order to maintain
blood glucose concentrations within appropriate boundaries.
OWENS et al.: RUN-TO-RUN CONTROL OF BLOOD GLUCOSE CONCENTRATIONS FOR PEOPLE WITH TYPE 1 DIABETES MELLITUS
TABLE I
SUMMARY OF LITERATURE ALGORITHMS
In run-to-run (or batch) control, information about product
quality from the previous run is used to determine the input for
the next run [12]. In this study, the daily cycle of the patient
with diabetes is treated as a batch process. For a person with
diabetes, each day represents a single run and, from the blood
glucose measurements obtained at the end of the day, the insulin
therapy can be adjusted for the next day.
The similarities between the protocol for a person with type
1 diabetes and a batch chemical reactor recipe which motivate
the application of the run-to-run technique include:
• the recipe (24-h cycle) for a patient consists of a repeated
meal protocol (typically three meals) with some variance
on meal type, timing, and duration;
• there are key quality variables or time points of the blood
glucose profile that can be measured, reflecting the impact
of the input insulin therapy regimen;
• since each patient is different, the insulin therapy needs to
be adjusted individually.
Hence, the run-to-run algorithm aims to employ readily avail-
able data to improve the insulin therapy of the patient.
997
II. VIRTUAL PATIENT MODELS
A number of virtual patient models for diabetes exist in the lit-
erature. The work that is described in the following sections em-
phasizes results evaluated on the Bergman model. In addition,
the Sorensen and Automated Insulin Dosage Advisor (AIDA)
models have been investigated, and those results are highlighted
to emphasize the applicability of the run-to-run algorithm to dif-
ferent virtual patient systems.
A. Bergman Model
The Bergman model is a three compartment minimal model
of glucose and insulin dynamics [13]. This third-order model is
comprised of a glucose compartment, , a remote insulin com-
partment, , and an insulin compartment, [14]. The remote in-
sulin compartment mediates glucose uptake within the glucose
space to the peripheral and hepatic tissues. The insulin distribu-
tion space combines the sinks and sources of insulin production
and consumption into a single pool. While the Bergman model is
simplistic in nature, it is able to capture certain dynamics of the
diabetic patient system. The insulin dynamics of the Bergman
model are driven by an intravenous infusion of insulin to the
system. The meal model is the same as that proposed in [15].
B. Sorensen Model
The Sorensen model is a 6 compartment model [16], [17]. The
compartments are physiological representations of the brain,
heart and lungs, liver, gut, and kidney peripheral tissue. Within
the brain and peripheral tissue, both the dynamics within the in-
terstitial fluid and capillary fluid are detailed. The glucose meal
disturbance is ingested directly into the gut accompanied by in-
travenous delivery of insulin and arterial blood glucose mea-
surements. This twenty-first-order model describes glucose, in-
sulin, and glucagon dynamics of the diabetic patient. Similar
to the Bergman model, the Sorensen model relies on an intra-
venous infusion of insulin to drive the insulin dynamics of the
system. Due to the incorporation of glucagon’s action to pro-
mote hepatic output of glucose into the model, blood glucose
levels do not remain in the hypoglycemic range for extended
periods of time.
C. AIDA Model
The AIDA is a fourth-order three-compartment model of in-
sulin and glucose dynamics. In contrast to the Bergman model,
the insulin dynamics of the model are driven by subcutaneous
injection of insulin [15]. The AIDA model was first proposed
as an educational tool. Hence, the model does attempt to model
the effects of different meal sizes on the rate of gastric emptying
in the system. The AIDA model was also originally designed to
reflect the use of several different insulin analogs on the insulin
therapy of an individual. In doing so, subcutaneous transport
dynamics were accounted for to simulate the effect of a subcu-
taneous injection of insulin to the body. The AIDA model was
developed in the spirit of the minimal model approach, resulting
in few patient specific parameters.
D. Model and Parameters Used in this Study
In this discussion and simulation results, the Bergman model
is employed (the results with the Sorensen and AIDA model re-
998
sults are highlighted in a subsequent table). With the Bergman
virtual patient it is feasible to simulate a 24-h day of meal in-
takes (breakfast, lunch, and dinner) and the corresponding meal
related bolus doses for the diabetic patient. For each meal, the
desired measurements are the maximum glucose concentration,
, and the minimum glucose concentration, . Each
meal requires the appropriate insulin bolus to mediate rising
blood glucose concentrations. Therefore, the timing, , and the
amount, , of the insulin dose must be adjusted to yield the de-
sired and value for each meal.
For clinical application of the run-to-run algorithm, and
may be translated into appropriate markers for quality of
glucose regulation that occur at fixed time points (to limit mea-
surement requirements). For example, with insulin aspart, the
maximum insulin concentration occurs 60 min after the start
of the meal, , and the 90-min blood glucose concentration,
, is lower than . These time points will depend upon both
the peak effectiveness and the elimination dynamics of the in-
sulin used. In general, the run-to-run algorithm only requires a
“quality index” for the feedback calculation, so one can realize
this in a number of ways, including measurements taken at fixed
time points (which has a minimum burden for data).
