978 3 540 74031 5

Lecture Notes
in Control and Information Sciences 368

Editors: M. Thoma, M. Morari
Frederick Chee, Tyrone Fernando
Closed-Loop Control of Blood

Glucose
ABC
Series Advisory Board
F. Allgöwer, P. Fleming, P. Kokotovic,
A.B. Kurzhanski, H. Kwakernaak,
A. Rantzer, J.N. Tsitsiklis
Authors
Frederick Chee
School of Electrical
Electronic and Computer Engineering
University of Western Australia
35 Stirling Highway
Crawley, WA 6009
Australia
Email: hf.chee@ieee.org
Tyrone Fernando
School of Electrical
Electronic and Computer Engineering
University of Western Australia
35 Stirling Highway
Crawley, WA 6009
Australia
Email: tyrone@ee.uwa.edu.au
Library of Congress Control Number: 2007932674
ISSN print edition: 0170-8643

ISSN electronic edition: 1610-7411
ISBN-10 3-540-74030-9 Springer Berlin Heidelberg New York
ISBN-13 978-3-540-74030-8 Springer Berlin Heidelberg New York
This work is subject to copyright. All rights are reserved, whether the whole or part of the material is
concerned, specifically the rights of translation, reprinting, reuse of illustrations, recitation, broadcasting,
reproduction on microfilm or in any other way, and storage in data banks. Duplication of this publication or
parts thereof is permitted only under the provisions of the German Copyright Law of September 9, 1965, in
its current version, and permission for use must always be obtained from Springer. Violations are liable for
prosecution under the German Copyright Law.
Springer is a part of Springer Science+Business Media
springer.com
c Springer-Verlag Berlin Heidelberg 2007
The use of general descriptive names, registered names, trademarks, etc. in this publication does not imply,
even in the absence of a specific statement, that such names are exempt from the relevant protective laws and
regulations and therefore free for general use.
Typesetting: by the authors and SPS using a Springer LATEX macro package
Printed on acid-free paper SPIN: 12077946 89/SPS 543210
To Jehovah God, my parents and siblings
Frederick
To my wife Yasmin and two daughters Melissa and Rochelle

Tyrone
Preface
Diabetes is a disease that is now regarded an epidemic in the world and a signif-
icant effort is directed towards finding better ways to manage diabetes. Keeping
blood glucose levels as close to normal as possible, leads to a substantial decrease
in long term complications of diabetes and can bring significant cost reductions
associated with the disease. Traditionally, managing diabetes has been through
intermittent monitoring of blood glucose and then administering an appropriate
dose of insulin into the blood stream. This method of intermittent monitoring
and administration of insulin cannot ensure blood glucose remains at near nor-
mal levels at all times and therefore, there is considerable interest in managing
diabetes on a continuous basis.
The development of artificial organs/apparatus that regulate human’s blood
glucose level has been in progress since 1960. The aim was to measure blood
glucose level ex vivo and then injecting an appropriate amount of insulin to the
hyperglycaemic patient, thereby correcting the high glucose level. This aim of
closing the “loop” is still being challenged by technological barriers even today,
and progress are being made constantly both in overcoming the challenges and
understanding more about the workings of glucose-regulatory system.
The purpose of this book is to introduce the field of closed-loop blood glu-
cose control, in a simple manner, to the reader. This includes the hardware
and software components that make up the control system (see Chapter 2).
The hardware components involved the different types of glucose sensor (in-
vasive, minimally-invasive and non-invasive) and the different types of insulin.
The software component represents the approaches that translate a given blood
glucose reading to an appropriate insulin rate (see Chapter 4). Some of these
approaches applied mathematical sciences to model the underlying system and
formulate mathematical solutions. Examples of how mathematical models are
formulated as well as the control algorithms that stem from mathematical exer-
cises are given, where possible.
This book also attempts to describe, from functional level, the basic physi-
ology of blood glucose regulation during fasting and meal (see Chapter 3). The
physiological changes in diabetic and critically ill patients are also described.
VIII Preface
This would help in understanding the nature of the glucose-regulatory prob-

lem that we are facing and tackling. Finally, an example design of a closed-loop
hardware using commercially-available equipment is given (see Chapter 5).
The ultimate goal in closed-loop control of blood glucose is not just finding
the optimal insulin rates that can effectively reduce the high blood glucose, but
to infuse it in such a way that the blood glucose level can mimic the body’s
natural excursion.
Perth Frederick Chee

June 2007 Tyrone Fernando
Contents
1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 Causes of Hyperglycaemia . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2.1 Physiological Background . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2.2 Historical Background . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.2.3 Diabetic Patients . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.2.4 Surgical Patients . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.3 Importance of Tighter BG Level Control . . . . . . . . . . . . . . . . . . . . . 3
1.4 Achieving Tighter Control . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
2 Glucose Control: Input and Output . . . . . . . . . . . . . . . . . . . . . . . . . . 5

2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
2.2 BG Measurement – Invasive Techniques . . . . . . . . . . . . . . . . . . . . . . 6
2.3 BG Measurement – Minimally-Invasive Techniques . . . . . . . . . . . . . 7
2.3.1 Implantation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
2.3.2 Fluid Extraction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
2.4 BG Measurement – Non-invasive Techniques . . . . . . . . . . . . . . . . . . 15
2.4.1 Infrared Absorptiometry . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
2.4.2 Mid-IR Emission Spectrometry . . . . . . . . . . . . . . . . . . . . . . . . 21
2.4.3 Scatter Change . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
2.4.4 Tissue Manipulation and Red/Near-IR Absorptiometry
(λ ∼ 630–1300 nm) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
2.4.5 Raman Spectroscopy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
2.4.6 Polarimetry . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
2.4.7 Photoacoustic . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
2.4.8 Light Diffraction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
2.4.9 Dielectric Spectroscopy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
2.5 Amperometric Sensor Calibration . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
2.5.1 General Calibration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
2.5.2 Regression Calibration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
X Contents
2.5.3 GlucoWatch Calibration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38

2.5.4 Differences Between the MiniMed’s CGMStm and
Cygnus GlucoWatch Biographer . . . . . . . . . . . . . . . . . . . . . 41
2.6 Clarke’s Error Grid Analysis (EGA) . . . . . . . . . . . . . . . . . . . . . . . . . 41
2.7 Insulin Infusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
2.7.1 Fate of Insulin . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
2.7.2 Insulin Type Factor . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
2.7.3 Routes of Insulin Administration . . . . . . . . . . . . . . . . . . . . . . 43
2.8 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
2.9 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
3 Glucose Control: Patient Dynamics . . . . . . . . . . . . . . . . . . . . . . . . . . 49

3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
3.2 Intrinsic Blood Glucose Regulation . . . . . . . . . . . . . . . . . . . . . . . . . . 49
3.3 Diabetic Patients . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
3.4 Importance of BG Control . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52
3.4.1 New Findings . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
3.5 Glucose Control in Critically Ill Patients . . . . . . . . . . . . . . . . . . . . . 53
3.5.1 Loss of the Inhibitory Effects of Elevated Glucose
Levels . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54
3.5.2 Stress-Induced Release of Counter-Regulatory
Hormones . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54
3.5.3 Increased Insulin Resistance . . . . . . . . . . . . . . . . . . . . . . . . . . 55
3.5.4 Medication that Induces Hyperglycaemia . . . . . . . . . . . . . . . 55
3.5.5 Effects of Feeding . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
3.5.6 GI Tract Functionality . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56
3.5.7 Stress in Liver . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56
3.6 BG Management in the ICU . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56
3.6.1 BG Measurement . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57
3.6.2 Insulin Infusion Adjustment . . . . . . . . . . . . . . . . . . . . . . . . . . 57
3.7 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57
3.8 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57
4 Mathematics of Glucose Control . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59

4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59
4.2 Model-Less (Empirical) Control Algorithms . . . . . . . . . . . . . . . . . . . 60
4.2.1 Control Algorithm Based on Curve-Fitting . . . . . . . . . . . . . 60
4.2.2 Control Algorithm Based on Lookup Table . . . . . . . . . . . . . 61
4.2.3 Control Algorithm Based on Rule-Based Control . . . . . . . . 62
4.2.4 Control Algorithm Based on PID Control . . . . . . . . . . . . . . 64
4.3 Model-Based Control Algorithms . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67
4.3.1 Theoretical . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
4.3.2 Empirical + Theoretical . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
4.4 Mathematical Models of Gluco-regulatory System . . . . . . . . . . . . . 69
4.4.1 Linear Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
4.4.2 Ackerman’s Linear Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71
Contents XI
4.4.3
Non-linear and Comprehensive Models . . . . . . . . . . . . . . . . . 72
4.4.4
Minimal Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73
4.4.5
Cobelli’s Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75
4.4.6
Candas and Radziuk . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78
4.4.7
Hovorka’s Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79
4.4.8
Sorensen’s Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82
4.4.9
Sub-models That Complements the Comprehensive
Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86
4.5 Mathematics of the Model-Based Control Algorithms . . . . . . . . . . 87
4.5.1 Pole Placement . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87
4.5.2 Optimal Control . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 88
4.5.3 Adaptive Control . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 91
4.5.4 Model Predictive Control . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 94
4.5.5 H∞ Control . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97
4.5.6 Note . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 106
4.6 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 106
4.7 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107
5 Closed-Loop Control Apparatus Example . . . . . . . . . . . . . . . . . . . . 109

5.1 CGMS Integration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 109
5.1.1 Hardware Integration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 109
5.1.2 Software Integration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 110
5.1.3 Clinical Trial Example . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 113
5.2 Miniaturisation – Glucose Sensing Circuit . . . . . . . . . . . . . . . . . . . . 117
5.2.1 Background . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 117
5.2.2 Constant Potential Application Circuit . . . . . . . . . . . . . . . . . 118
5.2.3 Nano-ampere Measuring Circuit . . . . . . . . . . . . . . . . . . . . . . . 119
5.2.4 Supply Potential Splitting Circuit . . . . . . . . . . . . . . . . . . . . . 119
5.3 Miniaturisation – Syringe Pump and Microprocessor . . . . . . . . . . . 120
5.3.1 Motor Driving Circuit . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 120
5.3.2 Motor Rotation Detector Circuit . . . . . . . . . . . . . . . . . . . . . . 121
5.3.3 LCD and Hardware Buttons . . . . . . . . . . . . . . . . . . . . . . . . . . 121
5.3.4 Microprocessor . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 121
5.4 Signal Processing and Control Algorithm Programming . . . . . . . . 121
5.5 Circuit Performance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 122
5.6 Basic Safety Issues . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 124
5.7 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 126
5.8 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 126
6 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 127
A Mathematical Derivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 129

A.1 Solving Recursive Least Square Estimation . . . . . . . . . . . . . . . . . . . 129
A.2 Linearisation of Non-linear Model . . . . . . . . . . . . . . . . . . . . . . . . . . . 130
XII Contents
B Model Parameters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 133

B.1 Ackerman’s Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 133
B.2 Minimal Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 134
B.3 Cobelli’s Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 134
B.4 Hovorka’s Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 135
B.5 Sorensen’s Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 136
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 139
Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 155
1
Introduction
1.1 Introduction
Hyperglycaemia refers to an elevated glucose concentration in the circulating

blood. While blood glucose level (BG) is often elevated after a meal, it usually
normalises to a range of 3.5–5.6 mmol/l within 3 hours of a meal in a healthy
individual. During fasting, BGs are also usually maintained within the normal
range of 3.5–5.6 mmol/l [1].
Patients with diabetes mellitus, however, exhibit impaired regulatory re-
sponses. BG remains elevated postprandially, and during the fasting state. Pa-
tients under surgical or medical stress may also exhibit a diabetogenic response,
even though they are not diabetics [2]. This is due to an elevated catecholamine
and other “counter-regulatory” hormones in their circulation.
1.2 Causes of Hyperglycaemia
1.2.1 Physiological Background
Blood glucose level is usually regulated by two hormones – insulin and glucagon –
secreted by the endocrine pancreas. Insulin is anabolic, and causes rapid uptake
and use of glucose by most tissues in the body. It also causes the storage of excess
glucose as glycogen, mainly in the liver and skeletal muscles. Excess glucose, that
cannot be stored as glycogen, is converted to fatty acids and stored in adipose
tissues. Insulin also promotes protein synthesis and storage.
Conversely, glucagon is catabolic, mobilizing glucose, fatty acids and amino
acids from stores to the circulation, primarily through a breakdown of liver
glycogen (glycogenolysis) and the generation of glucose from amino acids (glu-
coneogenesis).
The two hormones (insulin and glucagon) are reciprocal in their overall action
and are secreted appropriately in most circumstances to keep the blood glucose
concentration within the normal range [3]. When the glucose concentration rises
too high, insulin is secreted, which then lowers the blood glucose concentration
F. Chee & T. Fernando: Closed-Loop Control of Blood Glucose, LNCIS 368, pp. 1–4 , 2007.
springerlink.com
2 1 Introduction
toward normal. Conversely, a decrease in blood glucose stimulates glucagon se-

cretion; glucagon then functions in the opposite manner to increase the glucose
concentration towards normal [4].
1.2.2 Historical Background
Before insulin was discovered, patients with diabetes suffered from polyuria and
a catabolic state, which depleted them of strength, weight and fluid. As described
by Arestaeus of Cappadocia in the 2nd century,
“Diabetes is a dreadful affliction, not very frequent among men, being a
melting down of the flesh and limbs into urine. The patients never stop
making water and the flow is incessant, like the opening of aqueducts.
Life is short, unpleasant and painful, thirst unquenchable, drinking ex-
cessive, and disproportionate to the large quantity of urine, for yet more
urine is passed. One cannot stop them either from drinking or making
water. If for a while they abstain from drinking, their mouth becomes
parched and their body dry; the viscera seem scorched up, the patients
are affected by nausea, restlessness and a burning thirst and within a
short time, they expire.” (Adapted from [5]).
For centuries, the only available treatment for patient with diabetes was starva-
tion. If the diabetes worsened, then more starvation was prescribed.
The discovery of insulin at the University of Toronto in 1921–1922 was one
of the most important milestones in the history of medicine [6]. The team that
discovered insulin was quickly awarded a Nobel Prize in 1923.
1.2.3 Diabetic Patients
In patients with diabetes mellitus, the correlation between insulin delivery and
blood glucose concentration is impaired. Both Type I and Type II diabetics have
a dysfunctional endocrine pancreas, which produces little (as in Type II) or no
insulin (as in Type I). Apart from this, the insulin receptor on the tissue cells
in Type II diabetics also respond abnormally to the circulating insulin (“insulin
resistance”). Type I diabetes is also known as Insulin-Dependent Diabetes Mel-
litus (IDDM), while Type II is also known as Non-Insulin-Dependent Diabetes
Mellitus (NIDDM).
1.2.4 Surgical Patients
Surgical patients commonly enter a hypermetabolic stress state, induced by the

area of infection or injury, and promoted by organs involved in the immunologic
response to stress [7]. In this stressed state, insulin secretion is suppressed, and
the normal carbohydrate metabolism is altered. This results in increased glucose
production, depressed glycogenesis (i.e. reduced conversion of glucose into the
storable form of glycogen), glucose intolerance, and insulin resistance [7].
1.3 Importance of Tighter BG Level Control 3
1.3 Importance of Tighter BG Level Control
BG level should be kept within the normal range, because:

1. a high glucose concentration exerts an osmotic pressure in the extracellular
fluid, and can cause cellular dehydration.
2. too low a BG level carries the risk of hypoglycaemic coma. Glucose is the
only source of energy that can be used by the brain. Prolonged and profound
hypoglycaemia can produce severe brain damage.
3. too high a glucose concentration (>11.1 mmol/l) can affect wound healing
and interfere with human neutrophil function [8].
4. therapy that maintains BG level at below 11.9 mmol/l has been shown to
improve the long-term outcome of diabetic patients with acute myocardial
infarction [9].
Researchers in the last few decades have found that the mere use of insulin alone
is not enough to guarantee the well-being of the patient (see e.g. [5, 10]), as di-
abetic microvascular complications have emerged. These microvascular compli-
cations include retinopathy (visual impairment), nephropathy (kidney disease)
and neuropathy (nerve damage).
Studies in Type I [11] and Type II [12,13] diabetic patients have found that the
onset and progression of these serious complications can be delayed by intensive
insulin therapy. This intensive therapy includes the administration of insulin
three or more times daily by injection or an external insulin pump, with the goal
of achieving preprandial blood glucose concentrations of between 3.9 mmol/l
and 6.7 mmol/l, postprandial concentration of less than 10 mmol/l, a weekly
3 a.m measurement greater than 3.6 mmol/l, and haemoglobin A1c (glycosylated
haemoglobin) within the normal range (less than 6.05%) measured monthly.
The insulin dosage was adjusted according to the results of self-monitoring of
blood glucose (SMBG) performed at least four times a day, dietary intake, and
anticipated exercise.
The complications listed above occurred notably in diabetic patients, but also
impacted on the critically ill patient population:
1. Findings from [9] has shown that maintaining BG at a level that did not ex-
ceed 110 mg/dl (6.1 mmol/l) substantially reduced mortality and morbidity
in critically-ill patients in the ICU. In addition, a pronounced hyperglycaemia
in critically-ill patients, even those who have not previously had diabetes,
may lead to complications in such patients [9].
2. Patients with mean glucose concentrations >11.1 mmol/l within 36 h follow-
ing surgery were more likely to develop infectious complications than their
counterparts who were under better glycaemic control [14].
All these findings demonstrated that apart from the administration of exogenous
insulin being essential to control the blood glucose concentration and to main-
tain homeostasis in patients, the tightness of such control is crucial in avoiding
4 1 Introduction
long-term complications (in diabetic patients) and infectious complications (in

critically ill patient). The finding in [9] sets a desirable BG target for a closed-
loop algorithm to achieve.
1.4 Achieving Tighter Control

In an effort to achieve tighter control, attempts have been made to sample the
patients’ BG level regularly, and adjust the insulin infusion rate automatically,
so as to steer the BG towards a given target value (see e.g. [15], [16] etc). In
such early experiments, whole blood was sampled continuously for glucose mea-
surement, using the invasive method (see Section 2.2). Good glycaemic control
was reported using this approach.
On the other hand, clinical routines have progressed with manual BG level
control, which uses intermittent (and not continuous) BG level sampling. BG
samples are taken from finger pricks or in-dwelling cannula, hourly, 2 hourly or
4 hourly, depending on severity. Various algorithms for perioperative control of
BG level have been proposed (see Section 4.2), using hourly and three-hourly
whole blood BG level sampling. Closed-loop (automatic) systems were not prac-
tical for use in routine treatment, due to the cost and preparation associated
with their operation. Furthermore, long term use of the system posed safety
concerns, due to the invasive nature of the measuring technique.
1.5 Conclusion
It is important to tightly control blood glucose level, both in ambulatory and
hospitalised patients, due to the demonstrated benefits. However, achieving tight
control requires frequent sampling of BG value, which to-date has been painful,
difficult (and expensive). There are efforts in-progress to reduce or alleviate the
pain of BG sampling, and also to improve the regulation of BG. These progress
will be discussed in the chapters to follow.
2
Glucose Control: Input and Output
2.1 Introduction
Automatic regulation of a patient’s blood glucose (BG) level requires a minimum
of three components, namely, a continuous BG sensor, a controller that matches
BG level with an appropriate insulin delivery rate, and an infusion pump to
deliver the insulin to the subject.
Shown here is a simple model of the control loop:
Sensor Controller Pump
Patient
As the variable to be controlled is a patient’s BG level, a knowledge of the

BG is required. This is provided by a glucose sensor, and represents the input
to the control system. Since insulin is used to lower a high BG level, the rate of
insulin delivery represents the output of the control system. The patient is the
“plant” to be controlled in control system terminology.
The controller is the component of the system that regulates the blood glu-
cose levels in the patient. The formulation of the control rule depends on the
knowledge we have about the sensor, the pump and the patient - and specifi-
cally, the BG measurement methods, the type (or preparation) of insulin used,
the route of infusion, and the patient’s characteristics. Various BG measure-
ment techniques exist, and each has its unique characteristics. As for insulin,
F. Chee & T. Fernando: Closed-Loop Control of Blood Glucose, LNCIS 368, pp. 5–4 8, 2007.
springerlink.com
6 2 Glucose Control: Input and Output
each type of insulin has different kinetics, and different infusion routes exhibit
different characteristics.
This chapter reviews:
• the various BG measurement techniques in existence to-date, which may gen-
erally be categorised into:
1. Invasive,
2. minimally invasive, and
3. non-invasive techniques.
• the properties of the different types of insulin and their routes of infusion.
2.2 BG Measurement – Invasive Techniques