III. GENERIC RUN-TO-RUN CONTROL
A. State of the Art
Arimoto et al. proposed the “betterment process” concept
as a measure of making the inputs for certain systems, such
as mechanical robots, yield better outputs [18]. The concept
can be applied to multi-input–multi-output systems that have
a repetitive or cyclic behavior. The convergence properties of
this iterative learning control algorithm was studied by Amann
et al. [19]. Using an inverse process model, Lee et al. derived a
feedback-assisted iterative learning control scheme to attain the
maximum convergence rate [20]. This approach was evaluated on
a simulated bench-scale batch reactor, to demonstrate the ability
of the algorithm to control reactor temperature. For time-varying
linear constrained systems, prevalent in the chemical process
control area, Lee et al. proposed a model-based iterative learning
control scheme with a quadratic performance objective [21]. The
feasibility of this algorithm in the presence of disturbances and
noise was tested on several numerical examples. Lee et al. have
also combined model predictive control with iterative learning
control [12]. In doing so, the algorithm can accommodate both
within-run and run-to-run errors. Srinivasan et al. proposed
a constraint control scheme in the run-to-run framework to
optimize the operation of a batch process and this approach is
investigated in the present study for glucose control [22]–[24].
B. Run-to-Run Control Algorithm
The general run-to-run control algorithm is given below:
1) Parameterize the input profile for run , , as . blood coagulation [25]. In that work, they employed an expo-
Also, consider a sampled version, , of the output, , nentially weighted moving average (EWMA) controller, using
such that the input parameter vector, , and the controlled
variable vector have the same dimension. This gives
(1)
IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING, VOL. 53, NO. 6, JUNE 2006
2) Choose an initial guess for , .
3) Complete the run using the input corresponding to
. Determine from the measurements .
4) Update the input parameters using
(2)
where is an appropriate gain matrix and represents
the reference values to be attained. Set and
repeat steps 3)–4) until convergence.
The convergence of the run-to-run algorithm can be determined
by analyzing the dynamics of the error for the closed-loop
system, where the error is . For this analysis, a
linearized version of the system equations (1) is used
(3)
where is the sensitivity matrix. Using (2) and
(3), the linearized error dynamics are easily derived through the
following manipulations:
(4)
Hence, the eigenvalues of the matrix should be within
the unit circle for the algorithm to converge. Also, note that con-
vergence of the algorithm is determined by the controller gains
and the sensitivity matrix .
IV. APPLICATION OF RUN-TO-RUN CONTROL TO
DIABETES MANAGEMENT
The application of run-to-run control to help manage diabetes
provides an opportunity to properly adjust the current insulin
therapy of the patient. The run-to-run algorithm exploits the
repetitive nature of the insulin therapy regimen of the diabetic
patient. For each meal, an update law is prescribed to correct the
insulin bolus amount and timing for the next day. For the present
application, and correspond to the insulin and glu-
cose profiles, respectively, where is the continuous time vari-
able and the run number. The run index represents the rep-
etition of the 24-h daily routine of breakfast, lunch, and dinner
meals.
The closest related work to the one proposed in this paper
was performed in parallel by Good et al. in which they applied
run-to-run control to optimize drug dosage in order to regulate
a linear process model, and a variant of the optimizing adap-
tive quality controller (OAQC), using a second-order Hammer-
stein model, to optimize the anticoagulant medication of the pa-
tient. The methodology was evaluated using simple numerical
models.
OWENS et al.: RUN-TO-RUN CONTROL OF BLOOD GLUCOSE CONCENTRATIONS FOR PEOPLE WITH TYPE 1 DIABETES MELLITUS 999

A. Definition of Variables
Three different meals are considered over a 24-h period as the
basic cycle which is repeated.
• Manipulated variables: Timing of three insulin injections
for breakfast, , lunch, , and dinner, , and quan-
tity of insulin injection for breakfast, , lunch, , and
dinner, , i.e., .
• Controlled variables: Maximum value of glucose attained
after breakfast, , lunch, , and dinner, ,
and minimum value of glucose attained after the peaking
for breakfast, , lunch, , and dinner, , i.e.,
. The max-
imum and minimum glucose concentrations are selected as
convenient scalar measures of performance for a particular
insulin regime. As before, pragmatic considerations may Fig. 1. Plot of maximum glucose concentration for the breakfast meal
dictate that time points are used to approximate maximum as a function of the insulin amount and the insulin timing.
and minimum values. , , and . The
dark solid curve represents as calculated from the nonlinear model. The
The insulin sensitivity matrix of the person with diabetes, , is light grey plane represents the dynamics of from a local estimation of
characterized to aid in the selection of the controller gain matrix. the insulin sensitivity of the patient. The clear plane represents the dynamics of
Here, the patient sensitivity reflects the breadth of output vari- from a global estimation of the insulin sensitivity of the patient.
able responses, , as a function of changes in the input variable,
. This does not characterize the insulin sensitivity of the virtual
by varying the timing of the insulin bolus and the amount
patient in the traditional sense, as the transient glucose profile is
over a wider range ( 40%). Three sets of sample points were
not analyzed completely, only the desired output variables.
generated corresponding to the three meals. Each set had points
is composed of the sensitivity values for breakfast, lunch,
from a two-dimensional grid (18 18) where the axes were the
and dinner. The elements of the matrix represents the sensi-
insulin bolus quantity and timing of injection corresponding
tivity of the output in response to variations in the input.
to the meal considered. Thus, the number of points generated
For example, denotes the change in the maximum glucose
was . After obtaining the surface, a
concentration for breakfast to a unit change in the timing
least-squares fit was performed using the MATLAB fmincon
of the breakfast bolus . Accordingly, corresponds to a
routine with the cost function given below to determine the
noncausal relationship between the timing of the lunch bolus
corresponding best-fit sensitivity
and and, thus, is zero. The patient sensitivity was de-
termined, separately, using a local estimation, and global esti-
mation technique.
Local Estimation of Patient Sensitivity: The local insulin sen- (5)
sitivity of the virtual patient, which corresponds to approxi-
mating the derivative around the current operating The fit from the global determination of the insulin sensitivity
point, was computed numerically using a finite difference ap- is given by the clear plane in Fig. 1. Similar to the response of
proximation. The timing of the bolus was varied 5 min from the nonlinear model, increases as the bolus timing moves
nominal and the amount of the bolus was varied approximately away from the meal peak.
5–10% of the nominal for all three meals. The operating point It is important to note that in a clinical setting (as opposed
used was , , and to the present simulation study), there are many addition fac-
. The profile of the local estimation of the insulin tors that will complicate the identification of patient sensitivity
sensitivity can be observed by the light grey plane in Fig. 1. for the proposed algorithm. In such circumstances, it is possible
Global Estimation of Patient Sensitivity: Fig. 1 also shows to incorporate the intuition and knowledge of trained medical
the maximum blood glucose concentration for the breakfast professionals in determining a suitable sensitivity measure for
meal as a function of the insulin dose amount and insulin dose a given patient. Such an approach is described in a companion
timing covering a wider range ( 40%). The dark curved surface paper that details the findings of a clinical evaluation of the pro-
represents the actual response of the virtual patient model. As posed algorithm [26].
the amount of the bolus increases, decreases slightly, and
as the timing of the bolus moves forward well beyond its impact B. Controller Design
on the breakfast meal, begins to increase asymptotically In principle, a multivariable controller should be designed to
up to the open-loop concentration. This is done for all maintain the components of at their reference values . How-
three meals and all six output variables. As the timing of the ever, the sensitivity matrix shows that the effects of the inputs
insulin bolus moves past the peak, the concentration for are relatively decoupled for the case studies considered [as de-
breakfast increases. These nonlinearities necessitate a global termined by the relative gain array (RGA)]. The RGA for the
estimation of the patient sensitivity, which was determined Bergman model subject to breakfast (20 g, 8 am), lunch (40 g,
1000
12 noon), and dinner (60 g, 5 pm) is calculated below (based on
sensitivity derivation described in Section IV-A)
First of all, the RGA suggests that there is no coupling be-
tween the meals. For each meal, there are two outputs (
and ) and two inputs ( and ). Again, the RGA suggests
pairing, for each meal, the timing of the insulin bolus with
the maximum glucose concentration and the quantity
of the insulin bolus with the minimum glucose concentration
. This result may be distinct to the model employed in this
study, and other models may lead to more interconnected rela-
tionships between meals. However, all three of the models con-
sidered in this study, and the combined medical expertise of the
co-authors, strongly suggests that, for quick-acting insulin, the
meal responses will be largely decoupled. This is, indeed, an
approximation, but one that is typical in many process control
applications.
The appropriate controller gain matrix corresponds to a
diagonal matrix as
where is the controller gain for the timing control loop of
the meal [breakfast (B), lunch (L), and dinner (D)] and the
controller gain for the corresponding quantity control loop.
The important variables that need to be designed are the initial
guesses . In addition to and (desirable bounds
on glucose), there exist the hard limits and (absolute
bounds that, if violated, lead to serious medical problems). The
values should be chosen such that these hard constraints are
satisfied for all cases [22]. Also, the history of the patient can
facilitate the selection of initial guesses.