Historically, measurement of BG has been done via direct venous access, and
subsequent ex-vivo determination of whole blood glucose levels by a glucose sen-
sor (as used in [16,17]). As this was an invasive method, patients needed constant
medical supervision [18, 19], while a portion of their blood was channelled from
a vein into a dual-lumen catheter, mixed with heparin (to prevent blood from
clotting in the tubing), and sent to a measuring device [16,17]. The heparin does
not enter the patient’s circulation.
In the measuring device, the BG level was usually measured based on an
immobilised glucose oxidase combined with amperometric detection of either
peroxide, oxygen, changes in the pH or changes in colour [16,17]. Other detection
methods include potentiometric detection, calorimetric detection, or a number
of other variations of indicator techniques. More details of these techniques can
be found in [20] and [21].
The latency of the BG measurement process (i.e. time taken from venous
blood extraction to ultimate measurement) was initially 10 min, but was re-
duced to 1.5 min with an improved measurement technique. The promptness of
measurement was important in the 1980s, because the so-called “artificial en-
docrine pancreas” regulated the insulin delivery on a minute-by-minute basis.
It was reported that any delay of more than 6 min in the measurement would
prevent effective feedback action [22]. Nowadays, on-line glucose analysers have
a measurement latency of 50 sec (such as the Via Blood Glucose Monitor from
Metracor, San Diego, CA).
Advantages
• The advantage of this method is the accuracy of the measurement, as whole
blood is analysed for its glucose content, i.e. the measurement is made in the
actual fluid of interest.
Disadvantages
• Since invasive BG measurements continuously withdraw blood for analysis,
any studies conducted using the technique are limited in duration as described
2.3 BG Measurement – Minimally-Invasive Techniques 7
below. The withdrawn blood is usually discarded (as it is mixed with heparin,
and is not safe to be returned to the patient).
• The practicability of long-term invasive measurement for closed-loop glucose
control is mainly hindered by concerns for patient safety, because this tech-
nique carries the risk of infection and thrombosis. Furthermore, the fact that
the technique cannot be performed by patients themselves without proper
supervision [19, 23] limits the technique to be used only in hospitals.
2.3 BG Measurement – Minimally-Invasive Techniques

Minimally invasive methods provide a more favourable solution than invasive
methods, mainly because the measurement of glucose is now done outside the
vascular tree, minimizing the risk associated with vessel access (i.e. pain, infec-
tion and thrombosis).
These methods measure the blood glucose level indirectly — using the glucose
level in the interstitial fluid that surrounds all cells under the skin (or subcu-
taneous space) to reflect the plasma glucose level. This technique is based on
the assumption that a correlation exists between the subcutaneous (s.c.) glucose
level and the plasma (or whole blood) glucose level.
This assumption is valid to an extent, as many studies have reported on the
relationship of subcutaneous glucose to blood glucose [24, 25]:
• At equilibrium, the glucose concentration at s.c. sites correlated well with
those of plasma [25, 26].
• During periods when the blood glucose concentration is rising or falling, the
two concentrations differ [27] with a delay of an average of 10 min in the s.c.
tissue [28]. But delays as much as 45 min have also been reported [25].
Sites such as adipose tissue and s.c. tissues are regarded as most appropriate
for minimally-invasive BG determination because of the ease of accessibility
for surgery or sensor replacement [29]. Also, less pain receptors exist at these
sites [24].
Measurement of s.c. glucose can be done by direct implantation of a glucose
sensor in the s.c. tissue, or by means of fluid extraction from the s.c. space
(usually by microdialysis), as discussed below.
2.3.1 Implantation
In “implantation”, miniaturised glucose sensors were inserted directly into the

s.c. tissues [30, 31], and measurement was performed in situ. Once the sensor is
inserted, glucose level can be determined using methods like:
• amperometry
• fluorescence detection
Amperometric Glucose Sensor
Among the chemically-based methods of glucose sensing, the amperometric sen-

sor is one of the most simple, stable, and widely studied (and perhaps the most
successful) approaches [21]. As described in [32],
“amperometry is the determination of the intensity of the current cross-
ing an electrochemical cell under an imposed potential. This intensity is
a function of the concentration of the electrochemically active species in
the sample.”
Amperometric glucose sensors are almost entirely based on the glucose-oxidase
(GOx) enzyme, with the flavin-adenin-dinucleotide (FAD) radical as the active
part [33]. FAD is able to react with hydrogen to form FADH2 . The two forms of
GOx are thus marked as “GO(FAD)” and “GO(FADH2 )”. The chemical prin-
ciples underpinning amperometric methods of measuring glucose concentration
are described in Table 2.1.
Table 2.1. Glucose-oxidase reaction
Glucose-Oxidase Reaction
Glucose oxidase dehydrogenates glucose (C6 H12 O6 ), transforming it to glu-

conolactone (C6 H10 O6 ):
Glucose + GO(F AD) −→ gluconolactone + GO(F ADH2 )
And:
1. Gluconolactone reacts with water to form gluconic acid (C6 H12 O7 ),

which dissociates in solution.
Gluconolactone + H2 O −→ C6 H12 O7 ⇐⇒ C6 H11 O7− + H +
2. Glucose oxidase is then oxidised in the presence of oxygen, and hydrogen

peroxide, H2 O2 , is formed as a result.
GO(F ADH2 ) + O2 −→ GO(F AD) + H2 O2
By applying -600 mV to -700 mV (with respect to the working platinum

electrode), H2 O2 will be electrolysed into O2 and 2H + and electrons flow as
a result.
H2 O2 −→ O2 + 2H + + 2e−
This generated current is linearly dependent upon the concentration of glu-
cose.
Since oxygen is a limiting factor in the reaction, other methods that use
mediator to eliminate the dependence on oxygen also exist. The reaction
described above is only representative.
When amperometric sensors are used as implantable sensors, enzyme immo-

bilisation techniques such as cross-linking of glucose oxidase and Bovine Serum
Albumin (BSA) to the electrodes using glutaraldehyde (GA), are frequently used
to ensure maximal contact and response. Immobilisation has the advantage of
stabilising the protein so that it can be used repeatedly [32]. The use of BSA
improves enzymatic activity because of the better mass distribution of the var-
ious proteins. This configuration (i.e. GOx + BSA + GA) has a response time
of 10 sec [32], and this is the reason behind the CGMS sensor taking glucose
readings every 10 sec (because it uses the same configuration (see [34])).
Fig. 2.1. Basic circuit illustrating the operation of amperometric sensors (Adapted
from [33])
Implantable amperometric sensors have only recently become available as an

off-the-shelf product. Medtronic MiniMed Inc (Northridge, CA) is the first com-
pany to market a FDA-approved amperometric glucose sensor for use in human
subjects. Called MiniMed CGMS (Continuous Glucose Monitoring System), it
is designed for continuous BG monitoring in diabetic patient for periods of up to
3 days. CGMS takes a subcutaneous glucose reading every 10 sec and stores an
averaged reading every 5 min. The sensor is pre-sterilised, readily-implantable,
easily inserted, and features a short (1-hour) preparation time. The short prepa-
ration time is an advantage over the 2-hour run-in period required for other
earlier attempts [35].
Fluorescence Detection
Fluorescence-based glucose sensor involves measuring a change in light emission

when glucose binds to certain molecules (in affinity-binding process) or reacts
with certain molecules (in glucose-oxidase reaction).
• Affinity-binding glucose sensors uses competitive binding between glucose and
a suitably labelled fluorescent compound, to a common (and glucose-specific)
receptor site [38]. These fluorescence chemicals are immobilised within a se-
lectively permeable membrane at the tip of an optical fibre sensor probe.
Fig. 2.2. [Left] Con A is immobilised on the inside of a hollow dialysis fibre. A low
molecular weight cutoff of the dialysis fibre retains the dextran within the fiber lumen
while allowing glucose to pass freely through the dialysis membrane [36]. Excitation
light passes through the fibre optics and into the solution, fluorescing the unbound
FITC-dextran. The resulting small fluorescence intensities are measured through the
same fibre optics by a measuring system [37]. [Right] FRET mechanism allowing an
acceptor-labelled dextran to quench the fluorescence when a donor is nearby (see text).
1. Schultz et al [36] used a protein called Concanavalin A (or Con A)1 as the
common receptor, with a fluorescein-isothiocynate labelled dextran2 as the
competing ligand. An increased presence of glucose will displace with the
dextran from Con A, leading to an increase in fluorescence intensity (Fig
2.2).
2. In more recent work (e.g. McCartney et al [39]), the phenomenon of Flu-
orescence Resonance Energy Transfer (FRET) was used (Fig 2.2). In this
method, Con A is labelled with a fluorophore donor, and the dextran is
labelled with an acceptor. When the dextran binds to Con A, the donor is
closer to the acceptor. As the donor is excited at its specific fluorescence
excitation wavelength, this excited state is nonradiatively transfered to the
acceptor (FRET phenomenon), resulting in a diminished fluorescence. Con-
versely, the displacement of dextran by glucose reduces FRET and increases
fluorescence intensity and lifetime [41].
1
Con A is a plant lecitin [39], extracted from jack bean [40].
2
Dextran is a carbohydrate derivative [40], and is made of many glucose molecules
joined into chains of varying length.
3. Ballerstadt et al [42] has proposed subcutaneous implantation of a self-

contained semipermeable membrane chamber containing the glucose-
sensitive fluorescent reagents close to the skin surface. Glucose measurement
can be performed by illuminating the implant region. Similar idea has been
trialled and presented by PreciSense A/S (Horsholm, Denmark) as part of
the company’s goal to develop a stable and precise minimally-invasive com-
mercial FRET-based glucose sensor with 2 weeks continuous use (See [43]).
4. March et al [44] describe the incorporation of fluorescent-based glucose sen-
sor into intra-ocular lens. The lens contains tetramethylrhodamine isoth-
iocyanate Concanavalin A (TRITC-Con A) and FITC-dextran. When
TRITC-Con A binds competitively with FITC-dextran and glucose, the
bound FITC-dextran fluoresence is quenched through FRET. An increase
in glucose concentration will liberate FITC-dextran, producing a fluores-
cence proportional to the glucose concentration.
• Glucose oxidase catalytic reaction can be monitored by detecting oxygen con-

sumption or hydrogen peroxide production via fluorescent (see the reaction
equations in Table 2.1). Glucose concentration can be quantified:
1. using fluorescence that can be quenched by oxygen [40], or fluorophore that
is sensitive to oxygen concentration [38, 45];
2. using redox mediator dye [46], where the dye (instead of water) reacts with
gluconolactone to produce current;
3. detecting the formation of fluorescent oxidation product by hydrogen perox-
ide [38, 47].
There are other fluorescence and affinity-binding methods in existence. Interested

readers in the subject are encouraged to read [40].
Advantages of Implantable Sensors
• Implantable sensors are small in size, portable and do not require extraction
of fluids for operation.
Disadvantages of Implantable Sensors
• The most common problem inhibiting the longevity and reliability of im-
plantable sensors is membrane biofouling. Biofouling is the adhesion of pro-
teins and other biological matter on the sensor surface, and in some cases,
impregnation of the material [48], resulting in drifting and/or diminishing
of the sensor signal. Biofouling starts immediately upon contact of the sen-
sor with the body. Biofouling remains a problem in both amperometric and
fibre-optic based glucose sensor [49].
• Biological matter including the immune cells can attack the glucose sensor
assembly (especially the membrane) [50, 51], causing inflammation.
• In the case of amperometric sensor, the body’s warm and saline environment
can corrode the implanted metal electrodes and inactivate the enzymes.
• A person’s movement can create artifacts and noise in the measurement.
• Substances such as vitamin C and acetaminophen may react at the electrode,
depleting the hydrogen peroxide before it can react at the electrodes [51].
Other problems include the oxidation of other reactants at the same po-
tential as glucose oxidase, creating noise and inaccuracy in the measure-
ment.
• Although Con A and glucose react very rapidly with time constants in the
order of milliseconds [42], the response time of the affinity sensor was reported
to be 7 ± 2.5 min [52]. The response time are limited by the diffusion rate
of glucose into or out of the sensor, the diffusion of FITC-dextran within the
hollow fibre lumen, and the temperature.
2.3.2 Fluid Extraction
Some existing fluid extraction techniques include:

• microdialysis;
• transcutaneous harvesting by reverse iontophoresis;
• transdermal extraction by ultrasound;
• microporation with a laser beam (and subsequent fluid extraction).
Microdialysis
Microdialysis is a method that allows blood or interstitial fluid (ISF) to be

sampled and analysed by an ex-vivo sensor. Hollow fibres of dialysis membrane
are implanted, often subcutaneously, and perfused at low flow rate with isotonic
fluid (Fig 2.3). The dialysate is pumped to a flow chamber that incorporates a
glucose sensor [53].
Fig. 2.3. The inner cannula of the dialysis probe is perfused with an isotonic fluid
(perfusate), which is transported to the tip of the dialysis probe. The probe is placed
under the skin, and substances (e.g. glucose) having a concentration gradient compared
with the perfusate, diffuse through the dialysis membrane down the concentration
gradient (Adapted from Meyerhoff et al [26]).
Advantages with Microdialysis
• Microdialysis methods prevent many of the other substances present in the

ISF from entering the measuring cell and creating noise in the measurement.
Disadvantages with Microdialysis
• Although microdialysis filters off much of the unwanted substances preventing

them from entering the measurement chamber, the catheters used in micro-
dialysis probes are very fragile and difficult to use in the clinical environment.
In addition, biofouling will still happen at the “exchange” membrane of the
dialysis probe (rather than at the glucose sensor in the case of implanted
sensor).
• The use of microdialysis systems requires a method to dispose off the dialysate,
which may pose a health risk if not done properly (e.g. due to the potential
for viral transmission such as hepatitis B and C and HIV).
Reverse Iontophoresis
Reverse iontophoresis involves the application (or conduction) of a constant,

low-level electrical current to the skin between an anode and a cathode. The
applied potential draws sodium and chloride ions from beneath the skin toward
the cathode and anode, respectively. Glucose is also carried along with these ions
by electro-osmosis, and transported across the skin. Since the skin has a nega-
tive charge at neutral pH, there is a greater net transport to the cathode. As a
result, glucose is preferentially extracted at the cathode [54–56]. This transder-
mally extracted glucose is then measured by a standard amperometric technique
involving glucose oxidase.
Although the glucose concentration of the extracted fluid is minute (about
∼1/1000 that of blood glucose) [54], it is shown to be proportional to that of
plasma blood [55–57]. This technique is currently being marketed by Cygnus Inc
(Redwood City, CA) as GlucoWatch G2 Biographer, offering near non-invasive
measurement. See Section 2.5.3 for a technical description of GlucoWatch Biog-
rapher.
Advantages with Reverse Iontophoresis
• This technique is near non-invasive. Despite reports of mild discomfort when

using the device [51, 55], this technique is relatively painless compared to
implantable sensors.
Disadvantages with Reverse Iontophoresis
• Because of the use of a low current to extract glucose, it takes 10–20 min
from the beginning of each fluid extraction cycle until a blood glucose level
can be reported [56]. Since extracted glucose comes from the subcutaneous
space, there is a chance that any existing delay in glucose diffusion between
the subcutaneous space and the plasma will be added to the fluid extraction
cycle delay. There is potential for a long lag phase in the detection of blood
glucose changes.
• Excessive sweating can prevent accurate measurement, as the device depends on
minute glucose concentrations in ISF, which may be contaminated by sweat [56].
Ultrasonic Transdermal Extraction
Ultrasonic transdermal extraction uses ultrasound to facilitate the transport of

glucose from the ISF across the skin for non-invasive monitoring. It was reported
that sufficient amount of clinically relevant analytes including glucose can be ex-
tracted using low-frequency ultrasound. Ultrasound is defined as sound that has
frequency beyond 20kHz. The mechanism of improved transdermal transport by
ultrasound is still not well understood [58]. It is suggested that ultrasound can
induce cavitation in and around the skin. Oscillation and collapse of cavitation
bubbles disorder the lipid bilayers of the skin, resulting in enhanced skin per-
meability [59]. It is also reported that once treated with ultrasound, the skin
permeability remains high for up to 15 hours [58].
Suggested configuration of the glucose monitor would feature an ultrasonic
device to permeabilize the skin (for ∼2 min) and a patch to be placed on the
ultrasonically permeabilized skin for continuous extraction and detection of glu-
cose [60]. Sontra Medical Corporation (Franklin, MA, USA) is in the process of
developing an ultrasonic glucose monitor called “SymphonyTM Diabetes Man-
agement System” (see http://www.sontra.com).
Advantages with Ultrasound
• This technique is near-invasive, with no pain and no visible effect of ultrasound

on skin.
Disadvantages with Ultrasound
• A contact medium (e.g. oil, water oil emulsions, ointments) between the ultra-
sonic probe and skin is required. Otherwise, the ultrasound will be completely
reflected by air [60].
• A lag time exists between capillary blood glucose concentration and the ex-
tracted glucose concentration (due to diffusional delay) [60].
• Although no adverse effects of ultrasound was reported, further research fo-
cusing on safety issues is required [58].
Microporation
Microporation involves puncturing the skin layer using a micro laser beam. The
laser creates a micro-hole on the surface of the skin, and glucose is extracted by
2.4 BG Measurement – Non-invasive Techniques 15
means of suction (vaccum) or a patch [61]. The extracted glucose is then mea-
sured by a conventional method (such as amperometry). Products incorporating
this method have been described by companies, such as SpectRx Inc (Norcross,
GA) (http://www.spectrx.com) in [62], and PowderChek Diagnostics (Fremont,
CA).
2.4 BG Measurement – Non-invasive Techniques

Without argument, the most welcomed BG measurement would be a painless
and “finger-prick free” measurement. Such a measurement would require a means
of examining the glucose content in the tissue without puncturing the skin, or
causing discomfort (such as skin irritation etc). The non-invasive techniques
include:
1. Optical spectroscopy;
Glucose can be measured by optical means because it exhibits optical prop-
erties such as light absorption, reflection, polarisation, circular dichroism
and other responses to radiation. Methods that allow the optical properties
of glucose to be identified can be used to determine glucose concentration
non-invasively. Most measurement techniques used in analytical chemistry
have been used to measure glucose.
2. Dielectric spectroscopy (or impedance spectroscopy).
At the time of writing, none of the reported non-invasive technology has been
used in routine clinical practice [58], and progress are being made to bring these
non-invasive system to commercialisation.
2.4.1 Infrared Absorptiometry
One of the most studied methods for glucose determination by its optical prop-
erties is the infrared absorptiometry. Infrared absorptiometry measures the ab-
sorption of light intensity by a solute (e.g. glucose) as light passes through a
solution. By determining the light attenuation caused by absorption at a single
wavelength [63], the glucose concentration in a solution can be quantified. How-
ever, to obtain an accurate reading, the glucose solution has to be clear because
light scattering would result in additional attenuation of light.
In a complex mixture of substances (such as tissue), it is still possible to
quantify glucose by using various wavelengths and complex mathematical pro-
cedure like multivariate calibration. The more the spectra of the substances are
different from each other, the better the reliability of the quantification [63].
The ex vivo measurement of glucose in plasma, serum or whole blood (i.e. com-
plex structures) is feasible using high performance equipment and sophisticated
mathematical calibration procedures [63, 64].
Optical absorption technique are based on selective absorption of light by the

molecule. The absorption can be described by the Beer-Lambert law [38]:
I = Io e−CL
where I0 is the intensity of incident optical radiation, I is the transmitted inten-
sity, is the wavelength-dependent molar extinction coefficient3 [(mol/l)−1 cm−1 ],
C is the molar concentration, and L is the path length [cm].
The absorbance can be defined by
I0
A = log10
I
Infrared absorptiometry works because each molecule has specific resonance ab-
sorption peaks [66], i.e. the specific frequencies at which they vibrate correspond-
ing to the excitation source. In another word, the frequency of the absorbed light
is the molecular vibrational4 frequency actually responsible for the absorption
process [67]. Each organic molecule has an unique chemical composition and thus
an unique infrared spectrum. Optical measurement of glucose can be performed
using Near-infrared or Mid-infrared radiation.
Near-IR [700 nm–2.5 μm or 14000–4000 cm−1 ]

The near-infrared (Near-IR) spectrum extends from the end of the visible spec-
trum5 to the start of the fundamental infrared absorption bands at 2500 nm.
Absorption in the Near-IR region are most often associated with overtone and
combination bands of the fundamental molecular vibrations of C–H, O–H and
N–H functional groups [38, 64, 65].
Near-IR light can penetrate deeper into the tissue (than other infrared fre-
quency bands). In addition, within the spectral range of 600–1300 nm [41], there
is an “optical window” where tissues are “transparent” to light.
Glucose produces one of the weakest Near-IR absorption signals per concentra-
tion unit. In addition, the normal proportion of glucose in blood and tissue is only
about 0.1% of the water content [68]. Measurements of glucose in tissues by optical
means are thus limited by the signal-to-noise ratio that can be achieved at such low
concentrations. Furthermore, measurements can be confounded by issues such as
• interference by water, which absorbs very strongly in the Near-IR region;
Water exhibits very similar absorption spectra to glucose [63], and absorbs
very strongly in Near-IR region [64]. Glucose absorption peaks has small mag-
nitude compared to a large aqueous background spectrum, and this yields low
signal-to-noise ratio.
3
Also known as Molar absorptivity, molar extinction coefficient is the absorbance
A of a solution divided by the product of the optical path l [cm] and the molar
concentration C of the absorbing species [65], i.e. = A/(l × C).
4
There are many types of molecular vibrations. Readers are invited to consult general
textbooks on Analytical Chemistry for more details.
5
Visible spectrum ranges from 380 nm (violet) to 700 nm (red) approximately, with
no clear boundaries between colours.
0.002
~964μm
2
0.001
~943μm
in vitro glucose absorption
0
~921μm
1.5
−0.001
~901μm
Absorbance
Absorbance
−0.002
~870μm
~1004μm
1 −0.003
~697μm
~732μm
−0.004
~1087μm
0.5 −0.005
−0.006
0 −0.007
600 700 800 900 1000 1100 1200 2.05 2.1 2.15 2.2 2.25 2.3 2.35
Wavelength (μm) Wavelength (μm)
Fig. 2.4. Glucose absorption spectrum (Adapted from McNichols et al [38])
• strong light scattering by skin [63], as well as other substances and compounds
in tissues (such as blood cells) [41];
• overlaps in light absorption by molecules introduces interference into the ab-
sorbance measurement;
Since biological molecules have complicated structure, they exhibit a large
number of similar IR absorption peaks that are often overlapping [66] (e.g.
glucose has more than 20 absorption peaks in the wavelength region of 2.5–
10 μm). Not all of these peaks are specific for glucose [63]. The prominent
absorption peak of glucose is around 9.7 μm. In addition, haemoglobin is the
most dominant component in blood, with a concentration level of more than
100 times of glucose. Its absorbance becomes increasingly strong toward short
Near-IR and visible wavelengths. Although haemoglobin absorption peaks do
not interfere with those of other components, its influence is by no means
negligible due to its high concentration [69].
• pronounced temperature dependence of light absorption [63, 69].
Because of the strong absorbance of water, Near-IR transmittance measurement
should use regions between the water bands where sufficient amount of light can
be transmitted [70]:
• Near-IR region between 0.7–1.3 μm — This range contains higher orders of
glucose overtone regions [70]. Nevertheless, this region also shows very little
glucose absorption (i.e. less than 0.1% of the fundamental region of glucose
at 9–9.6 μm) despite having lower light absorption by water [69].
• Near-IR region between 1.5–2.5 μm — This range was reported to be a suitable
for noninvasive glucose monitoring [69], with highest glucose absorption at
1.575–1.7 μm (which corresponds to the overtone band of C-H stretching
mode [71]). Most of the analytes of interest display distinguishing absorption
features that do not suffer from excessive attenuation by water [72].
• Near-IR region between 2.0–2.5 μm –This range is also popular for aqueous
glucose measurements because it contains a relative minimum in the water
absorption spectrum and readily identifiable glucose peak information.
With all the confounding issues, extra signal processing (in conjunction with
Near-IR measurement) became necessary, e.g. digital filtering to correct for tem-
perature sensitivity of the Near-IR spectrum, and multivariate calibration tech-
niques. Sensing site for Near-IR measurement included fingertip [41], earlobe,
tongue, lip [73], forearm, oral mucosa surface and aqueous humour [38].
Mid-IR [2.5–50 μm or 4000–200 cm−1 ]
Mid-IR spectrum ranges from 2.5–50 μm [67], and this region is often called
“fingerprint” region. This is because the absorption bands in this region are
due to the fundamental stretching and bending modes of the molecule. Mid-IR
spectrum is very useful for spectral identification of compounds [38].
The Mid-IR range of 5–13 μm contains absorbance fingerprints generated by
biologically important molecules such as glucose, proteins and water [74]. The
prominent absorption peak of glucose is reported to be around 9.7 μm [75, 76].
However, water absorbs Mid-IR radiation very effectively [63,76]. Background
absorption bands of other solution constituents are also strong, and thus limits
the depth and path length of Mid-IR transmission spectroscopy to a few hundred
microns. [38,63,77]. Alternative measurement techniques using Mid-IR radiation
have been proposed (see Section 2.4.2).
Implementation
The basic setup of the IR measuring instrument consists of a light source, an

optical assembly (that directs the light at a sample and collects the transmitted
or reflected light), and a light detector unit (which contains a spectrometer if
measuring over multiple wavelength). The measured spectra were then analysed
using mathematical procedures, most commonly the Multivariate Calibration
methods like PLS (Partial Least Square) and PCR (Principal Component Re-
gression), or a Neural network [78]. Multivariate calibration (in the simplest
form) involves multiplying the (measured and processed) spectrum Ai by a vec-
tor of calibration coefficients bi , and summing the result over all wavelengths to
obtain the analytes concentration c [72]:

c= Ai bi + constant
i
Some examples of IR absorptiometry instrumentation (and not meant to be

exhaustive),
• Robinson et al [79] used an FTIR-spectrometer with a tungsten halogen source
and a liquid-nitrogen cooled InSb (indium antimonide) detector (or a gating
gating spectrometer, tungsten halogen source and Si array detector), over the
frequency range 870–1300 nm. PLS (Partial least square) and PCR (principal
component regression) calibration were applied to define an empirical model
relating changes in the calibration spectra to the differing reference glucose
concentrations of blood samples obtained during the calibration. Burmeister
et al [70] and Small et al [80] also used FTIR spectrometer with a tungsten-
halogen lamp to collect the Near-IR spectra.
• Hiese et al [81] used attenuated total reflection (ATR)6 and FTIR spec-
troscopy to estimate the blood glucose concentration in whole blood sample
(taken out of subjects) over the wavelength range 1500–750 cm−1 (or ∼660–
1300 nm). Multivariable calibration with the PLS algorithm was performed
on the spectral data.
• Tenhunen et al [82] used a fibre bundle as a transflectance probe to direct the
light from a halogan lamp to achieve an appropriate penetration depth into the
tissue. Temperature stabilised IR sensitive detector was used for measuring
the interferogram of the FTIR-spectrometer.
• Yoon et al [71] did not use a spectrometer, but used custom made detectors
to measure the absorbance of the Near-IR light. A minimum of three dif-
ferent wavelengths were selected for determination of glucose (λ1 =1625 nm,
λ2 =1365 nm, λ3 =1200 nm), based on satisfying a simple but efficient algo-
rithm that will estimate the glucose concentration. Instruments were setup
up carefully to minimise reflected light from skin surface and effects of tissue
scattering.
Fig. 2.5. Attenuated Total Reflection: When an infrared beam is directed onto an
optically dense crystal with a high refractive index at a certain angle, the internal
reflectance creates an evanescent wave that extends a few micrometer (0.5–5 μm) be-
yond the crystal surface into the sample held in contact with the crystal. As the sample
absorbs the relevant infrared spectrum, the evanescent wave will be attenuated. The
attenuated energy from each evanescent wave then passes back to the IR beam, which
then exits the opposite end of the crystal and is passed to the detector in the IR
spectrometer (Adapted from PerkinElmer Inc [83]).
6
ATR is a sampling technique that allows samples to be examined by a spectrometer
directly in solid or liquid state without extra preparation. It measures the changes
that occur in a totally internally-reflected infrared beam when the beam comes into
contact with a sample. Kaiser [73] was among the earliest teams to combine ATR
prism with laser absorption spectrometry for measurement of glucose on the lips.
Table 2.2. FTIR-spectrometer
Fourier-Transform Infrared Spectrometer (FTIR)
Spectrometer (or spectroscope) is an optical instrument that disperses light

into its spectrum, allowing measurement of the individual wavelengths and
intensities.
In the original spectrometer design, light entered a slit and a collimating
lens transformed the light into a thin beam of parallel rays. A prism then
separated the beam into its spectrum. The observer then viewed the spec-
trum through a tube.
Later design used diffraction gratings in place of the prism. Diffraction grat-
ings have fine parallel and equally spaced grooves on the surface. Each wave-
length of the incident beam spectrum is sent into a different direction when
the gratings is illuminated by light. (Every day examples of gratings are CD
and DVD media).
A Fourier-Transform Infrared (FTIR) spectrometer uses interferometric
method to measure the spectrum, in contrast to the classical spectrome-
ters. In its simplest form, an FTIR spectrometer consists of an infrared
light source, a beam splitter, a fixed mirror, a moving scanning mirror and
an infrared detector. This is the typical Michelson Interferometer design.
The IR beam is guided through the interferometer. After passing through

the sample, the amplitudes of the waves are combined either constructively
or destructively to form an interferogram (i.e. a plot of the output power
from the detector as function of the difference in path-length of the two
beams [67]). FTIR measured all the wavelengths simultaneously with a sin-
gle detector element [65, 72]. By performing a mathematical Fourier Trans-
form on this signal, a spectrum identical to that from conventional (disper-
sive) infrared spectroscopy can be obtained.
Different types of IR detector can be used. One common choice is indium
gallium arsenide (InGaAs) detector element because it has good specific
detectivity and excellent amplitude linearity over a large dynamic range
required for interferometric signals [72]. It also cost less than the array
detectors used in gating spectrometers.
It should be noted that as the IR wavelength is lengthened, the usable optical

path length shortens because of more scattering and absorption in the tissue [72].
Reflection measurements, or back-scattering become more effective than trans-
mission measurements in this scenario (see Section 2.4.3). Readers interested in
further details on infrared glucose measurements are encouraged to read articles
such as [84] and [85].
2.4.2 Mid-IR Emission Spectrometry
In contrast to infrared absorptiometry, the Mid-IR emission spectroscopy is

based on the principle that the human body naturally emits strong electro-
magnetic radiation (or blackbody radiation), and the natural Mid-IR emission
from the human body is modulated by the state of the emitting tissue [86]. The
light source in this technology is the human body’s natural emission of Mid-IR
light (no external light source needed). The measurement techniques reported
in the literature are:
• Temperature gradient (Cooling the surface [74, 87, 88])
In the Mid-IR range, glucose absorbs strongly with minimal interference from
other species [87]. However, glucose is not measureable at the same temper-
ature as the tissue that emit the blackbody radiation [74, 87]. To solve this
problem, a temperature gradient is introduced, where the surface is cooled
so that the cooler glucose will absorb more than it will emit. In another
words, when there is a temperature gradient, the radiation intensity (which
is wavelength-dependent) will deviate from the original Planck’s prediction
(of blackbody emission). At the wavelengths where glucose absorbs, the to-
tal energy emitted will fall below the Planck curve [74]. This would yield a
reasonable glucose absorption spectrum superimposed on the usual smooth
blackbody spectrum [87].
• Tympanic membrane measurement [86]

Malchoff et al described the detection of glucose from the tympanic membrane
using room temperature detectors and a specially designed non-dispersive filter-
based spectrometer. An IR wave-guide with gold-plated inner surface directs
IR radiation into a circular variable IR filter in close proximity. The circular
variable IR filter has transmission band of 7.7–14.1 μm, and when rotated in
the IR light path would generate variable passbands. A thermopile detector was
placed very close to the filter surface to capture the light output from the filter.
The detector window has two more filters – one that passes thermal emission
bands with glucose signature; the other passes radiation that does not include
emission bands characteristics of glucose at wavelengths in the range of interest.
2.4.3 Scatter Change
Light scattering is the next major optical interaction with tissue after light
absorption [89]. Light scattering occurs when light interacts with variations in
the refractive index7 or small particles acting as scattering centres in the medium,
resulting in light dispersing from a straight trajectory. When scattering centers
are grouped together, the light may get scattered many times (i.e. multiple
scattering).
In tissue, most of the light scattering can be attributed to Mie scattering,
which is related to:
• the the size of the scattering particles;
Mie scattering takes place when the size of the scattering particles and the
wavelength of light are in the same order of magnitude [63].
• the magnitude of the refractive index mismatch between the particles and
their medium [90].
The amount of the scattering depends on the ratio between the refractive
index of the scattering particles (e.g. cell membranes) and the surrounding
medium (e.g. extracellular fluids). If the difference between the refractive in-
dices is pronounced, light scattering is larger. With identical refractive indices,
the media is transparent [63, 91].
Fig. 2.6. Scattering of light (Adapted from Heinemann et al [63])
Due to the free exchange of glucose between blood plasma and interstitial
fluid, changes in glucose concentrations of these two fluids are closely correlated.
An increase in glucose concentration leads to a rise of the refractive indices
of blood plasma and ISF (as glucose can displace the water from the fluids
[92]), whereas the refractive index of the scattering particles (red blood cell,
cell membranes, collagen fibres etc) is assumed to remain relatively unchanged
[63, 91, 93]. Since the refractive index of ISF is lower than that of the scattering
particles, an increase in glucose concentration causes a reduction of the refractive
7
The refractive index (or index of refraction) of a medium is a measure for how much
the speed of light is reduced relative to vaccum when the light travels inside the
medium. For example, typical glass has a refractive index of 1.5, which means that
light travels at 1 / 1.5 = 0.67 times the speed in air or vacuum [65].
index mismatch between ISF and scattering particles. The increase in blood
glucose concentration effectively decreases the scattering coefficient of skin [63].
The scattering coefficient is a factor that expresses the attenuation caused by
scattering:
ncell
μs = f ρ, a, g,
cmedium
where ρ = number of density of scattering cells in the observation volume, a
= diameter of the cells, g = anisotropy factor (the average cosine of the angle
at which a photon is scattered), ncell = the refractive index of the cells, and
nmedium = the refractive index of ISF [94].
It should be noted that the scattering coefficient of the skin is not directly
influenced by glucose, but indirectly via the refractive indices [63]. The glucose
effect on scattering coefficient is not specific to glucose molecule only. The glucose
concentration measured from scatter change detection is derived from changes
in refractive index “induced by glucose”. Other blood analytes and physiological
factors may influence the scattering coefficient, for example:
• mitochondria, which is the main scatterers in the skin [95];
• temperature [63, 92];
• changes in the intracellular refractive index;
• changes in cell size due to by osmolarity changes [91].
Implementation
To measure the scattering coefficient, a diffuse reflectance measurement is usually

used. A narrow beam of light is directed through the skin surface into the tissue.
Some of the light is absorbed by the tissue while a certain portion of the light is
diffusely scattered and reflected back to the skin surface (after interacting with
the tissue, or exciting vibrational modes of the analyte molecule in the tissue).
The intensity of this diffuse reflectance is dependent on both the scattering
coefficient and the absorption coefficient of the tissue.
“By measuring the reflected intensities at different distances from the light
entry point, an intensity profile (i.e. diffuse reflectance vs. distance) can be
recorded. The reflectance measured close to the light entry point is mainly in-
fluenced by the scattering of the skin, while reflectance further from the light
source is affected by both the absorption and the scattering properties of the
skin.” (Adapted from Heinemann et al [89]).
The effects of light scattering and light absorption from the recorded light
intensity can be differentiated by algorithms designed based on diffusion theory
of light propagation in tissue [63]. A neural-network can be used to extract the
optical properties from the reflectance data [91].
Examples of the measuring instruments include diode-array spectrometer with
InGaAs detector [78], FTIR spectrometer with nitrogen-cooled InSb detectors
[96], LED arrays with photodetector arrays [63], and He-Ne laser beam with
CCD camera [97].
Fig. 2.7. Diffused reflectance for measuring scatter change (Adapted from Heinemann
et al [63])
Optical Coherence Tomography has also been used to measure glucose based
on the fact that OCT is very sensitive to the detection of changes in refractive
indices of the sample, and thus the scattering properties of the sample. A reduc-
tion in the sample scattering property lowers the intensity of the backscattered
photon. Larin et al [98] used OCT techniques to scan the skin up to a depth of
1 mm for measurement of tissue scattering properties at a specific layer in the
skin. However, since OCT measures only relative changes in the scattering prop-
erties, other osmolytes of the ISF that can change the refractive index mismatch
have the potential to interfere with glucose measurement [95].
2.4.4 Tissue Manipulation and Red/Near-IR Absorptiometry

( λ ∼ 630 -- 1300nm)
In most Near-IR absorptiometry techniques considered above, optical non-

invasive measurements were conducted using “DC measurement technique”,
where light illuminates blood-perfused tissue, and the reflection/transmission
effect is studied. All spectrally active components of the tissues (e.g. skin, blood,
muscle, fat etc) contribute to the observed effect. Orsense Inc (Nes Ziona, Isreal;
http://www.orsense.com) in [101] considered an “AC measurement technique”,
which focuses only on the “blood” component of the tissue, by (for example)
looking at the difference in light absorption of the tissue measured during the
systole (blood pressure rise) and the diastole (blood pressure fall) period.
Table 2.3. Time-domain Optical Coherence Tomography (OCT)
Optical Coherence Tomography
Time-domain OCT is based on low-coherence interferometry (i.e. interferome-

ter with a low coherence, broad bandwidth light source). It consists of a low-
coherent light source, a detector, a reference arm and and a sample arm. The
light is split into the reference and sample arms, and recombined at the detector.
The combination of reflected light from the sample arm and the reference light
from the reference arm give rise to an interference pattern. OCT is different
from conventional interferometry in that the interference occurs over a distance
of micrometers (rather than meters).
OCT detects the backscattered photons. By changing the length of the refer-
ence arm (i.e. scanning the mirror), light backscattered or reflected back from
the sample combines with light reflected from the mirror to result in either a
constructive or destructive interference [99]. If the length of the sample arm
is also changed, then photons from different depths of the sample can also be
detected.
Photon detection with OCT requires that the path-length difference of the pho-
tons coming from the sample arm and from the reference arm be smaller than
the coherence length of the light source. The achievable depth resolution is gov-
erned by the coherence length of the light source, and the length is inversely
proportional to the light’s bandwidth. Over small distances the OCT signal
is formed by single backscattered photons, but in deeper scattering samples
the multiple-scattering effect (i.e. “echo” [100]) has to be considered. (Adapted
from [95]).
“Occlusion spectroscopy” (or Occlusion Red Near-InfraRed Spectroscopy,