For all three meals, the desired reference value for was
set to 140 mg/dl and the target for to 90 mg/dl. The con-
troller gains and were computed in accordance to the
insulin sensitivity of the virtual patient. Though the coupling
is low, it might so happen that a fast decrease in may
cause an unacceptable response in and vice versa. Thus,
the gains reflect a compromise between speed and accuracy.
C. Simulation Results
Fig. 2 shows the convergence of the run-to-run control algo-
rithm for the lunch meal of the Bergman model. Specifically,
the convergence properties of for lunch are shown for the
linear model [Fig. 2(top)], employing the patient sensitivity, and
the full nonlinear model [Fig. 2(bottom)]. As can be seen from
IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING, VOL. 53, NO. 6, JUNE 2006
Fig. 2. Plot of for the lunch meal as a function of the pole location. (top)
Convergence speed of linear Bergman model as the pole moves along the real
axis: (dashed-dotted line) slow convergence, ; (dashed line) fast con-
vergence, ; (solid line) fast convergence with oscillations, ;
(dotted line) diverging with oscillations, . (bottom) Corresponding
convergence speed of the nonlinear Bergman model. As the controller gain is in-
creased, the virtual patient model eventually diverges, as shown from the dotted
line.
the top plot, the linear model quickly converges when the poles
lie within the unit circle. As these poles move away from the
origin in the left half plane, convergent oscillations develop and
eventually, the system diverges as the poles lie outside the region
of stability denoted by the unit circle. The bottom plot shows the
response of the nonlinear model as the control loop is tuned. As
the controller tuning becomes more aggressive, the system be-
gins to oscillate and eventually diverges. Table II lists the patient
sensitivity values for the Bergman model employed to evaluate
the error dynamics of the system as determined from the global
estimation of the insulin sensitivity of the virtual patient model.
The corresponding controller gains are given in Table III. The
control loops were tuned to converge as quickly as possible with
minimum oscillation in the virtual patient model response. In
Table III, the differences in the controller gains reflects the non-
linearity and the interactions within the system. All of the sub-
sequent results are based on these controller tunings. To demon-
strate the ability of the run-to-run control algorithm to regulate
blood glucose concentrations, the virtual patient model is sub-
ject to a 20-g breakfast, 40-g lunch, and 60-g dinner, with an
initial guess for the insulin regimen. The meal times were fixed
at 8 am, 12 noon, and 5 pm.
Fig. 3 shows the run-to-run algorithm’s ability to control
blood glucose concentrations and converge to the desired
preprandial goal of 90 mg/dl and postprandial goal of 140
mg/dl within 10 days. The dashed-dotted line on the top plot
shows the open-loop blood glucose concentration in the absence
of insulin boluses. Glucose levels rise above 200 mg/dl and the
patient remains in a hyperglycemic state for an extended period
of time. On day 1, given by the solid line, the initial guesses
for the insulin bolus amount and timing can be seen from the
bottom plot. The corresponding blood glucose concentration
OWENS et al.: RUN-TO-RUN CONTROL OF BLOOD GLUCOSE CONCENTRATIONS FOR PEOPLE WITH TYPE 1 DIABETES MELLITUS
TABLE II
TABLE OF INSULIN SENSITIVITY VALUES FOR VIRTUAL PATIENT MODEL
TABLE III
TABLE OF CONTROLLER GAINS FOR VIRTUAL PATIENT MODEL
Fig. 3. Plot of convergence of run-to-run algorithm within 10 days. The
virtual patient is subject to a 20-g breakfast, 40-g lunch, and 60-g dinner. The
dotted lines represent the target boundaries of 90 and 140 mg/dl for
and , respectively. (top) Blood glucose concentration as a function of
time: (dotted-dashed line) open-loop blood glucose concentration with no
bolus; (solid line) blood glucose concentration on day 1 from initial guesses
of insulin bolus amount and timing; (dashed line) blood glucose concentration
on the second day using the run-to-run algorithm; (thick solid line) blood
glucose concentration after 10 days of the run-to-run algorithm. (bottom)
Corresponding insulin injection input for the run-to-run algorithm: (solid line)
initial guess for insulin bolus injection; (dashed line) insulin bolus injection for
day 2; (thick solid line) insulin bolus injection for day 10.
still remains outside of the desired boundaries for and
. On day 2, the algorithm continues to compute the op-
timal insulin bolus amount and timing for each meal, as shown
by the dashed line in the top and bottom plots. Since was
largely out of the zone, the timing of the meal bolus changes
the most on day 2. Eventually on day 10, represented by the
thick solid line, and for each meal come within the
desired bounds.