O-RNIRS) (see [101, 102]) uses this concept of AC measurement, based on the
observation that a cessation of blood flow triggers a change of the optical charac-
teristics of blood [103]. It was discover that when blood flow is suddenly stopped,
the average size of scattering particles in blood changes, allowing more light to
transmit through the tissue.
Before a stop of blood flow is imposed, red blood cell (RBC) aggregates less
due to the force of the ongoing blood flow. When a temporary over-systolic
pressure is introduced, there is a stop of blood flow, and the RBC starts to
aggregate. The increase in the size of the blood scattering particles cause a
change in the scattering coefficient, which is related to the mismatch between
the RBC and the plasma/ISF refraction index [102–104] (see Section 2.4.3).
Blood parameters such as hemoglobin, glucose, oxygen saturation, etc. in-
fluence the light transmission following over-systolic occlusion. For example, an
increased glucose concentration reduces the refractive index mismatch, while a
decreasing glucose concentration increases the refractive index mismatch, caus-
ing light transmission to decrease.
Hence, one way to measure the glucose value (using the above observation) is
by looking at the difference in light absorption of the tissue measured during the
systole and the diastole period. The systole and diastole are generally caused
by the pumping action of the heart (or heartbeat), but can be simulated by
tissue manipulation (i.e. applying artificial pressure to the tissue to cause a
volumetric change in the blood, e.g. by pressing on the finger). It was reported
that the glucose in arterialized whole blood correlated with glucose prediction
by O-RNIRS [103].
Yamakoshi et al [105] used the concept of AC measurement in “Pulse glucom-
etry” technique, where Near-IR transmittance spectra was taken during normal
blood volume pulsations throughout the cardiac cycle. PLS calibration is used
to predict the BG values. Their approach differs to “occlusion spectrometry” in
that no pressure was applied to the tissue during measurement.
2.4.5 Raman Spectroscopy
Raman spectroscopy is a spectroscopic technique based on inelastic scattering

of monochromatic light [106]. There are generally two types of scattering8 :
• elastic scattering – Also known as Rayleigh scatter, the incident photon
“bounces” off the molecule with no gain or loss of energy, and thus the
scattered photon has the same energy (frequency and wavelength) as the
incident photons. In Rayleigh scattering, the dimension of the molecules or
aggregates of molecules are significantly smaller than the wavelength of radi-
ation [67].
• inelastic scattering – Also known as Raman scatter, the incident photon
exchange energy with the molecule, leading to the emission of another
photon with a different frequency than that of the incident photon [67,
107].
8
If a radiation is extinguished when it interacts with a molecule, then the process is
known as “absorption”.
When the photon transfers energy to (or from) the molecule during an inelastic
collision, the energy difference between the incident light (Ei ) and the Raman
scattered light (Es ) is equal to the energy involved in getting the molecule to
vibrate. This energy difference is called the Raman shift [108]:
Ev = Ei − Es
The loss (Strokes shift) or gain (anti-Stokes shift) of photon energy owing to
the transitions of rotational and vibrational energy states within the scatter-
ing molecule causes the frequency of the scattered photon to shift up or down
in comparison with original monochromatic frequency. The vibrational spectra
produced as a result provide specific information about the chemical structure
of the sample [65, 109], and reveals the interactions between the molecule and
its local chemical environment [110].
Note that the observed Raman shifts are independent of the excitation fre-
quency [67], and thus an excitation frequency may be chosen which is appro-
priate for a particular sample. However, the intensity of Raman scattered peaks
generally falls off with decreasing frequency [109].
Most incident photons undergo elastic Rayleigh scattering, and about 0.001%
of the incident light produces Raman signal [65]. Since this signal is very
weak [108], special measures should be taken to distinguish it from the dom-
inant Rayleigh scattering [106]. Using Raman spectroscopy for in vivo transcu-
taneous glucose measurement is difficult because whole blood and most tissue
are highly absorptive and containing many fluorescent and Raman-active con-
founders [110].9
Instrumentation
Fig 2.8 shows an example of the Raman spectroscopy setup used in [111]. Typi-
cally, the instrument consists of:
1. Excitation source (laser)
2. Wavelength selector (bandpass filter, holographic gating)
3. Sample illumination assembly and light collection optics
4. Detector (photodiode array, CCD or photon-multiplier tube).
A sample is normally illuminated with a laser beam in the ultraviolet, visible or
near infrared range. Scattered light is collected with a lens and is sent through
the detector to obtain Raman spectrum of a sample.
To improve the Raman signal intensity, there are many other variations of
Raman spectroscopy that have been developed, e.g. Surface Enhanced Raman
Spectroscopy, Resonance Raman spectroscopy, to name a few (See [67]).
9
Other site could be chosen, for example, the aqueous humor, which is relatively non-
absorptive and contains few Raman-active molecules. Raman excitation wavelength
in the Near-IR region can be used to minimise biological fluorescence and tissue
damage [110].
Fig. 2.8. Raman spectroscopy used in Berger et al’s experiment (Adapted from [111])
2.4.6 Polarimetry
Polarimetric measurement of glucose concentration is based on the phenomenon

of optical rotation (or optical activity), whereby a solution containing a chiral
molecule rotates the plane of polarisation for linearly polarized light passing
through it.
This rotation is the result of a difference in refractive indices ηL and ηR for
left and right circularly polarised light traveling through the electron cloud of
a molecule. It occurs because of the chirality or “handedness” of the molecule
(i.e. the molecule has at least one centre about which its mirror image cannot
be superimposed upon itself) [38].
In pure liquids, the angle of the rotation depends linearly on the optical path-
length L [dm], the density of the liquid D [g/ml] and a constant for the liquid
[α]Tλ :
T
φ = [α]λ × L × D
The constant [α]Tλ is known as the specific rotation10 , for a wavelength λ [nm]
and measurement temperature T [degree Celsius].
10
Specific rotation [α]Tλ is defined as the number of degrees of optical rotation φ ob-
served when plane-polarized light is passed through a material with a path length
of 1 decimetre and the sample concentration of 1 gram per 100 ml [65]. For solution,
its unit is [ ◦ dm−1 (g/100 ml)−1 ]. For pure liquid, the liquid’s density replaces the
solution concentration, i.e., the unit is [ ◦ dm−1 (g/ml)−1 ].
Table 2.4. Optical activity
Optical Activity
A linearly polarised light can be regarded as an equal combination of right

and left circularly polarised light with equal amplitude. As illustrated in
[112],
where θR = − 2πL c
λ vR
, θL = 2πL
λ vL
c
,λ = wavelength of the light (in vaccum),
c = speed of light, L = length of material, vR and vL = velocity of the
right and left circularly polarised components in the medium. The angle of
rotation is
1 πL c c
Δθ = (θL + θR ) = −
2 λ vL vR
In an optically active material, the two circular polarisations will experi-
ence different travelling velocities. This difference (also called optical activ-
ity strength) is characteristic of the material. Since the refractive index is
defined as η = c/v, the angle of rotation of the light that passes through
the material can be expressed as
πL
Δθ = (ηL − ηR )
λ
A circularly polarised light traces out a circle as the electric wave propagates
[113]. The final polarisation is rotated to angle θ + Δθ. From this equation,
the degree of rotation depends on the colour (frequency) of the light, the
path length L and the properties of the material.
In a large aggregate of randomly oriented molecules, the net rotation by the
material averages to zero. However, in the case where the molecule is chiral,
the rotation is additive [109].
In solution, the equation is [114–116]:
T L×C
φ = [α]λ
100
where L = optical path length [dm] and C = concentration [g/100 ml] for a sam-
ple at wavelength λ [nm] and temperature T [degree Celsius]. If the wavelength
of the light used is Yellow Sodium D line at 589 nm, and the temperature is
20◦ C, then the specific rotation can be written as [α]20 o

589 = +52.6 (see [65]) or
20
[α]D = +52.6o (for Sodium “D” line).
Glucose in the body is dextro-rotatory (rotates light in the right-handed direc-
tion) and has a specific rotation of +52.6◦ dm−1 (g/ml)−1 at the sodium D-line
of 589 nm [38, 109].
Fig. 2.9. Glucose is an equilibrium mixture of 2 enantiomers (i.e. nonsuperimposable

mirror images of each other [65]): α-D-glucose and β-D-glucose. When pure α-D-glucose
is dissolved in water, the specific rotation of the sample decreases from +112◦ to +52.6◦ .
When pure β-D-glucose is dissolved in water, the specific rotation increases from +19◦
to +52.6◦ .
T
Generally, [α]λ for glucose decreases with increasing wavelength across the
visible light spectrum and has a rise in magnitude near the optical absorption
bands of a particular molecule [38, 114]. The variation in the optical rotation of
a substance with a change in the wavelength of light is called Optical rotatory
dispersion (ORD) [65], and the observed rotation is
π
φ= (ηL − ηR )
λ
where λ = wavelength, ηL = refractive index of left-circularly polarised light, and
ηR = refractive index of right-circularly polarised light.
Instrumentation
The general setup for polarimetric measurement is illustrated in Fig 2.10 and
2.11.
In the simplest form, the light travels from the source is linearly polarised by
polariser 1. After passing through the sample, the polarisation plane will rotate
due to the optical activity of the sample. The second polariser (also known as
the analyser) is placed perpendicular to the first polariser. If there is no sample,
then theoretically no light will be transmitter to the detector from the second
polariser [117]. If an optically active sample is introduced, then the intensity
of the light incident on the detector will be proportional to the square of the
amplitude of the Electric-field passing through the second polariser. The E-field
is proportional to the sine of the rotation angle ϕ by the sample [109].
An optical modulator (such as a Faraday rotator) can also be placed in be-
tween the sample and second polariser. The modulator causes the angle of the
Fig. 2.10. Polarimetry glucose measurement (Adapted from McNichols et al [38]).

Linearly polarised light is passed through a material, and the material “twisted” the
polarisation angle of the light. This results in a varied light intensity when measured
at the detector.
polarization vector to vary periodically [109]. The introduction of rotation of the

linear polarisation vector of the light beam would eliminate amplitude variations
due to fluctuations in the light source [118, 119].
Apart from measuring the light amplitude, the phase of the incident and
transmitted light can be measured instead [118]. There are other implementa-
tions of the polarimetry system in existence, e.g. using a coil wound around the
solution instead of a Faraday rotator to reduce cost [120], signal enhancement
using digital closed-loop control of the polarimetric system to improve system
reliability and stability [114], using optical heterodyne polarimeter [121, 122], or
using the effect of Magnetic Optical Rotatory effect (MORE) where the optical
rotations are multiplied on reflecting back through a medium in the presence of
magnetic field [112]. MORE was used by Jang et al [123] to double the angle of
rotation.
Obstacles to Polarimetric Glucose Measurement in Human
Human skin exhibit significant light scattering [124] and can depolarise the light
passing through it. Hence it is not a feasible site for polarimetric measure-
ment. The aqueous humour in the anterior chamber of the eye has relatively
Fig. 2.11. Polarimetry glucose measurement by phase difference (Adapted from McNi-
chols et al [38]). Instead of measuring the polarisation angle directly as light intensity,
the angle is measured as a difference in the phase of incident and transmitted light.
low scattering properties, and has shown close resemblance of glucose levels in
the blood [117], with a time lag of minutes [109, 114]. Nevertheless, glucose is
not the only optically active component in the aqueous humour, and there are
other potential problems with polarimetric glucose sensing in the eye including
optical rotation due to the cornea (in addition to aqueous humor), birefringence
of the corneas, and motion artifacts [38].
2.4.7 Photoacoustic
Photoacoustic (PA) refers to the generation of acoustic waves by modulated

optical radiation [125]. The PA effect was discovered by A.G. Bell (1880) when
he observed that audible sound is produced when chopped sunlight is incident
on optically absorbing materials [75].
Most PA generation is due to “photothermal” (PT) effect, i.e. the heating of
the sample after the absorption of optical energy [125]. PT heating of a sample
is frequently produced with the use of laser beams.
In PA glucose measurement, the tissue is illuminated by pulses of low-energy

laser set at a wavelength that is absorbed by glucose. The light pulses are fo-
cused to a relatively small region. When the light pulses are absorbed by glucose,
the energy is converted into kinetic energy, which causes the temperature and
pressure of the absorbing tissue to increase. This generates acoustic waves that
radiate out from the region [126]. The PA waves can be detected using acoustic
sensor such as microphone, piezoelectric transducer [127], or capacitance trans-
ducers. Other ways to detect the PA waves are, for example, shining a light
(so-called a probe beam) onto the surface and detecting the deflection of the
beam reflected from the surface, or interferometry, or detecting the refractive-
index change associated with the PA pulse [128, 129].
The peak pressure of the PA wave can be described by:
βv n
P =K E0 μa
Cp
where K = system constant, β = volumetric thermal expansion coefficient, v =

sound velocity in the medium, n is between 1 and 2 (depending on experimental
condition), Cp = specific heat capacity, E0 = incident laser pulse energy, and μa
= optical absorption coefficient [76].
When glucose concentration increases, Cp decreases while the acoustic velocity
v increases. At glucose absorption wavelength, changes in PA signals are the
result of changes in μa , v, and Cp [94]. Hence, the intensity of the PA waves can
be related to the concentration of glucose in the absorbing tissue region.
The PA wave can be generated with:
• one excitation wavelength (e.g, Zhao & Myllylá [129] used wavelength of
905 nm and short low-energy pulse)
• over a range of excitation wavelength (e.g. Quan et al [130] scanned across
the wavelength of 0.75–1.75 μm)
using
• point-by-point measurement – where the wavelength is shone on a spot on the
tissue, and the PA signal is recorded.
A range of wavelengths can also be scanned, with the corresponding PA signal
for each wavelength recorded. Mackenzie et al [76] used 800–1200 nm wave-
length at 7 ns pulse repeated at 10 Hz.
• multiplexing – the tissue is simultaneously excited by many wavelengths [85,
131].
The multiplexed signal can then be transformed back into the conventional PA
spectrum using suitable transform techniques (e.g. Fourier transform). (E.g.
Spanner et al [132] used three wavelengths 834, 1304 and 1554 nm continuous-
wave output powers at one focal point).
Fig. 2.12. Photoacoustic glucose measurement instrument setup, as used by Glucon

Inc [126]
Fig 2.12 shows the instrumentation aspect of a photoacoustic glucose measure-

ment system (AspriseTM sensor) by Glucon Inc (Boston, MA, USA), described
in [126] and [133].
PA response may still be interfered by other factors such as physiological
change, blood circulation, water content and body temperature drift [128], but
preliminary results from clinical testing showed promising result (See [131]).
Readers interested in further details on photoacoustics are encouraged to read
[125] and [128].
2.4.8 Light Diffraction
Measurement of glucose based on light diffraction has been reported in the lit-
erature. The technique used material that can swell or shrink when interacting
with glucose [134, 135]. Light diffraction is governed by Bragg’s law [134, 136]:
mλ = 2nD · sin θ
where m = order of diffraction, λ = wavelength of the light in vaccum, n =
average refractive index of the system (hydrogel, colloid), d = spacing between
the diffracting planes/fringes, and θ = glancing angle between the incident light
direction and the diffracting planes.
Glucose sensor constructed using this technique used:
• Crystalline colloidal array in hydrogel [134]
Glucose oxidase and an enzyme were incorporated into a colloidal crystalline
array (CCA). When glucose oxidase reacts with glucose, the enzyme became
negatively charged, causing the hydrogel to expand. Because the brightly
coloured CCA diffracts visible light according to the Bragg condition, any
change in the size of the hydrogel can cause a change in the wavelength or
colour of the diffracted light.
• Holographic sensor in hydrogel [135–137]
This sensor is based on planar small-volume polymer hydrogels containing
silver halide, which acts as a holographic recording material. The hydrogel
contains the reflection hologram that reflects a narrow band of wavelengths
when illuminated with white light. By incorporating ligands that interact with
glucose (via competitive-binding), the hydrogel can change its swelling state
upon analyte binding, which in turn causes a change in the wavelength of
light diffracted by the holographic grating within the hydrogel. The wave-
length is determined by the spacing between the fringes of the hologram
(where an increase in the fringes increases would cause longer wavelengths
to be diffracted). A holographic contact lens glucose sensor has been clini-
cally trialled (see [137]).
2.4.9 Dielectric Spectroscopy
It was known that small changes in glucose concentration in blood can induce
significant reactions in tissues involved in the metabolism of carbohydrate (such
as liver, pancreas and red blood cell) [138]. One of the response of these tis-
sues to increasing glucose concentration is a decrease of sodium and increase of
potassium due to water movement from the tissue and red blood cell into the
vascular system [138]. The variation of the electrolyte balance in the blood leads
to a change in membrane potentials on the cellular level. Changes in membrane
potentials can be recorded with dielectric spectroscopy.
In dielectric spectroscopy (or impedance spectroscopy), a small AC current
is applied and the impedance of the tissue to the current flow is recorded as a
function of frequency [139]. For non-invasive glucose monitoring, changes in skin
impedance is measured. Skin impedance is sensitive to changes in membrane
potential, which is influenced by the interaction of glucose with red blood cell.
Caduff et al [140] reported a non-invasive measurement of glucose using
impedance spectroscopy. Fig 2.13 shows the equivalent circuit of the sensor
mounted on the skin, attached to the impedance measuring network. The
impedance of the sensor at a given resonance frequency depends on impedance
changes within the human skin and underlying tissue, and also the skin temper-
ature [140]. A frequency sweep across 1–200 MHz and temperature measurement
were performed every minute to calculate the impedance minima |Z|min . The
sensor system was also incorporated into a wristwatch device.
Although the measured signal have a closer correlation to glucose changes in
blood than to those in ISF, a few factors remain to be addressed, e.g. sweat
and relocation of the sensor system (in order to obtain reliable glucose related
signals), temperature variations, and spikes in registered signals due to move-
ment [140].
Fig. 2.13. Sensor for measurement of skin impedance (Adapted from Caduff et al
[140]). R is the averaged resistance of the skin and underlying tissue. The voltage VREF
from a Voltage-Controlled-Oscillator is fed to a resistive-divider network consisting of
a series resistor RS and the sensor impedance Z. The measured sensor voltage is VSENS
and the impedance can be approximated by ZSENS ≈ Rs VREF VSEN S
−VSEN S
.
2.5 Amperometric Sensor Calibration

2.5.1 General Calibration
Since the current produced in the amperometric method is proportional to the

level of glucose present, a calibration factor (or constant of proportionality) can
be determined, and then used to convert all subsequent current measurements
into an estimate of glucose concentration.
In implantable sensors, calibration methods based on “in-vitro”, “in-vivo”
and a combination of both have been suggested. In-vitro calibration determines
the signal-to-concentration conversion factor using external glucose solutions
with known concentration, while in-vivo calibration attempts to establish the
relationship by matching the sensor signal to the glucose concentration in the
blood. Examples of calibration methods are:
• “one-point in-vitro, one-point in-vivo” calibration [26];
i.e. an in-vitro calibration is done before sensor implantation, followed by an
in-vivo calibration after implantation. The in-vitro calibration is performed
with the whole sensor system perfused in sterile phosphate buffered saline.
• “two-point in-vivo” [141];
i.e. testing of the sensor’s response using two in-vivo BGs and then applying
a linear conversion to adjust the signal. Two points can be taken when an
externally infused insulin or glucose caused BG to reach a new plateau or
steady-state. The points are used to calculate an “in vivo sensitivity” coeffi-
cient which is expressed as nA/(mmol/l).
2.5 Amperometric Sensor Calibration 37
• “Wick” technique [142];

i.e. the simultaneous measurement of glucose concentration in peripheral ve-
nous plasma blood and in centrifuged wick fluid. The wick is made from a
double-stranded cotton threads and is positioned subcutaneously to absorb
the glucose, and then removed after an appropriate interval. The fluid ab-
sorbed by the wick is then centrifuged, and the extracted glucose is correlated
with those obtained from the plasma blood.
2.5.2 Regression Calibration
In contrast, MiniMed uses “Regression calibration” to calculate the calibration

factor, or rather, a set of calibration constants. Regression calibration is a modi-
fied form of linear regression. A regression slope is first calculated with an initial
fixed intercept of zero, by a linear regression of a minimum of four calibration
BG level entries, paired with corresponding valid sensor readings (called Isig
with unit in nA) for each 24-hour period (see [143]).
Mathematically, if we define:
xi = i-th current measured by sensor (i.e. Isig) [nA]
x̄ = background current of sensor in the absence of glucose [nA]
yi = i-th calibration BG value (from glucometer) [mg/dl]
ȳ = calibration BG value in the absence of sensor current [mg/dl]
Then least square regression line of y on x is

Sxy
y = a + bxy · x with bxy = and a = ȳ − bxy · x̄
Sx2
where

n
(xi − x̄) (yi − ȳ)
i=1
Sxy =
n
and

n
2
(xi − x̄)
i=1
Sx2 =
n
with n-pairings of glucometer BG value and Isig each day, i.e. (x1 , y1 ), (x2 , y2 ),...,
(xn , yn ) etc.
If the resulting slope bxy is less than 7 mg/dl per nA, then an offset of 3 nA
is subtracted from each valid Isig, and the regression slope is re-calculated. The
BG level is then estimated from each Isig by:

bxy · Isig, if bxy ≥ 7 mg/dl per nA
BGestimate =
bxy · (Isig − 3) , if bxy < 7 mg/dl per nA
The calibration BG level entries are based on BG measurements taken using

a conventional glucometer. The user enters the glucometer readings into CGMS
every day, while the monitor is in use. Isig values are deemed to be valid if they
are within the range 10–100 nA. This calibration method can be viewed as a
multi-point in-vivo calibration, with the exception that MiniMed has presented
it for use in an “offline” manner, rather than for real-time estimation.
2.5.3 GlucoWatch Calibration