Table IV compares the performance of the Bergman model
with the AIDA and Sorensen virtual patient models. The AIDA
model was subject to a 20-g breakfast, 50-g lunch, and 65-g
dinner. The Sorensen model was subject to a 20-g breakfast,
40-g lunch, and 70-g dinner. The target values for the AIDA
1001
TABLE IV
TABLE OF MODEL COMPARISON OF SETTLING TIME
model were 80 to 140 mg/dl while those for the Sorensen model
were 70 to 140 mg/dl. All three models were tuned, without
severe oscillations, to converge within a two-week period.
Different operating conditions were chosen for each model as
the meal and insulin dynamics manifest themselves differently
in the glucose profile of each virtual patient model (e.g., sub-
cutaneous versus intravenous administration). Therefore, the
settings were chosen so that each model had similar glucose
excursions.
V. ROBUSTNESS OF RUN-TO-RUN ALGORITHM
There are several assumptions underlying the run-to-run al-
gorithm.
1) The timing of the meals remains the same from day to day.
2) The size of the meals remains the same from day to day.
3) Any patient variability that may occur, i.e. insulin sensi-
tivity, is negligible from day to day.
For being applicable to people with Type 1 diabetes, this algo-
rithm must be able to remain effective in the presence of sig-
nificant variability. Therefore, the robustness of the run-to-run
algorithm to variations in the aforementioned items must be ad-
dressed. In evaluating the robustness of the algorithm, the ac-
curacy of the characterization of the patient sensitivity, , is
called into question. In essence, this is reflected in the output
variables,
(6)
where represents the uncertainty in the patient. For the fol-
lowing results, employing the Bergman model, the settling time
is defined as the minimum number of days required for
and for all three meals to reach and
.
A. Variation in Meal Timing
In this case study, the nominal meal timing was fixed at 8
am, 12 noon, and 5 pm for breakfast, lunch, and dinner, re-
spectively. Keeping all other variables constant (initial condi-
tions, controller tunings, meal amounts, and patient insulin sen-
sitivity), the timing of the meals was simultaneously varied more
than 1 h from nominal, as shown in Fig. 4.
1002
Fig. 4. Plot of settling time versus variation in meal timing for breakfast, lunch,
and dinner. The thick dashed line represents the desired window for convergence
within 10 days. The (x) represents the maximum blood glucose concentration
and the (o) represents the minimum blood glucose concentration. Values for
breakfast are denoted by the solid lines, values for lunch are denoted by the
dotted lines, and values for dinner are denoted by the dashed lines.
Fig. 5. Plot of standard deviation of the output variables and for
breakfast, lunch, and dinner versus standard deviation in the timing of the meal.
The (x) represents the maximum blood glucose concentration and the (o) rep-
resents the minimum blood glucose concentration. Values for breakfast are de-
noted by the solid lines, values for lunch are denoted by the dotted lines, and
values for dinner are denoted by the dashed lines.
For all cycles, the deviation in the meal timing remained con-
stant. For variations less than 20 min, all meal parameters con-
verge within a window of 5 to 10 days. As the deviation from
nominal increases, the for lunch begins to diverge quickly.
This may be due to the interactions in the system from the break-
fast meal.
To further investigate the robustness of the algorithm, random
variations in the meal timing were introduced from day to day.
Random values were selected from a normal distribution that
has been truncated to 3 standard deviations to avoid outliers.
Again, all other parameters remained at their nominal values.
Fig. 5 displays the standard deviation of the output variables,
IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING, VOL. 53, NO. 6, JUNE 2006
Fig. 6. Plot of settling time versus variation in meal amount for breakfast,
lunch, and dinner. The thick dashed line represents the desired window for con-
vergence speed of within 10 days. The (x) represents the maximum blood glu-
cose concentration and the (o) represents the minimum blood glucose concen-
tration. Values for breakfast are denoted by the solid lines, values for lunch are
denoted by the dotted lines, and values for dinner are denoted by the dashed
lines.
and , for breakfast, lunch, and dinner as a function
of the standard deviation of the variation in the meal timing.
The results represent the average response over 30 realizations
of random meal timing. For these simulations, the meal timing
is randomly varied, from run to run positively and negatively,
from the nominal setting. As more variation is introduced into
the system, the variation in the output variables increases. The
variation of for all three meals is higher than the cor-
responding values with the variation on the dinner meal
being the largest. This may be due to the size of the dinner meal
in comparison to breakfast and lunch and the initial conditions
that were set for the nominal case. However, overall, the algo-
rithm remains robust to these variations.
B. Variation in Meal Amount
In this case study, the nominal meal amount was fixed at 20,
40, and 60 grams (g) for breakfast, lunch, and dinner, respec-
tively. Again, all other settings remained constant to observe
the convergence of the algorithm in the presence of variations
on the meal amount. Fig. 6 shows a plot of the settling time
as the meal amount varies from its nominal value. For meal
amounts approximately 10 g less than nominal, the algorithm di-
verges. The initial insulin dosage over-boluses the patient as the
meals become smaller and hypoglycemia occurs. Consequently,
this moves the timing of the bolus in the other direction (gain
change) as becomes significantly less than . On
the other hand, as the meal gets larger, the initial meal bolus
significantly under doses the patient but does not change the di-
rection of the bolus timing as remains significantly greater
than . Once the meal becomes more than 25–30 g larger
than nominal, the algorithm begins to diverge. This indicates
that there is some flexibility in meal amount that the algorithm
can tolerate. To further test the robustness of the algorithm to
OWENS et al.: RUN-TO-RUN CONTROL OF BLOOD GLUCOSE CONCENTRATIONS FOR PEOPLE WITH TYPE 1 DIABETES MELLITUS
Fig. 7. Plot of standard deviation of the output variables, and
for breakfast, lunch, and dinner versus standard deviation in the amount of the
meal. The (x) represents the maximum blood glucose concentration and the (o)
represents the minimum blood glucose concentration. Values for breakfast are
denoted by the solid lines, values for lunch are denoted by the dotted lines, and
values for dinner are denoted by the dashed lines.
changes in the meal amount, random variations were introduced
to the size of the meal, from day to day, using the previously de-
scribed procedure. Fig. 7 shows a plot of the standard deviation
of and as a function of the percent variation in the
meal amount. As before, 30 random realizations of meal amount
were averaged to generate the results. As the size of the meal
is allowed to randomly vary from 10% to 40%, there is a
subsequent increase observed in the variation of the output vari-
ables. The values show the largest variation with dinner
being the highest. Again, the trend of increasing variation in
output variables with increasing variations introduced into the
system is expected. Depending upon the starting point for the
run-to-run control algorithm, the degree of the increase may
vary slightly.
C. Variation of Insulin Sensitivity Parameter
During the day and from day-to-day, the patient may expe-
rience changes in their insulin sensitivity. In this context, sen-
sitivity refers to the classical biomedical definition, which in-
fluences the value of the run-to run sensitivity, , employed in
this work. Fig. 8 gives an example of how convergence speed
may vary as a function of the insulin sensitivity. The insulin
sensitivity parameter for the Bergman model, , is varied ex-
tensively within the range of values described in [13]. As be-
fore, 30 random realizations of the insulin sensitivity were av-
eraged to generate the results. The extremum values represent
the patient becoming highly sensitive to insulin and insensitive
to insulin, respectively. In general, within 50% of the nom-
inal insulin sensitivity, the algorithm is able to converge within
a reasonable time frame. This indicates that the algorithm can
tolerate gradual changes in the insulin sensitivity of the patient
or mildly abrupt changes in the insulin sensitivity of the indi-
vidual. As the patient insulin sensitivity increases, the insulin
has a larger impact on the glucose profile. Consequently, it will
1003
, Fig. 8. Plot of settling time versus variation in insulin sensitivity parameter .
The thick dashed line represents the desired window for convergence speed of
within 10 days. The (x) represents the maximum blood glucose concentration
and the (o) represents the minimum blood glucose concentration. Values for
breakfast are denoted by the solid lines, values for lunch are denoted by the
dotted lines, and values for dinner are denoted by the dashed lines.
continue to converge quickly because the correction in the up-
date laws will get smaller until the initial guess is too much for
the meal. Hence, the poles of the error dynamics will
be pushed outside of the unit circle. On the other hand, as the
insulin sensitivity of the individual decreases with the initial
guess, the correction on timing will continue to grow because
is less affected by the insulin bolus.
VI. CONCLUSION
The feasibility of a run-to-run control algorithm to effectively
normalize blood glucose concentrations is demonstrated. By pa-
rameterizing the insulin input ( and ) and the glucose output
( and ) for breakfast, lunch, and dinner, appropriate
control laws are derived. To tune the controller, the insulin sen-
sitivity of the virtual patient model is determined by examining
and as a function of the insulin dose timing and
amount. To address the questions of robustness, variations in the
meal timing, meal amount, and insulin sensitivity of the virtual
patient model have been analyzed.
In general, the algorithm is able to converge when the meal
timing is varied within 40 min. If the meal size is underesti-
mated by approximately 10 g, the algorithm is still able to con-
verge within a reasonable time frame for breakfast, lunch, and
dinner. If the meal size is overestimated by 20–25 g, the algo-
rithm is able to converge. When random variations in the meal
timing and the meal amount are introduced, the variation on
the output variables, and , scales according to the
amount of variation allowed. Along with this, the insulin sen-
sitivity of the virtual patient model is varied. The algorithm is
robust for differences in insulin sensitivity less than 50% of
the nominal value.