GlucoWatch is a wrist-watch device containing the electronic circuits and am-
perometric biosensors (called AutoSensor). The working electrode (WRK) is
made of screen printed layer of Pt/C composite, while the reference (REF) and
counter (CNT) electrodes are made of Ag and Ag/AgCl screen-print layers. The
AutoSensor contains two hydrogel discs, which serve both as the electrolyte and
reservoirs into which the glucose is collected [144]. Glucose oxidase is dissolved in
these hydrogel discs. The glucose sample are first extracted from the skin via re-
verse iontophoresis into the hydrogel discs, and then measured by amperometric
biosensor. The data are then processed prior to display.
The amount of glucose extracted during a 3-min iontophoresis cycle is esti-
mated to be 50–500 picomol, which corresponds to 2.5–25 μM of glucose con-
centration in each hydrogel pad [144].
Fig. 2.14. A simplified view of the GlucoWatch mechanism (Adapted from [145] and
GlucoWatch Patent US6233471 B1 [146])
The sequence of events in the operation of GlucoWatch are (as explained

in [144]):
1. Iontophoresis current of 0.3 mA is delivered for 3 min to collect glucose at
the cathode. The concentration of H2 O2 increases in the hydrogel during
this period.
2. Biosensors are biased at 0.42 V against the Ag/AgCl electrode. This low
potential of 0.42 V for the detection of H2 O2 improves selectivity towards
glucose.
2.5 Amperometric Sensor Calibration 39
3. The current at the cathode is integrated for 7 min during which the concen-
tration of H2 O2 in the gel is depleted.
4. The polarity of the iontophoresis current is then reversed, and the whole
process is repeated.
The integrated currents from the two biosensors are summed and input to a
signal processing algorithm, where the GlucoWatch Biographer filters the current
signals and then estimates BG from the signal. As described in [146], the signal
processing consists of:
• Baseline background signal processing
“Baseline background” is the current (nA) generated by the sensor indepen-
dent of the presence or absence of the analyte of interest. This baseline back-
ground can vary with time, temperature and other variable factors. To remove
or correct background information present in the sensor raw signal, a subtrac-
tion method can be used:
i(τ ) = iraw (τ ) − ibckgd (τ )
where iraw (τ ) = raw signals measured by sensor [nA] at time τ , ibckgd (τ )

= background current [nA], i(τ ) = corrected current [nA], and τ = time
after activation of the sensor. To compensate for temperature, a different
background current can be used:

−K1
ibckgd = A exp
T
where A = a constant, K1 = “Arrhenius slope” – a measure of how sensitive
the current is to changes in temperature, T = temperature in ◦ Kelvin. 11 The
temperature corrected background current is thus:

1 1
ibckgd,Correct = ibckgd,t=τ0 exp −K1 −
Tτ T τ0
where ibckgd,Correct = temperature corrected baseline current, ibckgd,t=τ0 =
baseline current at t = τ0 , Tτ = temperature at t = τ .
• Sensor calibration
The calibration can be carried out by giving a reference BG measurement
(BGcal , in [mg/dl]) to work out the conversion factor bgain in [mg/dl per nC]:
BGcal + ρ
bgain =
Ecal + δ
where Ecal = blank-subtracted smoothed sensor signal (nC) at calibration, ρ
= calibration offset [mg/dl], δ = offset calibration factor constant [nC] which
11 1
K1 can be found out by plotting the natural log of ibckgd versus . The slope of the
T
function is −K1 .
can be calculated using standard regression analysis. Blank subtraction is

another technique described in [146] where the signal from the inactive sensor
reservoir (i.e. blank, no glucose loading) is used to compensate for interference
from other analytes in the active sensor reservoir. The offset factor δ is used to
account for a non-zero signal at an estimated zero BG. This is needed because
sometimes a non-zero intercept is obtained in the correlation between signal
and reference glucose value.
To compensate for time-dependent decline in sensor signal, additive decay
parameters αi and multiplicative decay parameters i can be included:
BGcal + ρ − αi · tcal
bgain =
(1 + i · tcal ) Ecal + δ
where tcal is the time when calibration is performed, αi and i are time-
dependent signal decline and can have multiple time segments (i.e. i = 1, 2,
or 3). The BG at time t can be estimated using
BGest = bgain [(1 + i · t) Et + δ] − ρ + αi · t
• Mixture of Expert (MOE)
The MOE algorithm has been reported in [147] and proposed for GlucoW-
atch biographer. MOE is described as a generalized predictive data analysis
method that superimposes multiple linear regressions, along with a switch-
ing algorithm to predict outcomes. For GlucoWatch, the MOE inputs are
the elapsed time (time), integrated current (Iavg ), blood glucose values at
the calibration point (BGcal ) and a calibrated sensor signal (Ical ). The BG is
estimated using the equations

3
BG = wi · BGi
i=1
where wi are weighting factors and BGi are individual experts. Each expert
is described as a linear summation of appropriate input parameters Pj , plus
a constant (see [147]):

4
BGi = aij Pj + zi
j=1
= ai1 · time + ai2 · Iavg + ai3 · Ical

+ ai4 · BGcal + zi , i = 1, 2, 3
The weights wi are determined from
exp(φ1 )
wi = , i = 1, 2, 3
exp(φ1 ) + exp(φ2 ) + exp(φ3 )
where

4
φi = αij Pj + ζi
i=1
2.6 Clarke’s Error Grid Analysis (EGA) 41
˙ + αi2 · Iavg + αi3 · Ical

= αi1 time
+ αi4 · BGcal + ζi , i = 1, 2, 3
To determine the unknown coefficients (aij , zi and αij , ζi ), a technique called

Expectation Maximization is used as a two-step process: (1) The weights asso-
ciated with each data point are fixed, and the parameter values are optimised;
(2) the parameter values are fixed and the weights are optimised. These two
steps are iterated until convergence is reached [147].
For a good technical review of GlucoWatch Biographer, see [56].
2.5.4 Differences Between the MiniMed’s CGMStm and Cygnus

GlucoWatch Biographer
Both CGMS and GlucoWatch use the same principle of current generation from
the oxidation of hydrogen peroxide. The hydrogen peroxide is created by the
enzymatic action of glucose oxidase on glucose. However, the signal processing
in GlucoWatch requires a more comprehensive mathematical formulation (as
seen above) than the simpler Finite Impulse Response filter used in CGMS (as
reported in [148]). This is due to the differences in physiological pathways, and
the concentration of the glucose being measured by the two devices. Nevertheless,
both products require users to enter a glucometer reference value to perform a
single or multi-point calibration.
2.6 Clarke’s Error Grid Analysis (EGA)
Self-monitoring of blood glucose (SMBG) has become important in the man-

agement of diabetes mellitus. In order to assess the accuracy of a particular
glucometer, Clarke et al devised an analysis method that evaluates the accuracy
of a particular system over the entire range of BG value, and the clinical and
statistical significance of the system’s accuracy. It was designed to take into ac-
count the absolute value of the glucometer’s glucose measurement, the absolute
value of the reference blood glucose value, the relative difference between these
two values, and the clinical significance of this difference [149].
EGA defines the x -axis as the reference blood glucose and the y-axis as the
value generated by the monitoring system. The diagonal represents the perfect
agreement between the two. Data points situated above or below the diagonal
represents overestimates and underestimates, respectively.
EGA is based on the assumptions that (from Clarke et al):
1. the target blood glucose range, or the range of glucose values that patients
were taught to attain and maintain is 70–180 mg/dl (3.88–10 mmol/L);
2. patients will attempt to correct BG readings that are above or below the
target range, but not those readings within the target range;
3. corrective treatment by the patient is considered inappropriate if such treat-

ment causes the BG reading to fall outside of the target range;
4. failure to treat BG values <70 or >240 mg/dl is inappropriate.
Error Grid Analysis (EGA)

25
E C B A
A
20
SMBG determined BG (mmol/l)
15
10
D D
C E
0
0 5 10 15 20 25
Reference BG (mmol/l)
Fig. 2.15. Error Grid Analysis (Clarke et al)
EGA is divided into five zones of glucose estimations (Fig 2.15):

• Zone A – Glucose values deviate from the reference by no more than 20%,
or glucose values are in the hypoglycemic range (<70 mg/dl or 3.88 mmol/l)
when the reference is also <70 mg/dl. Values falling within zone A are clini-
cally accurate in that they would lead to clinically correct treatment decisions;
• Zone B – Values that deviate from the reference by >20% but would lead to
benign or no treatment (based on the assumptions described above);
• Zone C – Values that would result in overcorrecting acceptable BG levels, and
such treatment might cause the actual BG to fall outside of 70–180 mg/dl
(3.88–10 mmol/l) range;
• Zone D – Actual BG values are outside of the target range, but glucometer
values are within the target range. This would cause “Dangerous failure to
detect and treat” errors;
• Zone E – Reference BG values and glucometer BG values are opposite to each
other, and corresponding treatment decisions are therefore opposite to that
called for. This is the “erroneous treatment” zone.
In summary, values in Zone A and B are clinically acceptable, while values in
zones C, D and E are potentially dangerous and are clinically significant errors.
EGA has since been used quite extensively to compare the accuracy of glu-
cometers (commonly against a reference meter such as Yellow Springs Instrument
(YSI) whole blood glucose analyser).
2.7 Insulin Infusion 43
2.7 Insulin Infusion
2.7.1 Fate of Insulin
Physiological insulin delivery from the endocrine pancreas is a continuously regu-

lated process, which responds to changes in substrate concentration in a matter
of seconds [150]. In a normal individual, endogenously produced insulin is se-
creted directly into the portal venous circulation, and undergoes partial (40% to
80%) extraction by the liver. After this clearance, it is diluted into the systemic
insulin pool [151,152], and is distributed within the plasma as “free” (unbound)
insulin. This biologically active free insulin is then distributed by diffusion into
the extravascular compartment to reach insulin receptors on the surface of the
target cells. The interaction of insulin with its receptor enables fuel (mostly car-
bohydrate and amino acid) to enter cells and activate enzymes for the storage
or metabolism of those fuels. Insulin that does not combine with receptor cells
is degraded mainly in the liver. Insulin is also cleared from the plasma by the
kidney.
2.7.2 Insulin Type Factor
In contrast to endogenous insulin, some exogenously infused insulins require fur-

ther intermediate steps before being active at the target cells. Because insulin is
more readily absorbed as dimers and monomers, insulin preparations that con-
tain the larger hexamers (such as regular injectable insulin) must first dissociate
into dimers and monomers to be absorbed [28].
Conversely, insulin preparations that contain monomers and dimers are ab-
sorbed more rapidly [28]. Shorter-acting insulin neutral and rapid-acting insulin
Lispro are some examples of readily absorbable insulins. Figure 2.16 shows time
of activity of some commonly used insulin.
Exogenously infused insulin is degraded at the usual sites: liver, muscle, and
kidney. However, the kidney plays a larger role in clearing exogenously adminis-
tered insulin (than endogenous insulin), and it clears approximately 30–80% of
exogenous insulin from the systemic circulation.
2.7.3 Routes of Insulin Administration
As insulin is normally cleared rapidly from the plasma, circulating insulin levels
and its ultimate delivery to the target tissues depends mainly on the entry of
insulin to the circulation.
In a healthy individual, insulin enters the circulation mainly from pancreatic
release. But if there is an impairment of endogenous production (i.e. diabetes),
exogenous insulin must be administered. Exogenous insulin can be administered
by different routes:
• Intraperitoneal (i.p.).
• Subcutaneous (s.c.).
Time of Activity of Human Insulins*

(in hours) 0 2 4 6 8 10 12 14 16 18 20 22 24 26 28 30
Rapid-Acting
(Clear and colourless)
Insulin Lispro
Humalog®
Short-Acting
(Clear and colourless)
Neutral (soluble)
Humulin® R
Intermediate-Acting
(Uniformly cloudy or milky)
Isophane
Humulin® N
Intermediate-Acting
Lente
Humulin® L
Long-Acting
UltraLente
Humulin® U
Premized
Humalog® Mix 75/25

Premized
Premixed Insulin
Humulin® 70/30
Premized
Humulin® 50/50
Description *Typical profiles of insulin activity, based on glucose

(Appearance) utilisation from clinical trial data.
Generic name
Brand name Adapted from:
LillyDiabetes.com (Eli Lilly and Company, Indiana, USA)
Fig. 2.16. Human insulin action activity chart. (Source: Eli Lilly & Co., Indiana,
USA).
2.7 Insulin Infusion 45
• Intravenous (i.v.).
• Intramuscular (i.m.).
• Alternative non-invasive routes, such as oral, nasal, pulmonary or transder-
mal/transcutaneous.
The intraperitoneal, subcutaneous and intravenous routes are the three most
common routes for continuous administration of insulin.
Intra-peritoneal Delivery
In a normal individual, insulin is first delivered to the hepatic circulation by the

pancreas. Hence, it follows that delivery of insulin to the liver should produce
more normal metabolism of glucose and other metabolites, than peripherally
delivered insulin.
The peritoneum is a possible site for insulin administration [150] because the
visceral peritoneal membrane has the same blood supply as the small and large
bowel, and is drained via the portal vein to the liver. Therefore, the delivery
of insulin intraperitoneally should, theoretically, have the advantage of insulin
being absorbed preferentially into the portal system, producing a therapeutic
state more closely resembling the normal physiology.
Numerous studies on insulin-dependent diabetes mellitus patients have demo-
nstrated that insulin delivered intraperitoneally is rapidly and predictably
absorbed, with most of it going into the portal system [153]. Intraperitoneal
delivery:
• provides sufficient insulin to the liver to control hepatic metabolism;
• reduces peripheral insulin levels, and the risks of arteriosclerosis, and other
adverse effects of hyperinsulinemia (e.g. increased risk of developing Alzheimer
disease [154, 155]);
• results in carbohydrate and particularly lipid metabolism that more closely
mimics the normal physiological state, than those produced by s.c. injections
of insulin.
Nevertheless, intraperitoneal insulin delivery requires an insulin pump to be
surgically implanted in the peritoneum. Any mechanical failure to the pump
poses risks to the patient by the need to go through surgery for pump removal.
Subcutaneous Delivery
S.c. injection is the most preferred route by patients to receive their daily insulin
doses, either by direct injection at the site, or via insulin pump therapy.
When insulin is injected subcutaneously, it forms a depot at the injection site.
The compound then transforms to an absorbable state through dissociation and
dissolution [156], and diffuses from the depot into the circulation. The diffusion
process is slow, taking 50–90 min in general [28]. Similar to endogenous insulin,
the injected insulin is distributed within the plasma as “free” (unbound) insulin,
diffusing further into the extravascular compartment to reach the target cells.
Some shortcomings of s.c. delivery are:

1. The s.c. absorption rate is highly unpredictable, mainly attributed to changes
in blood flow around the injection site [151]. In most insulin-treated subjects,
part of the circulating insulin is reversibly bound to insulin antibodies, which
changes the free insulin levels unpredictably;
2. Variation exists in insulin absorption due to factors such as depth and loca-
tion of injection, exercise, and vascularity of the tissue. These are common
causes of unpredictable wide swings in blood glucose levels;
3. Although the insulin peaks can be adjusted to match food absorption by
giving subcutaneous insulin in advance of meals, the difficulty of estimating
when exactly to eat increases when insulin absorption could vary [153];
4. As s.c. insulin forms a depot at the injection site, once injected, any excess
insulin cannot be withdrawn from the injection site. This may lead to hypo-
glycaemia if the patient does not ingest carbohydrate after the injection;
5. An inflammatory reaction might develop at the injection sites, necessitating
frequent rotation of injection sites;
6. There is a long lag time between the injection of insulin and its appearances
in the plasma (30 min to 1 hour).
Intravenous Delivery
Unlike subcutaneously injected insulin, intravenously administered insulin, es-

pecially as the soluble form, enters the bloodstream immediately, and is rapidly
absorbed, thereby acting faster. Since i.v. insulin is usually injected into the
veins at a peripheral site, and not directly into the hepatic circulation, the in-
sulin avoids the first pass through the liver, and this effectively leads to an
increase of 50% of insulin concentration, compared to i.p. or endogenous insulin
availability.
Peripheral hyperinsulinemia may be a consequence of this route of infu-
sion [16], and is not desirable, as emphasized by the mounting evidence that
circulating insulin levels may be directly related to the risk for atherosclero-
sis [153].
The primary shortcoming of i.v. insulin administration is the requirement for
venous access, which usually requires medical supervision.
Continuous Versus Intermittent Intravenous Delivery
Intravenous insulin may be given as either repeated boluses or continuous in-

fusion. The bolus ensures insulin absorption, and permits rapid adjustment of
insulin to suit the patients needs. However, the effect of each bolus of short-acting
insulin dissipates within 30 minutes to 1 hour, necessitating frequent BG mea-
surement and the delivery of insulin bolus. Also, it is possible for the patient to
form anti-insulin antibodies, and thus be immunized against the intermittently
delivered insulin.
2.8 Conclusion 47
In contrast, continuous intravenous insulin delivery allows a longer period

between BG measurement, and permits fine-tuning of BG levels to within a
chosen range, although it may increase the risk of hypoglycaemia.
Other Delivery Routes
The latest in insulin delivery technique has focused on non-invasive delivery:

• Oral – Capsules containing insulin for oral administration (or so-called
Emispheretm delivery agent combination for example, marketed by Emisphere
Technologies Inc. (Tarrytown, N.Y.)).
• Nasal – The inhaled insulin delivery system provides insulin as a dry powder
inhaled through the mouth directly into the lungs from where it passes into
the bloodstream.
• Transdermal - Delivery of insulin through a transdermal “patch”.
One product is called “U-Strip” by the founding company Encapsulation Sys-
tems Inc (http://www.encsys.com). The device uses ultrasound to enhance
transport of insulin through the skin pores (See US Patent 6,908,448). It is
wearable and portable, and it features programmable drug delivery through
its drug sonic applicator device. Another rivaling company is Sontra Medical
Corporation (http://www.sontra.com) (See [58, 60]).
The advantages with insulin delivered through transdermal route is that the
insulin suffers no degradation (in contrast to orally administered insulin that
can be destroyed by the GI tract) and no painful needle stick is required.
Other Factors/Mode of Delivery
Apart from the type and infusion route of insulin, insulin action is also influenced
by factors such as:
1. insulin resistance
2. insulin sensitivity — the sensitivity of the insulin receptors
3. the action of the counter-regulatory hormones,
4. physical exercise and diet
There has also been a report that oscillatory delivery of exogenous insulin with
a period of 120 minute into the systemic circulation leads to a greater reduction
in the plasma glucose concentration as compared to constant delivery [157].
This phenomenon was discovered in normal weight, non-diabetic subjects. Its
implication in diabetic subjects is yet to be determined.
2.8 Conclusion
Various BG measurement methods exist, each with their advantages and disad-
vantages. Although the choice of a method depends on the target population
and the clinical setting, it is undoubtedly that non-invasive method is the most
preferable choice, and the way forward for the future. But until non-invasive
glucose sensing techniques have reached maturity, minimally-invasive methods
would remain the choice for longer-term continuous glucose monitoring. Invasive
glucose monitoring remains in use only in clinical research settings (and not for
day-to-day use by the general public).
Similarly, for the insulin infusion, the choice of insulin type and administration
routes depends on the target population. The subcutaneous route is the preferred
route for administration in ambulatory patients, despite possible discomfort,
while invasive route is used mainly in clinical routine (due to its faster action).
There are progress being made in non-invasive insulin delivery, and commercial
products have started to appear. With the emergence of non-invasive insulin
delivery techniques, the future of pain-free delivery appears hopeful.
2.9 Summary
This chapter discussed the three categories of BG measurement techniques, i.e.