These results demonstrate that the run-to-run control algo-
rithm is able to handle day-to-day variations that may occur in
a patient with type 1 diabetes. Moreover, the algorithm aims to
properly adjust the current insulin therapy of the individual by
1004
trying to make the same decisions as the physician. With knowl-
edge of the insulin sensitivity of the patient and the appropriate
glucose references, the run-to-run control algorithm can help
manage diabetes. A clinical trial evaluating the efficacy of the al-
gorithm to control glucose concentrations for people with Type
1 diabetes has recently been completed, and these results sug-
gest the algorithm described here can prescribe appropriate meal
insulin doses [26].
REFERENCES
[1] T. Mandrup-Poulsen, “Recent advances: diabetes,” Br. Med. J., vol.
316, pp. 1221–1225, Apr. 1998.
[2] American Diabetes Association, “Standards of medical care for pa-
tients with diabetes mellitus,” Diabetes Care, vol. 26, pp. S33–S50,
2003.
[3] Diabetes Control and Complications Trial Research Group, “The effect
of intensive treatment of diabetes on the development and progression
of longterm complications in insulin-dependent diabetes mellitus,” N.
Engl. J. Med., vol. 329, pp. 977–986, Sep. 1993.
[4] American Diabetes Association, “Standards of medical care for pa-
tients with diabetes mellitus,” Diabetes Care, vol. 26, pp. S33–S50,
Jan. 2003.
[5] L. Jovanovic and C. Peterson, “Home blood glucose monitoring,”
Comp. Ther., vol. 8, pp. 10–20, Jan. 1982.
[6] L. Chanoch, L. Jovanovic, and C. Peterson, “The evaluation of a pocket
computer as an aid to insulin dose determination by patients,” Diabetes
Care, vol. 8, pp. 172–176, Mar./Apr. 1982.
[7] A. Schiffrin, M. Mihic, B. Leibel, and A. Albisser, “Computer-assisted
insulin dosage adjustment,” Diabetes Care, vol. 8, pp. 545–552, Nov./
Dec. 1985.
[8] F. Chiarelli, S. Tumini, G. Morgese, and A. Albisser, “Controlled study
in diabetic children comparing insulin-dosage adjustment by manual
and computer algorithms,” Diab. Care, vol. 13, pp. 1080–1084, 1990.
[9] C. Mao, M. Riegelhuth, D. Gundy, C. Cortez, S. Melendez, and E.
Ipp, “An overnight insulin infusion algorithm provides morning nor-
moglycemia and can be used to predict insulin requirements in nonin-
sulin-dependent diabetes mellitus,” J. Clin. Endo. Metab., vol. 82, pp.
2466–2470, Apr. 1997.
[10] A. Albisser, “Toward algorithms in diabetes self-management,” Diab.
Tech. Ther., vol. 5, pp. 371–373, 2003.
[11] J. Moyne, E. del Castillo, and A. Hurwitz, Run-to-Run Control in Semi-
conductor Manufacturing. New York: CRC Press, 2001.
[12] K. Lee, I. Chin, and H. Lee, “Model predictive control technique com-
bined with iterative learning for batch processes,” AIChE J., vol. 45,
pp. 2175–2187, 1999.
[13] R. Bergman, L. Phillips, and C. Cobelli, “Physiologic evaluation of
factors controlling glucose tolerance in man,” J. Clin. Invest., vol. 68,
pp. 1456–1467, 1981.
[14] R. Bergman, Y. Ider, C. Bowden, and C. Cobelli, “Quantitative estima-
tion of insulin sensitivity,” Am. J. Physiol., vol. 236, pp. E667–E677,
1979.
[15] M. Berger and D. Rodbard, “Computer simulation of plasma insulin
and glucose dynamics after subcutaneous insulin injection,” Diab.
Care, vol. 12, pp. 725–736, Nov./Dec. 1989.
[16] J. Sorensen, “A Physiologic Model of Glucose Metabolism in Man and
its Use to Design and Assess Improved Insulin Therapies for Diabetes,”
Ph.D. dissertation, Massachusetts Inst. Technol., Boston, MA, 1985.
[17] R. Parker and F. Doyle, III, “Control-relevant modeling in drug de-
livery,” Adv. Drug Deliv. Rev., vol. 48, pp. 211–228, 2001.
[18] S. Arimoto, S. Kawamura, and F. Miyazaki, “Bettering operation of dy-
namic systems by learning: a new control theory for servomechanism
or mechatronics systems,” in Proc. IEEE 23rd Conf. Decision and Con-
trol, Las Vegas, NV, Dec. 1984, pp. 1064–1069.
[19] N. Amann, D. Owens, and E. Rogers, “Iterative learning control for dis-
crete-time systems with exponential rate of convergence,” in Proc. Inst.
Elect. Eng. Control Theory Applications., Dearborn, MI, Mar. 1996,
pp. 217–224.
[20] K. Lee, S. Bang, and K. Chang, “Feedback-assisted iterative learning
control based on an inverse process model,” J. Proc. Cont., vol. 4, pp.