invasive, minimally-invasive and non-invasive techniques. The methods in each
category were described, as well as their advantages and disadvantages. The
chapter also paid particular attention to one glucose sensing method – amper-
ometry – from how the signal is obtained from the sensor, to the steps involved
in converting the signal to BG level estimations. Amperometric glucose sensor
has reached successful commercialisation ahead of other measuring techniques,
and this progress would certainly help in advancing glucose control by provid-
ing the necessary tools for further advancement in the field. This chapter also
presented the types of insulin that were commonly used, and the effects of the
routes of delivery on the insulin activity. Non-invasive delivery of insulin remains
the ultimate goal by many to achieve totally pain-free insulin delivery.
3
Glucose Control: Patient Dynamics
3.1 Introduction
The role of the control algorithm in a closed-loop insulin delivery system is to reg-
ulate the patient’s BG level, replacing the intrinsic glucose regulatory function,
which is abnormal in diabetics. To develop an effective algorithm, a knowledge
of how glucose is intrinsically regulated in a healthy person is essential.
The human pancreas has between 1 and 2 million islets of Langerhans. These
islets contain three major cell types: alpha, beta and delta. The beta cells con-
stitute about 60% of the cells, and secrete insulin. The alpha cells, about 25%
of the total, secrete glucagon. The delta cells, about 10% of the total, secrete
somatostatin. The remaining 5% of cells are made up of other cell types which
secrete hormones of uncertain function [4]. Insulin and glucagon play the most
important roles in the glucose-regulatory system.
The glucose-regulatory mechanism is not an isolated system, but has connec-
tions with many other metabolic pathways in the body.
This chapter reviews the various mechanisms involved in the glucose-regulation
in normal individuals, diabetic patients and critically ill patients.
3.2 Intrinsic Blood Glucose Regulation

In a normal healthy person, blood glucose concentration is controlled to within
a narrow band, usually between 3.5 mmol/l and 5.6 mmol/l in the fasting state.
During fasting, insulin secretion is reduced to a basal level, and glucagon is
released to allow the liver to:
• mobilize glucose from its glycogen stores (glycogenolysis) and
• synthesize glucose from amino acid (gluconeogenesis).
In addition, when insulin levels are low, the uptake of glucose by muscle is
minimized, and there is lipolysis of stored fats to release free fatty acids. When
fasting persists longer than 12 to 18 hours, these free fatty acids then become
the main energy substrate, used by essentially all tissues of the body, except
F. Chee & T. Fernando: Closed-Loop Control of Blood Glucose, LNCIS 368, pp. 4 9–57, 2007.
springerlink.com
50 3 Glucose Control: Patient Dynamics
the brain. Gluconeogenesis still supplies glucose for obligatory glycolytic tissues,
notably the brain.
This mechanism effects a stable fasting blood glucose concentration so that
the brain, which has no energy stores, has a sufficient supply of nutrients for nor-
mal activity. Glucose is an essential nutrient for the brain, retina, and germinal
epithelium of the gonads. Insulin is always present, and a low level of circulating
insulin regulates the rate of lipolysis, glucose transport and gluconeogenesis at
all times [158].
When a person prepares to eat a meal, two phases of insulin secretion occur:
an anticipatory phase (first phase) and a glucose-sensitive phase (second phase).
In the anticipatory phase, the sight of food and the first bite of a meal cause
the brain to send signals to the pancreas. These signals cause the pancreas to
release insulin into the hepatic circulation (Fig 3.1). Once the insulin is in the
hepatic circulation, the liver stops breaking down glycogen into glucose.
Signals from brain Inferior vena cava
Vagus nerve trunk
Portal vein
Esophagus
Pancreas
Liver
Insulin released
Stomach directly into
Intestine portal vein.
Nutrients from the

intestine is absorbed
into the hepatic circulation.
Fig. 3.1. Anatomy of the intrinsic blood glucose regulation mechanism
As the food enters the stomach, the release of insulin is further facilitated by
gastrointestinal hormones. These hormones increase the sensitivity of the islet
cells to glucose [159]. As nutrients are absorbed into the circulation, the glucose-
sensitive phase begins, and there is continuous secretion of insulin. These two
phases are sometimes termed the biphasic response of insulin secretion.
After absorption of all the carbohydrates, the feedback system for control of
blood glucose returns the glucose concentration rapidly back to the control level,
usually within 2 hours.
Although glucose is an important physiological stimulant of insulin secre-
tion, nutrients other than glucose, particularly amino acids, are also capable of
3.3 Diabetic Patients 51
stimulating insulin release. Nevertheless, amino acids stimulate a much greater

insulin response when accompanied by hyperglycaemia, than the modest degree
expressed in the presence of normal plasma glucose level.
Biphasic response from pancreas

Insulin level (arbitrary unit)
−20 0 20 40 60 80 100 120 140 160

Time (min)
Fig. 3.2. Biphasic insulin response (illustration)
The biphasic response (Fig 3.2) is also observed in the plasma compartment
during glucose clamp studies, where a square wave of hyperglycaemia is intro-
duced intravenously under normal conditions in a normal individual. The first
phase of the insulin response consists of a rapid rise in the insulin level during
the first 10 min, with a peak response at 4 min; the plasma insulin level then
falls and reaches a nadir at 10 min [160]. After 10 min, the plasma insulin level
is then observed to increase gradually according to the degree of hyperglycaemia
and persists for the duration of the stimulus [161].
3.3 Diabetic Patients

In patients with Type I and Type II diabetes mellitus, beta cells have either been
partially (as in Type II) or completely (as in Type I) destroyed. The destruction
of beta cells results in reduced or no insulin secretion, and also the disappearance
of the first phase (anticipatory phase).
With no anticipatory phase, the liver does not receive the message to stop
breaking down glycogen into glucose, resulting in continued hepatic glucose pro-
duction. Added the glucose absorbed from the meal, and a lack of insulin release,
hyperglycaemia ensued [162]. The significance of the first phase release has been
demonstrated in [163] where a 15-minute delay in initiation of insulin delivery,
relative to the start of the meal resulted in a greater meal-induced hypergly-
caemia in patients.
The major difference between Type I and Type II diabetes mellitus is that
Type I diabetic patients cannot survive without exogenous insulin, contrary to
Type II patients who suffer from either a reduced (but present) insulin secre-
tion, or abnormal insulin response (increased peripheral insulin resistance) or
both. Hence, Type I diabetes is also known as insulin-dependent diabetes melli-
tus (IDDM), where as Type II is also known as non-insulin-dependent diabetes
mellitus.
The first phase insulin release is somewhat harder to achieve by a closed-loop
insulin delivery system in an automatic manner. This is because the anticipatory
phase in a normal individual is initiated by neuro- and gastro-intestinal signals,
which are not measurable by a glucose sensor. The word “automatic” is used,
because in open-loop delivery systems, this signal can be given by the patient
by injecting insulin boluses manually before meal.
A solution to “artificially” achieve the first phase insulin release using a closed-
loop system would be to use sensors that have a very fast response time (e.g.
1 min at least), and to immediately deliver insulin when BG level starts rising
(with the assumption that the rise of BG level signifies the start of a meal (as
inferred from [163])). In fact this method has been used in the early external
artificial endocrine pancreas clinical experiments, such as those reported in [17]
and [16].
As for the second phase insulin release, the matching of the insulin dose to
the blood sugar intake depends on:
• A knowledge of how much glucose is ingested. In a healthy pancreas, beta
cells also function as “fuel-sensors” and are capable of adapting the rate of
insulin secretion to variations in plasma glucose level.
• The responsiveness of the insulin receptors on target cells to enable glucose
to enter and be utilised by the cells.
3.4 Importance of BG Control

It is important for BG level to be kept within the normal range, because:
1. a high glucose concentration exerts an osmotic pressure in the extracellular
fluid, and causes cellular dehydration. This excessive BG level causes loss
of glucose through urination (glycosuria), leading to osmotic diuresis that
depletes the body further of fluids and electrolytes.
2. too low a BG level carries the risk of hypoglycaemic coma. The BG level
should not drop below a certain level because glucose is the only nutrient
that can be used for energy by the brain, retina, and germinal epithelium of
the gonads.
3. too high a glucose concentration (>11.1 mmol/l) can affect wound healing
and interfere with human neutrophil function [8].
4. therapy that maintains BG level at below 11.9 mmol/l improves the long-
term outcome in diabetic patients with acute myocardial infarction [9].
3.5 Glucose Control in Critically Ill Patients 53
5. pre-prandial blood glucose concentrations between 3.9 mmol/l and 6.7 mmol/l,
and postprandial concentration of less than 10 mmol/l was found to delay
the onset of diabetic microvascular complications [11]. These microvascular
complications included retinopathy (visual impairment), vision-threatening
lesions, nephropathy (kidney disease) and neuropathy (nerve damage).
Complications not only occurred in diabetic patients, but also has an impact on
critically ill patient population:
1. A recent finding (see [9]) has shown that the use of intensive insulin therapy
to maintain BG at a level that did not exceed 110 mg/dl (6.1 mmol/l)
substantially reduced mortality and morbidity in critically-ill patients in
the ICU (8.0% with conventional treatment to 4.6% with intensive insulin
therapy).
2. The finding also reported that a pronounced hyperglycaemia in critically-
ill patients, even those who have not previously had diabetes, may lead to
complications in such patients [9].
3. Patients with mean glucose concentrations >11.1 mmol/l within 36 h follow-
ing surgery were more likely to develop infectious complications than their
counterparts who were under better glycaemic control [14].
All these findings demonstrated that apart from the administration of exogenous
insulin being essential to control the blood glucose concentration and to maintain
homeostasis in patients, the tightness of such a control is crucial in avoiding
long-term complications (in diabetic patients) and infectious complications (in
critically ill patients).
3.4.1 New Findings
According to [164], it is believed that sensory nerves of pancreas release a neu-

ropeptide called “substance P” which is responsible for ensuring that islet cells
produce the right amount of insulin. On lacking of substance P, islets cells (in
mice) overproduce insulin, leading to insulin resistance and ultimately islet-cell
death. By injecting substance P to diabetic mice, it was found that the diabetes
in mice disappeared and the mice remained diabetes free for months.
3.5 Glucose Control in Critically Ill Patients

Critically ill patients are different from non-critically ill patients in many re-
spects. Major surgery and critical illness are physiologically stressful events that
provoke complex metabolic responses in the patient. In general, the greater the
degree of surgical trauma, the greater the endocrine upset.
Critical care medicine uses the term “stress” to describe the systemic re-
sponse to severe injury or infection [7]. Some of the responses to stress include
alterations in carbohydrate metabolism, and a state of hypermetabolism. The
hypermetabolic state is induced by sepsis or injury, as well as by organs involved
in the immunologic response to stress [7]. It is manifested as an increase in glu-

cose production that appears to be directed toward maintaining glucose delivery
to wound and immune tissues [7]. These alterations make the control of plasma
glucose more difficult than in ordinary diabetic patients.
Factors that affect glucose regulation in critically ill patients include:

• Loss of the inhibitory effects of elevated glucose levels.
• Stress-induced release of counter-regulatory hormones
• Increased Insulin resistance
• Medication that induces hyperglycaemia
• Effects of Feeding
• Gastro-intestinal (GI) tract functionality
• Stress in liver
3.5.1 Loss of the Inhibitory Effects of Elevated Glucose Levels

In the non-stressed state, elevated serum glucose levels exert an inhibitory effect
on gluconeogenesis. In critically ill patients, this important feedback mechanism
is blunted, often resulting in continued endogenous hepatic glucose production
and hyperglycaemia. Although there may be an increase in glucose production
and turnover, the reliance on glucose as an energy source is reduced.
3.5.2 Stress-Induced Release of Counter-Regulatory Hormones

In response to stress, the counter-regulatory or anti-insulin hormones are se-
creted. These hormones include epinephrine, norepinephrine, cortisol, growth
hormone, and glucagon.
• Both norepinephrine and epinephrine cause constriction of essentially all blood
vessels of the body (vasopressor activity), and an increased activity of the
heart (inotropic activity). The constriction of the vessels increases the total
peripheral resistance and elevates arterial pressure. Epinephrine (also known
as adrenaline) also mobilises glucose from glycogen and raises blood glucose
level.
• Cortisol (a gluco-corticoid) stimulates hepatic glycogen and glucose produc-
tion, and inhibits insulin action on peripheral tissues. It stimulates protein
synthesis, providing substrate for gluconeogenesis. Gluco-corticoids also sup-
press the inflammatory and immune response [4].
• Growth hormone increases the rate of protein synthesis in all cells of the body,
and increases mobilisation of fatty acid from adipose tissues into the blood.
There is also increased use of the fatty acid for energy, with a decreased rate
of glucose utilization throughout the body [4].
In all, the net effect of these hormones is the raising of blood glucose concentra-
tion, mobilisation of alternative fuels, reduction in glucose tolerance and increase
in peripheral resistance to the effects of insulin.
3.5 Glucose Control in Critically Ill Patients 55
3.5.3 Increased Insulin Resistance
Patients with Type II diabetes (also know as NIDDM) are usually insulin resis-
tant. Surgical stress potentiates this insulin resistance, mainly due to the release
of the counter-regulatory hormones. Enhanced gluconeogenesis during stress is
also resistant to inhibition by insulin and glucose. Although skeletal muscle has
traditionally been implicated as the major site of peripheral insulin resistance,
stress may also induce insulin resistance in adipose tissue, liver, and heart [7].
3.5.4 Medication that Induces Hyperglycaemia
Concurrent administration of exogenous vasopressors and gluco-corticoids can

further amplify the effects of stress-induced counter-regulatory hormones in ICU
patients as described above. Drugs with vasopressor and/or inotropic activity
(such as epinephrine and norepinephrine) are useful for resuscitation of critically
ill patients, and are administered when fluid resuscitation alone fails to reverse
hypotension.
3.5.5 Effects of Feeding
Patient in Intensive Care are usually sedated. During this period, nutrition is
delivered either enterally or parentally.
• Enteral nutrition delivery is always the preferred route whenever the GI tract
is functional. Nutrition is delivered proximal or distal to the pylorus through
a naso-enteric tube. Since gastric emptying is often impaired in critically ill
patients, feeding distal to the pylorus maybe preferable [165].
• Total parenteral nutrition (TPN) is given in situations where the patient has
contra-indications to enteral feeding (such as bowel obstruction, overwhelm-
ing intra-abdominal sepsis, or nectronizing pancreatitis), or when the patient
is unable to absorb nutrients via the GI tract. TPN in the ICU setting is
delivered through a central venous catheter, since most critically ill patients
already have central venous monitoring in place.
The plasma glucose concentration represents a balance between the influx of
glucose into the circulation and the rate at which it is cleared from the circulation
(either by the action of insulin, peripheral uptake or due to renal excretion).
Taken together with the varying rates of intravenous dextrose-containing fluids
ordered by the medical team, the ultimate glucose level in the patient depends
on how well this balance is being maintained.
3.5.6 GI Tract Functionality
In terms of meal intake, perhaps the difference between an outpatient and a

critically ill patient is in “how” and “when” they receive their daily nutrition.
The direct intubation to the pylorus implies that food ingestion does not
go through any peristaltic action of the esophagus, nor the normal stomach
stimulation by the presence of food, as would occur in a normal individual.

There is the effect of no chewing, no visual stimulation by food, and no activity
of the stomach which pre-empts insulin release in sedated patients.
Food absorption in critically ill patients remains dependent on gut motility
and other functions of the GI tract. The meal pattern is also abnormal in that
nutrient is given continuously over hours rather than as discrete meals.
3.5.7 Stress in Liver
The liver functions as an important blood glucose buffer system. During the
blood glucose rise after a meal, the liver stores as much as two thirds of the
glucose absorbed from the gut in the form of glycogen [4]. The stored glycogen
is later released back into the circulation as glucose when required.
This action of the liver decreases the fluctuations in blood glucose level. How-
ever, in patients with severe liver disease, it becomes difficult to maintain a
narrow range of blood glucose level [4].
3.6 BG Management in the ICU

Critically ill patients are treated by the method of continuous administration
of intravenous insulin by infusion pump, although intermittent subcutaneous
injections may also be given in certain cases. Continuous intravenous infusion
obviates concerns over the patient’s state of perfusion (which influences subcu-
taneously injected insulin), and permits fine-tuning of BG level within a chosen
range:
• 8.3–13.8 mmol/l (150–250 mg/dl) [158]
• 6.7–10 mmol/l (120–180 mg/dl) [8, 166]
• 6–10 mmol/l (108–180 mg/dl) [167].
In the current clinical setting, BG level less than 6 mmol/l (108 mg/dl) carries
the potential risk of hypoglycaemia, while BG level in excess of 13.8 mmol/l
(250 mg/dl) would also require intervention1 .
Continuous infusion of insulin may increase the risk of hypoglycaemia, but
the continuous monitoring of patients in an ICU minimises this risk and enables
BG to be controlled safely.
3.6.1 BG Measurement
Current clinical practices examine BG two- to four-hourly, with hourly checks

during critical cases. BG determination is by conventional glucometer with the
blood sample obtained from arterial or venous cannula.
1
These “higher” ranges would now have to be re-considered in light of the new finding
in [9] that emphasized the importance of tighter BG level in critically-ill patients.
3.8 Summary 57
3.6.2 Insulin Infusion Adjustment
Insulin infusions are administered in saline or colloid solution, and mixed such
that 1ml/hr of delivery equates to 1U/hr of insulin. A rate of 0.5 to 1 unit
of insulin per hour is the recommended starting dose. This dose is usually for
patients who are not severely stressed. Higher rates may be needed for adequate
BG control.
Adjustments to insulin rate are made hourly, two-hourly or four hourly based
on arterial blood (from arterial cannula) or fingerstick glucose determinations.
The adjustment frequency depends on the stability (or severity) of the BG eleva-
tion. Patients may vary greatly in their sensitivity to insulin and some, especially
long-term type 1 diabetic patients are quite sensitive to small changes in insulin
dose.
Various rules for selecting the appropriate insulin infusion rate exist, such as
the sliding scale table, titration, and variable-rate intravenous insulin infusions
(see Section 4.2 in Chapter 4). These methods were designed and mainly used for
critically ill patients, whose physiological situation is different from ambulatory
patients. These method also used intermittent BG determination, rather than
continuous measurements.
3.7 Conclusion
Biological systems involve complex and interdependent processes. High blood
glucose levels in patients are often a result of metabolic derangement at various
levels. Glucose control in critically ill patients differs from diabetic ambulatory
patients in that they face additional metabolic stress from critical illness or
surgery. This renders their blood glucose level difficult to predict and sometimes
control.
3.8 Summary
In this chapter, the glucose-regulatory mechanisms in a normal and diabetic
individuals were compared and contrasted. The challenges of controlling the BG
levels in the critically ill population were also discussed, followed by an overview
of the clinical practices currently used in managing BG in the Intensive Care
Unit.
4
Mathematics of Glucose Control
4.1 Introduction
We have looked at blood glucose measurement, insulin infusion (Chapter 2), and
the characteristics of the patient in terms of the blood glucose control (Chap-
ter 3). In this chapter, we look at the control algorithms (or the “smarts”) that,
when worked together with the glucose sensor and insulin infusion pump, would
ideally re-balance a patient’s blood glucose level.
Recall that the beta cells (which being the fuel sensor as well as the insulin
production source) were destroyed in a diabetic patient, causing an impairment
of the ability to self-regulate glucose level. Glucose level regulation must then be
restored by means of carefully calculated external insulin infusion. The goal of
a closed-loop control system is thus to mimic the functionality of the pancreas
in providing automatic regulation of blood glucose level in patients.
To be precise, the closed-loop control systems (with its algorithm) should
really answer the question: “How much insulin should be given such that the
person blood glucose is restored, as closely as possible, to that of a healthy
individual?”
Numerous researches were conducted to address this question, and many so-
lutions were proposed. This chapter aims to give a brief introduction to:
1. the control algorithms proposed in the literature, categorised based on the
two different approaches to the design of closed-loop control algorithm,
namely,
• model-less approach,
• model-based approach, where the model is linear and/or non-linear.
2. the relationships between the control algorithms and the formulation of
glucose-insulin model for the purpose of blood glucose level control in di-
abetic patients.
This chapter only touches upon some basics of glucose-insulin modelling. Readers
interested in detailed mathematical modelling techniques are invited to consult
other dedicated textbooks (e.g. [168] and [169]).
F. Chee & T. Fernando: Closed-Loop Control of Blood Glucose, LNCIS 368, pp. 59–10 8, 2007.
springerlink.com
60 4 Mathematics of Glucose Control
4.2 Model-Less (Empirical) Control Algorithms

In the model-less (empirical) approach to control algorithm design, the rela-
tionship between the input (insulin) and output (desired glucose level) are de-
termined based on experimental data, not on a theory. A control rule is then
formulated using the experimental data as the basis.
In practice, it is not often easy (or feasible) to completely understand how the
body works. Hence, the simplest way to arrive at a control algorithm would be to ob-
serve the cause-and-effect (or dose-and-response) in the real system, and formulate
a control rule based on the observed relationship between the variables of interest.
4.2.1 Control Algorithm Based on Curve-Fitting

In this method, the relationship between the inputs and outputs are obtained by
fitting simple curve equations to the experimental data. As an example, experi-
ments could be conducted to observe the glucose level measured from a patient
when different amounts of insulin injections were given over a period of time. A
curve would then be fitted to arrive at a simple glucose-insulin response curve
(see Table 4.1), which would then be used as the control rule.
The earliest literature featuring the use of glucose-insulin response curve for
glucose control was authored by Albisser et al [10, 17, 23, 172].
The control equation formulated by Albisser et al consists of a sigmoidal dose-
response curve, with an incorporated predictive equation.
The predictive portion of Albisser et al’s algorithm takes into account:
• the trend of BG (i.e. rise, fall)
• the delay between blood extraction and the ultimate measurement
The control equation is likened to a Proportional-Derivative controller. The
derivative action is based upon the rise or fall of BG, while the dose-response
curve provides the Proportional action.
In those time, the patient’s BG was measured invasively on a minute-by-
minute basis (see Section 2.3), and insulin was infused intravenously. However,
a delay of 5–10 min were required before a BG reading was available due to
the then-available BG analysis method. It was reported that any delay in the
BG measurement could throw the controller action out of synchronisation [22].
Compensation was required by predicting BG based on the rates of change over
the previous minutes.
Response to a meal was usually normalised by the combined action of a rapid
insulin delivery on first detection of BG rise (i.e. the derivative action of the
controller), and the action of a second phase secretion performed by the algo-
rithm. The timeliness of the first phase insulin release was compensated for,
through minute-by-minute BG measurement, and the fact that insulin entered
the circulation much more rapidly when administered intravenously (compared
to subcutaneous infusion) [156].
Unlike other methods that relied on variables that correlated with the whole
blood glucose, invasive BG measurement has the advantage of being more accu-
rate, because it measured glucose content in the whole blood itself.
4.2 Model-Less (Empirical) Control Algorithms 61
Table 4.1. Glucose-insulin response curve (Sigmoidal dose-response curve)
Glucose-Insulin Response Curve (Sigmoidal Dose-Response

Curve)
The glucose-insulin response curve came from the discovery that insulin
secretion did not respond as a linear function of glucose concentration. The
relationship between the extracellular glucose concentration and the rate
of insulin secretion in vitro is sigmoidal, with a threshold corresponding to
the glucose level normally seen under fasting conditions, and with the steep
portion of the dose-response curve corresponding to the range of glucose
levels normally achieved postprandially.
Dose-response curve
2.5 11
10
2
9
Glucose Utilisation (mg/kg/min)

Glucose production (mg/kg/min)
8
1.5
7
6
1
5
4
0.5
0 2
10 100 1000
Plasma Insulin (uU/ml)
This sigmoidal nature of the dose-response curve could be attributed to a

Gaussian distribution of threshold for stimulation in beta-cells. This also
means that linear model is not adequate to describe the glucose-insulin
interaction in human (Text adapted from [159, 170]; Graph adapted from
[171]).
Albisser’s algorithm formed the basis for the algorithm used in the Biostator
bedside glucose control system (Table 4.3) used in further clinical investigations
of glucose-insulin interaction in humans (e.g. see [17], [16], [173] and [174]).
4.2.2 Control Algorithm Based on Lookup Table
In this method, the relationship between the inputs and outputs is obtained by
mapping the inputs and outputs in the form of a lookup table. In the day-to-day
care of patients in the hospital settings, blood glucose control is commonly achieved
using model-less approach in an open-loop manner. BG samples are taken intermit-
tently (at defined internals), and insulin delivery rate is adjusted manually using:
• lookup table control, such as a sliding scale table. The table has a continuous
BG “partitioned” into ranges, with an insulin rate assigned to each range.
Table 4.2. Albisser et al’s infusion algorithm
Albisser et al’s Infusion Algorithm
• Glucose infusion rate, Rd = 12 Md [1 − tanh Sd (G − Bd )]

• Insulin infusion rate, Ri = 12 Mi [1 − tanh Si (Gp − Bi )]
• Projected BG, Gp = G + K1 exp A
K2
−1
where M = maximum infusion rate; S = slope; B = BG level at which

half maximum infusion rate is chosen to occur; subscript d,i = for glu-
cose/dextrose and insulin respectively; A = rate of change of BG, which is
minute-to-minute changes of BG averaged over the preceding four minutes;
K1 is chosen to adjust the magnitude of the difference factor, and K2 se-
lected to establish its sensitivity to changes in A.
Insulin is given in accordance with the range in which the BG sample resides
(Table 4.4). The insulin delivery can either be intravenous or subcutaneous,
and different tables are prescribed for different routes of delivery. Tables that
used 1-hourly [175], and 3-hourly BG sampling [18] for subsequent adjust-
ment of intravenous insulin delivery rate have been reported to achieve good
normalisation of high BG.
• linearised lookup table, where the “step-wise” insulin increase is replaced by
a “slope”. Furler et al [176] used such an algorithm, where insulin rates are
0.5 U/hr for BG<4 mmol/l and 2.5 U hr for BG > 8 mmol/l, with a linear
transition between these rates over the range of BG of 4–8 mmol/l. Oller-
ton [177] subsequently improved on Furler et al’s algorithm by using grid
search algorithm. These slope-based algorithms were mainly used in clinical
investigation (e.g. [176]).
Nowadays, experienced nursing staff would generally be able to control an in-
patient’s elevated glucose level more efficiently than a prescribed sliding table.
This is because a prescribed sliding table can lose it effectiveness when the
patients glucose tolerance changes with their state of well-being.
4.2.3 Control Algorithm Based on Rule-Based Control
Model-less approach to blood glucose control could also use expert rules (rules
that are based on experience) as the control rules. They can be seen in clinical
techniques, which include:
• titration control, where the intravenous (IV) insulin infusion rate is started
at an empirical level, and progressively tuned to an appropriate rate which
lowered and maintained the BG in the target range;
• variable-rate IV insulin infusion algorithm, where insulin rates are
increased/decreased by 0.5 U/hr (or remain unchanged) every 2 hours, based
Table 4.3. Biostator algorithm at a glance
Biostator algorithm
Throughout the evolution of Biostator, a series of control algorithms were

developed. These newer algorithms addresses the various observations en-
countered in their predecessor algorithms, and attempts to fine-tune the
BG response. Broekhuyse et al [175] compared the various algorithms used
in Biostator (or so called “GCIIS - Glucose-controlled Insulin Infusion
System”). The original Albisser’s algorithm used both exogenous glucose
and exogenous insulin infusion in the control action. Later, modifications
in algorithms permitted a blunting of insulin delivery when glucose con-
centrations decline. This has eliminated the requirement of such a glucose
counter-infusion. The general structure of the earlier algorithm has the form:
1
I= MI [1 + tanh SI (Gp − BI )]
2
where I is the insulin infusion rate [mU/min], MI is the maximum insulin
infusion rate, SI is the sigmoidal steepness factor, BI is the glucose con-
centration at the half-maximum insulin infusion rate, Gp is a calculated
quantity called “projected glucose”:
A
G0 + K1 exp K2
−1 (Albisser)
3
Gp = G0 + K1 · A + K2 · A , A = (4G0 − G1 − G2 − G3 ) /10 (Toronto)
i=4
G0 + K1 exp RC
K2 −1
, RC = 1
15 i=1 24−i ΔG
Δt t−i
(Kraegen)
where A is the linearly weighted average rate of change of glucose concen-

tration based on the previous four minutes of monitoring; G0 = last minute’s
average glycaemia [mg/dl]; Gk = the average glycaemia k+1 minute previ-
ously [mg/dl] for k = 1, 2, 3, 4; K1 and K2 are constants (see [175]).
Biostator algorithm later introduced three different settings [16, 22, 176,
177]:
• Static: insulin release is dependent upon the static value of BG:

n
G0 − BI
IR = Ri · +1
Qi
• Dynamic: insulin infusion is solely controlled by the rates of change of

BG: n
G0 + GD − BI
IR = Ri · +1
Qi
• Dynamic+Static: the static with the dynamic control mode for insulin
infusion:
n n
G0 + GD − BI G0 − BI
IR = Ri · +1 − Ri · +1
Qi Qi
KR A2
+ 6A forA > 0
where GD = 10
KF A2
and n = 2 or 4 depending on how
10
+ 6A forA < 0
much insulin release is required [176].
The development of the control algorithm saw the use of KR when BG
is increasing and KF when BG is decreasing ( KR > KF ), instead of
maintaining a fixed control constant for both the increase and decrease
in BG.
Table 4.4. Example sliding scale table
BG range (mmol/l) Insulin infusion rate (U/hr)

>20.0 4
15.1 – 20.0 3
10.1 – 15.0 2
6.1 – 10.0 1
0 – 6.0 0
on the BG measurement. This method is similar to a sliding table, except that

the insulin rates are given in terms of increment or decrement to the previous
rate, instead of a fixed insulin rate for each BG ranges. Proposed by Watts
et al [8] to maintain BG in patients within a certain range, this method was
reported to be safe and efficacious [2, 8];
• experience, where control is achieved by means of accumulated experiences in
controlling glucose level;
• neuro-fuzzy method, where insulin rates are increased/decreased every 4 hours,
based on a simple nomogram (similar to a look-up table). The nomogram
details the insulin infusion rate variation to be added or subtracted from the
current rate based on “preceeding” and “present” BG values. The nomogram
is obtained by training a Back Error Propagation Neural Network using 1000
paired BG–insulin values, and then providing the neural network with 400 pairs
of BG values that represents every possible combination of glycaemic values be-
tween 3.3 and 13.9 mmol/l to obtain 400 corresponding variations of the insulin
infusion rates [166].
Note that the glucose-insulin response curve (and other algorithms such as Furler
et al’s infusion slope) can be discretised (or partitioned) to form a sliding table.
How closely the discretised sliding table approximates the original curve depends
on the selection of the BG ranges and the accuracy of the conventional infusion
pump (i.e. what is the smallest reliable infusion rate available).
4.2.4 Control Algorithm Based on PID Control
Proportional-Integral-Derivative control (or any combination of P, I or D) is an

easy-to-use feedback control system. It does not require advanced mathematics
to design and can be easily “tuned”. The controller takes a measurement from a
plant process and compares it with a setpoint (reference) value. The difference
(or “error” signal) is then used to adjust the input to the plant in order to
bring the measured value back to its desired setpoint (Fig 4.1). PID controller
can adjust process outputs based on the history and rate of change of the error
signal. However, PID controller are sensitive to dead-time in the measurement
(i.e. delay in measurement could upset the control action, bringing the control
loop into possible oscillations).
Fig. 4.1. PID controller
• Proportional (P): The error signal is multiplied by a constant Kp for imme-

diate correction.
• Integral (I): To learn from the past, the error signal is integrated (added up)
over a period of time, and then multiplied by a constant Ki . The result is
then added to the controller output signal.
• Derivative (D): The slope of the error signal (i.e. the change of the error signal
over a pre-defined interval) is calculated, and multiplied by a constant Kd ,
and the the result is added to the controller output signal. The larger the
derivative term, the more rapidly the controller response to changes in the
plant output.
PID controller can be divided into two forms:
1. “Position” Form
In Position form, the new value of the controller output is calculated directly,
i.e.: n
[e(t) − e(t − Δt)]
u(t) = Kp e(t) + Ki · Δt e(i) + Kd
i=1
Δt
where Δt = time between data samples (which must be smaller than the
process response time).
2. “Velocity” Form
In Velocity form, only the change in the controller output is calculated. The
actual controller output is then calculated from the previous value, i.e.:
Δu = u(t) − u(t − Δt)

= Kp [e(t) − e(t − Δt)] + Ki Δt · e(t)
P I
[e(t) − 2e(t − Δt) + e(t − 2 · Δt)]
+ Kd
Δt
D
Implementations in Literature
• Palazzo et al [178] simulated a PID controller on Bergman’s model.

• Chee et al [167, 179] implemented a step-wise PID controller on critically-ill
patient population.
• Steil et al [180] implemented a PID controller, which is subsequently used in a
“Long Term Sensor System, LTSS” system (MiniMed-Medtronic, Northridge,
CA, USA) (see [181]). LTSS combines an intravenous glucose sensor and an
implanted insulin pump that uses intraportal infusion route).
• Gupakumaram et al [182] implemented a P-D controller that has a feed-forward
component. This component models the pancreatic beta-cell response in addi-
tion to present and previous glucose values. The beta-cell insulin secretion is mod-
elled on a fading history of glucose levels and trends, where current and recent
glucose levels are weighted more heavily than older glucose values. The recent
trends in glucose variations are also weighted more heavily than those in the past:
t t
I(t) = K0 W0 (t) · G(t)dt + K1 W1 (t)Ġ(t)dt
0 0
t
+ K2 W2 (t)G̈(t)dt + higher order terms
0
where I(t) = insulin infusion rate; K0 ,K1 ,K2 = Gain factors; W0 ,W1 ,W2 =
weighting functions, and G(t) = blood glucose levels.
• Marchetti et al [183] simulated a PID controller in the velocity form on Hov-
orka’s model. The PID controller has the form

1 t de(t)
c(t) = c0 + Kc e(t) + e(t)dt + τD
τI 0 dt
where Kc and τD are tuned to minimise the objective function
t
∞
J= |e(k)| Δt + wi + c1 + c2
k
with wi = exp (−100(Gmin,i − 60)) being a constraint to avoid hypoglycemia;

c1 = exp (−10000(−0.01 − Kc )) being a constraint to avoid Kc values larger
than −0.01 mU dl/mg min, and c2 = exp (−10000(1500 − τD )) being the con-
straint to avoid τD values larger than 1500 min. The index i = c, u, o refers
to the response for a correct, under-estimated and over-estimated carbohy-
drate content of the meal; e(t) = control error (mg/dl); t∞ = duration of the
simulation, and Δt = sampling period.
Advantages and Disadvantages of Model-Less Approach
The advantages with model-less approach is that:

1. It is a simple and easy way to relate the input and output variables.
4.3 Model-Based Control Algorithms 67
2. Compared to model-based approach, the patient parameters do not need

to be worked out in advance, and the control algorithm can be deployed
instantly.
Nevertheless, as this approach is empirical in nature:
1. It assumes no knowledge of the underlying system and is likened to the Black
box approach.
2. It provides no insight into the functioning of the underlying system.
3. Its predictive power is very much restricted.
4. Successful operation of control algorithm still relies on human expertise to
some degree, for example, in the adjustment of the starting insulin rate (as
in [8]), or adaptation for use in patients with differing sensitivity to insulin.
4.3 Model-Based Control Algorithms

As the name implies, the model-based approach involves the use of a model in
the control of blood glucose level. This model is the human glucose-insulin in-
teraction. If this complex interaction can be captured and described in terms of
mathematics, then the glucose control problem becomes a mathematical prob-
lem, and mathematical problem can be solved using various mathematical tech-
niques (Note: We are yet to discover that it is not as easy as once thought to
model the glucose-insulin interaction).
Other advantages of using a model:
• It offers useful description and insight into the underlying process. For exam-
ple, the observations obtained from accessible variables (e.g. glucose at tissue)
can be used to measure those system quantities that are of interest, but were
inaccessible to direct measurement (e.g. glucose at hepatic artery).
• It can help in predicting overall system behaviour under a variety of pertur-
bations [168]. In another word, it can serve to determine how a system would
respond to a stimulus or changes in the system.
• It allows the testing or simulation of the control algorithm to be performed
without involving real patients.
• It allows the study into the effect of insulin therapy regime to be undertaken
without risking patient safety.
Nevertheless, we must understand that whether or not a certain model is valid
in terms of its representation (approximation) of a real system really depends
on the context (or purpose) to which it is designed for.
Furthermore,
• Lack of understanding about the underlying process limited the extent of
derivation of the model.
• The predictive power of the model is limited by the extent to which the model
is valid or accurate.
• Some parameters in the model may not be always observable, and if they
cannot be measured in anyway (or are too difficult to be measured), then this
will limit the usefulness and applicability of the model.
The mathematical model used in BG regulation can be divided into two groups:
4.3.1 Theoretical
Theoretical modelling attempts to model the underlying process by means of a

theory. The theory is derived using the knowledge about the function or structure
of the underlying system, chemical process, physical laws, and quantitative data
(from observable measurements).
The extent to which a pure theoretical model can be derived depends on
current understanding of the system or process. It may not always be possible to
fully model the underlying process, either due to a lack of measurable variables,
or simply lack of detailed knowledge about the system.
4.3.2 Empirical + Theoretical
In metabolic system, it is rare to have adequate detailed knowledge of the un-

derlying system to allow full derivation of the theoretical models.
Hence, a combination of theoretical and empirical knowledge is used. Cer-
tain parts of the system are modelled empirically, whereas others are modelled
theoretically.
We could postulate about the internal structure of a system based on empirical
observation, and fit the hypothesis to the observation.
When deriving the glucose-insulin model, one could conceptualise the model
based on
• physiological knowledge of the system,
• functional description of the relevant processes
• the interconnection between the processes, and
• Relationship between the processes to the observable glucose (or insulin) mea-
surements in practical situation
Certain elements of the model could also be, as according to [168]:
• Abstractised, i.e. only certain aspects of the blood glucose regulation system
are considered in a model. For example, bloodstream is regarded as containing
only glucose and hormones involved in its regulation. Features that do not
directly relate to glucose regulation are omitted.
• Aggregated, i.e. lumping different components into a single entity for the pur-
pose of modelling. For example, the tissues that constitute the glucose uti-
lization space are normally lumped together, thus ignoring individual cellular
properties.
4.4 Mathematical Models of Gluco-regulatory System 69
• Idealised, i.e. some structures or behaviours that are difficult to describe or

treat are approximated by more simple idealized ones. For example, insulin
injected directly to the bloodstream takes effect instantaneously although in
fact its action is delayed.
Before continuing with a discussion of model-based control algorithms, it is useful
to have some knowledge of the mathematical models that were used in the design
of the algorithms.
4.4 Mathematical Models of Gluco-regulatory System

The attempts to capture the glucose-insulin mechanism have resulted in the
formulation of various glucose-insulin kinetic models. These models range from
simple expressions that relate glucose and insulin, to very complete mathematical
models. The three general groupings of mathematical models are:
1. Linear (see e.g. Bolie et al [184], Ackerman et al [15, 185–187], Ceresa et
al [188], Chorbajian et al [189], Cerasi et al [190], Nomura et al [191], Bajaj
et al [192, 193], Jansson et al [194] and Salzsieder et al [195])
2. Non-linear (see e.g. Bergman et al [196–198], Furler et al [176], Candas &
Radziuk et al [199], Doyle III et al [200] and De Gaetano et al [201])
3. Comprehensive (see e.g. Cobelli et al [202–204], Parrish et al [205], Sorensen
[206–209], Hovorka et al [210,211], Giugliano et al [212], and Guyton et al [4]).
4.4.1 Linear Models
Linear models are adequate when the intrinsic dynamics of the metabolic system
are essentially linear1 . In linear modelling of glucose-insulin kinetics, the models
are described by linear time-invariant equations:
ẋ(t) = Ax(t) + Bu(t)

y(t) = Cx(t) + Du(t); t0 < t < T (4.1)
where the state variable x(t) and its derivative appear in linear combination
only, and u(t) represents the input (or disturbances) into the system. Models
of glucose-insulin kinetics can be derived using technique like “compartmental
analysis”.
1
Quoted from [213]: A system is called “linear” if the Principle of Superposition
applies. The Principle of Superposition states that the response produced by a si-
multaneous application of two different forcing functions is the sum of the two in-
dividual responses. Hence, for the linear system, the response to several inputs can
be calculated by treating one input at a time and adding the results.
Conversely, a system is “nonlinear” if the Principle of Superposition does not
apply. Thus, for a nonlinear system the response to two inputs cannot be calculated
by treating one input at a time and adding the results.
Compartmental Analysis (Adapted from [168])
Compartmental analysis is the simplest method in biomathematics to describe

the transfer of materials in biological systems, and it can quickly lead to math-
ematical relationships.
A compartment is fundamentally an idealised store of a substance. Compart-
mental analysis consists of studying the exchanges of matter between the stores
(i.e. compartments) as a function of time, t. The material exchange between
compartments takes place either by physical transport from one location to an-
other, or by chemical reactions (Fig 4.2). The mathematical model then consists
of the mass balance equations for each compartments and relations describing
the rate of material transfer between compartment:
dQij
= Rij Rji
dt
where Qij = quantity of substance
in compartment i that interchanges matter
with other compartments; Rij = summation of the rates of mass transfer into
compartment i from all relevant compartments; Rji = summation of the rates
of mass transfer from compartment i to other compartments of the system.
For example, consider two compartments 1 and 2. Figure 4.2 shows the flow
of material between the two compartments. Kij indicates the rates at which
the materials in i is transferred to compartment j and vice versa. Let Q1 , Q2
Fig. 4.2. Compartmental analysis
= quantity of materials in compartment 1 and 2 respectively, and J(t), K(t)