77–89, May 1994.
[21] J. Lee, K. Lee, and W. Kim, “Model-based iterative learning control
with a quadratic criterion for time-varying linear systems,” Automatica,
vol. 36, pp. 641–657, May 2000.
IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING, VOL. 53, NO. 6, JUNE 2006
[22] B. Srinivasan, C. Primus, D. Bonvin, and N. Ricker, “Run-to-run op-
timization via control of generalized constraints,” Contr. Eng. Prac.,
vol. 9, pp. 911–919, 2001.
[23] B. Srinivasan, S. Palanki, and D. Bonvin, “Dynamic optimization of
batch process. I. Characterization of the nominal solution,” Comp.
Chem. Eng., vol. 27, pp. 1–26, 2002.
[24] B. Srinivasan, D. Bonvin, E. Visser, and S. Palanki, “Dynamic opti-
mization of batch processes. II. Role of measurements in handling un-
certainty,” Comp. Chem. Eng., vol. 27, pp. 27–44, 2002.
[25] R. Good, J. Hahn, T. Edison, and S. Qin, “Drug dosage adjustment
via run-to-run control,” in Proc. American Control Conference, An-
chorage, AK, May 2002.
[26] H. Zisser, L. Jovanovic, F. Doyle, P. Ospina, and C. Owens,
“Run-to-run control of meal-related insulin dosing,” Diab. Technol.
Ther., vol. 7, pp. 48–57, 2005.
Camelia Owens received the B.S. degree from the
University of Maryland Baltimore County, Balti-
more, MD and the Ph.D. degree from the University
of Delaware, Newark, all in chemical engineering.
She is an ExxonMobil Scholar-in-Residence at the
National Academy of Engineering, Washington, DC.
Howard Zisser received the B.S. degree from the
University of Florida, Gainesville, and the M.D. de-
gree from the Johns Hopkins University, Baltimore,
MD.
He is the Director of Clinical Research at the
Sansum Diabetes Research Institute, Santa Barbara,
CA.
Lois Jovanovic received the M.D. degree from
Albert Einstein College of Medicine, Bronx, NY.
Her postgraduate training included a residency in
internal medicine at The New York Hospital, Cornell
University Medical College and a fellowship in
endocrinology and metabolism at Cornell University
Medical College.
She is Chief Executive Officer and Chief Scientific
Officer at Sansum Diabetes Research Institute, Santa
Barbara, CA. She is a Clinical Professor of Medicine
at the University of Southern California-Keck School
of Medicine, Los Angeles, and on Faculty in Internal Medicine, Santa Barbara
Cottage Hospital. In addition she is an Attending Physician at the Santa Barbara
County Diabetes/Endocrine Clinic, and an Adjunct Professor in the Biomolec-
ular Science and Engineering Program at the University of California, Santa
Barbara.
Dominique Bonvin received the Diploma in chem-
ical engineering from the ETH, Zurich, Switzerland,
and the Ph.D. degree from the University of Cali-
fornia, Santa Barbara.
He is Professor of Automatic Control at the Swiss
Federal Institute of Technology (EPFL), Lausanne,
Switzerland. He worked in the field of process con-
trol for the Sandoz Corporation, Basel, Switzerland,
and with the Systems Engineering Group of the ETH
Zurich. He joined the EPFL in 1989 where his cur-
rent research interests include modeling, identifica-
tion, and optimization of dynamical systems. He is presently Dean of Bachelor
and Master studies at EPFL.
 OWENS et al.: RUN-TO-RUN CONTROL OF BLOOD GLUCOSE CONCENTRATIONS FOR PEOPLE WITH TYPE 1 DIABETES MELLITUS 1005

Francis J. Doyle, III, received the B.S.E. degree
from Princeton University, Princeton, NJ, in 1985,
the C.P.G.S. degree from Cambridge University,
Cambridge, U.K., in 1986, and the Ph.D. degree from
the California Institute of Technology (Caltech),
Pasadena, in 1991, all in chemical engineering.
He holds the Duncan and Suzanne Mellichamp
Chair in Process Control in the Department of
Chemical Engineering at the University of Cali-
fornia, Santa Barbara (UCSB), where he is also the
Associate Director of the Institute for Collaborative
Biotechnology. Prior to coming to UCSB, he has had positions at DuPont,
Purdue University, the University of Delaware, and the University of Stuttgart.
His research interests are in particulate process control, systems biology, and
control of biomedical systems.
Dr. Doyle is the recipient of several research awards [Humboldt Research Fel-
lowship (2001–2002), NSF NYI (1992), and ONR Young Investigator (1996)] as
well as teaching awards [ASEE Indiana Section (1996) and Tau Beta Pi Teaching
(1996)].

All in-text references underlined in blue are linked to publications on ResearchGate, letting you access and read them immediately.