= flow of material from exogenous sources, the mass balance equations can be
written as follows:
dQ1
= −k11 Q1 − k12 Q1 + k21 Q2 + J(t)
dt
dQ2
= k12 Q1 − k21 Q2 − k22 Q2 + K(t)
dt
which can be simplified to:

dQ1
= −m1 Q1 + k21 Q2 + J(t)
dt
dQ2
= k12 Q1 − m3 Q2 + K(t) (4.2)
dt
where m1 = (k11 + k12 ) and m3 = (k21 + k22 ).
The form taken by equation 4.2 above is the most common form of models
for linearised model of glucose-insulin, where the equation consist of glucose and
insulin as the primary compartments.
4.4.2 Ackerman’s Linear Model
There are many linear models proposed in the literature to-date. Among them,
Ackerman’s model has been by far the most commonly cited linear model. Ack-
erman’s model consists of a system of equations in which the parameters have
been lumped into two dependent variables, G and H, and four rate constants:
dg
= −m1 g − m2 h + J
dt
dh
=−m3 h + m4 g + K (4.3)
dt
where G= glucose concentration; G0 = fasting glucose concentration; g ≡ G−G0 ;
H= blood hormone concentration (including insulin). This is the effective hor-
mone concentration which included consideration of the role of epinephrine and
other hormone; H0 = fasting blood hormone concentration (including insulin);
h ≡ H − H0
The rate constants are:
• m1 = rate constant for the removal of glucose above the initial fasting level due
to its own excess above the initial level. Also known as glucose effectiveness,
this term normally has a value of 0.01–0.02/min but it is likely that its true
value is reduced at chronic hyperglycemia due to glucose toxicity;
• m2 = rate constant for the removal of glucose above the initial level due to
blood hormone concentration above the initial level
• m3 = rate constant for the removal of hormone above the initial fasting level
due to its own excess above the initial level (duration of insulin action)
• m4 = rate constant for the removal of hormone above the initial level due to
blood glucose concentration above the initial level.
Note the similarities between equation 4.3 and the compartmental model (equa-
tion 4.2) ). The same sets of equation 4.3 can also be derived using compart-
mental analysis.
The model parameters for the model were obtained from fitting the glucose
and insulin concentration profile to the model, following glucose tolerance tests
in patients. Some typical values can be found in Appendix B.
Table 4.5. Bolie vs Ackerman
Bolie’s Model
In 1961, Bolie published his pioneering work on mathematical modelling

of the glucose-insulin interaction, which was presented to show that “the
normally functioning glucose homeostasis mechanism exhibits physiological
coefficients which approximate the well-known critical damping criteria of
servo-mechanism theory” [184].
The model is represented by the equations:
dH 1 F1 (H) F2 (G)
= U̇ − +
dt V V V
dG 1 F3 (G, H) F4 (G, H)
= Ṗ − −
dt V V V
where t = time [hours]; G = extracellular glucose concentration; H = ex-
tracellular insulin concentration; V = extracellular fluid volume [L]; U̇ =
rate of insulin injection [U/hr]; Ṗ = rate of glucose injection [g/hr]; F1 (H)
= rate of insulin destruction; F2 (G) = rate of insulin production; F3 (G, H)
= rate of liver accumulation of glucose; F4 (G, H) = rate of tissue utilization
of glucose.
Assume that the fluctuations in the insulin and glucose concentrations are
limited to small physiological variations, then the non-linear model can be
linearised using Taylor series:
y = f (x)
df 1 d2 f
= f (x̄) + (x − x̄) + (x − x̄)2 + . . . (4.4)
dx 2! dx2
to produce
dh 1 ∂F1 1 ∂F2
=u− ·h+ ·g
dt V ∂H H0
V ∂G G0
dg 1 ∂F3 ∂F4 1 ∂F3 ∂F4

=p− + ·h− + ·g
dt V ∂H ∂H H0 ,G0 V ∂G ∂G H0 ,G0
where h = H − H0 , g = G − G0 , u = V1 U̇ , and p = V1 Ṗ .
The equation above can be identified with Ackerman’s model in terms of
the parameters m1 , m2 , m3 and m4 .
4.4.3 Non-linear and Comprehensive Models
The linear model has the disadvantage that it is a gross oversimplification of

the underlying glucose-insulin interaction in actual human (which is far more
complex than the linear model). Biological system dynamics are often non-linear
in nature, and low-order models may not adequately describe the real process
and therefore could contain both unacceptable levels of modelling error and
significant process-model mismatch [200].
Non-linear model ranges from less complex ones (e.g. [196, 197], [176] and
[214]) to comprehensive ones (e.g. [202–204], [215, 216] and [217]). Comprehen-
sive model attempts to coalesce the knowledge of metabolic regulations into a
generally large, non-linear model of high order, with a large number of model
parameters. This included the modelling of the distribution and metabolism of
glucose and insulin, hepatic glucose balance (i.e. glucose production and dis-
posal), renal excretion, glucose utilization, and insulin release and degradation,
to describe the system thoroughly ( [202–204]). Some investigators even took the
molecular approach, modelling individual isolated beta-cells, followed by popu-
lations of beta-cells (see e.g. [212]). Comprehensive models, in general, cannot
be easily identified.
In this section, only a few of the nonlinear models commonly referred in the
literature are described.
4.4.4 Minimal Model
Also known as Bergman’s model, the minimal model was proposed by Bergman
et al [196, 197] in the early 1980’s for the interpretation of glucose and in-
sulin plasma concentrations following the intravenous glucose tolerance test
(IVGTT). This model has been popular among past physiological researches
on the metabolism of glucose.
It is composed of two parts:
1. Minimal model for glucose disappearance: Two differential equations
which describe the glucose plasma concentration time-course, accounting for
the dynamics of glucose uptake dependent on, and independent of circulating
insulin. It treated insulin plasma concentration as a known forcing function
[201]
dG(t)
= − (p1 + X(t)) G(t) + p1 GB (4.5)
dt
dX(t)
= −p2 X(t) + p3 [I(t) − IB ] (4.6)
dt
where the term p1 GB accounts for the body’s natural tendency to move
towards basal glucose levels [177].
2. Minimal model for insulin kinetic: Single equation which describes the
time-course of plasma insulin concentration, accounting for the dynamics of
pancreatic insulin release in response to the glucose stimulus. The glucose
plasma concentration is to be regarded as a known forcing function.

dI(t) γ[G(t) − h]t − n[I(t) − IB ] f or G(t) − h > 0
= (4.7)
dt −n[I(t) − IB ] f or G(t) − h ≤ 0
where the state variables:

• G(t) [mg/dl] = the blood glucose concentration at time t [min];
• I(t) [μU/ml] = blood insulin concentration at time t [min];
• X(t) [min−1 ] = a function representing insulin-excitable tissue glucose
uptake activity, proportional to insulin concentration in a “distant” com-
partment;
• GB [mg/dl] = the subject’s basal glucose level (or pre-injection value of
glucose concentration);
• IB [μU/ml] = the subject’s basal insulin level (or pre-injection value of
insulin concentration);
• Go [mg/dl] = the theoretical glucose level at time 0 after the instanta-
neous glucose bolus, i.e. glucose concentration that would be obtained
immediately after glucose injection if there were instantaneous mixing of
glucose in the extracellular fluid compartment [218].
The parameters are:
• γ [(μU/ml)/(mg/dl)−1 min−1 ] = the rate of pancreatic release of insulin
after the bolus, per minute and per mg/dl of glucose concentration above
the “target” glycaemia;
• h [mg/dl] = the pancreatic “target glycaemia” [201]. It is a threshold value
such that only glucose level above h will effect a secretion of insulin [219].
It represents the critical value of plasma glucose at which glucose begins
to have a marked influence on the magnitude of the second phase;
• n [min−1 ] = the time constant for insulin disappearance [197], or fractional
disappearance rate constant for endogenous insulin [198];
• Io [μU/ml] = the theoretical plasma insulin concentration at time 0, above
basal insulinemia, immediately after the glucose bolus.
The two parts are to be separately estimated on the available data. That is, the
model parameter fitting has to be done in two steps:
1. using the recorded insulin concentration as input data, in order to derive the
parameters in the first two equations
2. then, using the recorded glucose as input data to derive the parameters in
the third equation.
The glucose and insulin recordings are obtained by first injecting a bolus of glu-
cose into the bloodstream of the experimental subject, to induce an (impulsive)
increase in the plasma glucose concentration G(t) and a corresponding increase
of the plasma insulin concentration I(t), secreted by pancreas. These concen-
trations are measured during a three-hour time interval beginning at injection,
until G(t) and I(t) have returned to normal [201].
To represent a person in a diabetic state, the original model has taken the
general form:
dG(t)
= − (p1 + X(t)) G(t) + p1 GB + p(t)
dt
dX(t)
= −p2 X(t) + p3 [I(t) − IB ]
dt
dI(t)
= −n[I(t) − IB ] + u(t)
dt
where endogenous insulin secretion (i.e. the term γ[G(t) − h]t) in equation 4.7
was removed, and a term of exogenous infusion of glucose p(t) and insulin u(t)
was added.
There were variations to the Minimal Model, for example, Furler et al [176]
adapted the original Minimal model to represent the diabetic state and included
insulin antibodies in the description of insulin dynamics; Van Herpe et al [220]
modified Minimal model for the Intensive Care Unit (ICU) population (called
“ICU-MM” model by the authors).
4.4.5 Cobelli’s Model
Cobelli et al’s model consists of a metabolic plant (glucose), and two-hormone

controller (insulin and glucagon) [202–204]. The glucose subsystem is described
by a one-compartment model of distribution and metabolism (extracellular flu-
ids), involving net hepatic glucose balance (i.e. the difference between liver glu-
cose production and liver uptake), renal excretion of glucose, insulin-dependent
glucose utilization (mainly by muscle and adipose tissue), insulin-independent
glucose utilization (mainly by the central nervous system and the red blood cell).
Glucose Subsystem
ẋ1 (t) = NHGB(x1 , u12 , u2 ) − F3 (x1 ) − F4 (x1 , u13 ) − F5 + Ix (t), x1 (0) = x10
u̇1p (t) = −k21 u1p + k12 u2p + W (x1 ), u1p (0) = u1p0
u̇2p (t) = k21 u1p − (k12 + k02 (x1 )) u2p , u2p (0) = u2p0
where NHGB = F1 (x1 , u12 , u2 ) − F2 (x1 , u12 ) is the net hepatic glucose balance;
F1 = the liver glucose production; F2 = the liver glucose uptake; F3 = the re-
nal excretion; F4 = the peripheral insulin-dependent glucose utilization; F5 =
the peripheral insulin-independent glucose utilization; Ix (t) is the rate of exoge-
nous glucose given intravenously; W (x1 ) = insulin synthesis controlled by BG
concentration; x1 = quantity of glucose in plasma and extracellular fluids [mg];
u1p = quantity of pancreatic stored insulin [μU]; u2p = quantity of pancreatic,
promptly releasable insulin [μU]. The constants are k12 =0.01, k21 =4.34×10−3
(values for a normal state).
• Liver glucose production
F1 (x1 , u12 , u2 ) = a11 G1 (u2 )H1 (u12 )M1 (x1 )
where
Fig. 4.3. Cobelli et al’s compartmental model. Solid line represents material flow;
Dashed line represents control signal (Adapted from Cobelli et al [202]).
G1 (u2 ) = 0.5 {1 + tanh [b11 (e21 + c11 )]}

H1 (u12 ) = 0.5 {1 − tanh [b12 (e12 + c12 )]}
M1 (x1 ) = 0.5 {1 − tanh [b13 (ex + c13 )]}
Baseline parameter values for a normal state: a11 = 1.51; b11 = 2.14; b12 =
0.0728; b13 =0.0275; c11 =−0.85; c12 =7; c13 =20.
• Liver glucose uptake
F2 (x1 , u12 ) = H2 (u12 )M2 (x1 )
where
H2 (u12 ) = 0.5 {1 − tanh [b21 (e12 + c12 )]}
M2 (x1 ) = a221 + a222 0.5 {1 + tanh [b22 (ex + c22 )]}
Baseline parameter values for a normal state: a221 = 0.00195; a222 = 0.00521;
b21 = 0.0111; b22 =0.0145; c12 =51.3; c22 =−108.5.
• Renal excretion of glucose
F3 (x1 ) = M31 (x1 )M32 (x1 )
where
M31 (x1 ) = 0.5 {1 + tanh [b31 (y1 + c31 )]}
M32 (x1 ) = a321 y1 + a322
Baseline parameter values for a normal state: a321 = 1.43×10−5; a322 = −1.31×
10−5 ; b31 = 20; c31 =−180.
• Insulin-dependent peripheral glucose utilization
F4 (x1 , u13 ) = a41 H4 (u13 )M4 (x1 )
where
H4 (u13 ) = 0.5 {1 + tanh [b41 (e13 + c41 )]}
M4 (x1 ) = 0.5 {1 + tanh [b42 (ex + c42 )]}
Baseline parameter values for a normal state: a41 =0.0287; b41 =0.031;
b42 =0.0144; c41 =−50.9; c42 =−20.2.
• Insulin-independent glucose uptake
F5 (x1 ) = M51 (x1 )M52 (x1 )
where
M51 (x1 ) = a51 tanh [b51 (ex + c51 )]
M52 (x1 ) = a52 ex + b52
Baseline parameter values for a normal state: a51 =1.01×10−3; a52 =4.6×10−6;
b51 =0.0278; b52 =4.13×10−4.
Insulin Subsystem
The insulin subsystem is described by a five-compartment model, involving pan-
creatic insulin storage, liver and portal plasma insulin, plasma insulin and insulin
in the interstitial fluid.
u̇11 (t) = − (m01 + m21 + m31 ) u11 + m12 u12 + m13 u13 + Iu (t), u11 (0) = u110
u̇12 (t) = − (m02 + m12 ) u12 + m21 u11 + k02 (x1 )u2p , u12 (0) = u120
u̇13 (t) = −m13 u13 + m31 u11 , u13 (0) = u130
where u11 = the quantity of insulin in plasma [μU]; u12 = the quantity of
insulin in the liver [μU]; u13 = the quantity of insulin in the interstitial fluid
[μU]; Iu (t) = the insulin test input. The constants are m01 =0.125, m02 =0.185,
m12 =0.209, m13 =0.02, m21 =0.268, m31 =0.042 (values for a normal state). The
term k02 (x1 )u2p = F6 (u2p , x1 ) represents the insulin secretion rate.
• Insulin synthesis
W (x1 ) = 0.5aw {1 + tanh [bw (ex + cw )]}
Baseline parameter values for a normal state: aw =0.287; bw =0.0151; cw =−92.3.

• Insulin secretion
F6 (u2p , x1 ) = 0.5a6 {1 + tanh [b6 (ex + c6 )]} u2p
Baseline parameter values for a normal state: a6 =1.3; b6 =0.0923; c6 =−19.68.
Glucagon Subsystem
The glucagon subsystem is described by a one-compartment model:
u̇2 (t) = −h02 u2 + F7 (x1 , u13 ), u2 (0) = u20
where u2 = the quantity of glucagon in the plasma and interstitial fluid [ng]; F7
= the endogenous release of glucagon, dependent on BG and interstitial fluid
insulin; h02 =0.086.
• Glucagon secretion
F7 (x1 , u13 ) = a71 H7 (u13 )M7 (x1 )
where
H7 (u13 ) = 0.5 {1 − tanh [b71 (e13 + c71 )]}
M7 (x1 ) = 0.5 {1 − tanh [b72 (ex + c72 )]}
Baseline parameter values for a normal state: a71 =2.35; b71 =6.86×10−3;
b72 =0.03; c71 =99.2; c72 =40.
W and F1 – F7 are nonlinear functions, mij , hij , and kij are constant rate param-
eters [min−1 ] with the exception of k02 which is a function of x1 . See [202–204] for
details on the identification of individual parameters, and their nominal values.
The nominal values are shown in Appendix B. See also [215, 216] for extensions
based on Cobelli et al’s model.
4.4.6 Candas and Radziuk
Candas & Radziuk [199] used this non-linear model to aid in the design of an
adaptive controller for the maintenance of basal glycaemia during euglycaemic
hyperinsulinemic clamps:
dG(t)
= − [ko + k(t)] G(t) + RG(t)
dt
dk(t)
= −a1 · k(t) + a2 · i(t)
dt
di(t)
= −a3 · i(t) + a4 · k(t) + a6 · i3 (t) + RI(t)
dt
di3 (t)
= −a6 · i3 (t) + a5 · i(t)
dt
where G(t) = plasma glucose concentration [mg/dl]; ko = insulin-independent
fractional removal rate of glucose [min−1 ]; k(t) = insulin-dependent fractional
removal rate of glucose [min−1 ]; i(t) [μU] = insulin mass in the central com-
partment; i3 (t)= insulin mass in a peripheral compartment non-active in glu-
cose removal [μU] ; a1 –a6 = fractional transfer rates of the three-compartment
model of insulin kinetics (a1 , a3 , a5 , a6 have units in min−1 , and a2 has unit of
min2 /μU, and a4 is in μU/(min2 )). The values for {ai } are 0.394, 0.142, 0.251,
0.394, 3.15 × 10−8 , and 2.8 × 103 , while ksc was assumed to be 0.03 min−1.
RG(t) accounts for the rate of appearance of glucose in the systemic circulation
from both the endogenous and exogenous sources (in [mg/ml · min]), and RI(t)
represents the entry of both the endogenous and exogenous insulin into the
systemic circulation (in [μU/min]).
4.4.7 Hovorka’s Model
Hovorka et al [211] has originally developed a glucoregulatory model to represent

glucose kinetics during basal conditions and during the intravenous glucose tol-
erance test in healthy subjects [210, 221]. The model was extended to represent
the relationship between subcutaneous insulin infusion as input and intravenous
glucose concentration as output, as part of ADICOL project (an European Union
initiative).
The extended model consists of a glucose subsystem (glucose absorption, dis-
tribution/transport and disposal), an insulin subsystem (insulin absorption, dis-
tribution and disposal) and an insulin action subsystem (insulin action on glucose
transport, disposal and endogenous production).
Hovorka’s model, as according to [211] and [210] is presented below:
Glucose Subsystem
The core model is a two-compartment representation of glucose kinetics

c
dQ1 (t) F01
=− + x1 (t) Q1 (t) + k12 Q2 (t) − FR + UG (t) + EGP0 [1 − x3 (t)]
dt VG G(t)
dQ2 (t) Q1 (t)
= x1 (t)Q1 (t) − [k12 + x2 (t)] Q2 (t); y(t) = G(t) =
dt VG
where Q1 = masses of glucose in the accessible compartment (i.e. where mea-
surements are made) [mmol]; Q2 = masses of glucose in the non-accessible com-
partment [mmol]; k12 = transfer rate constant from the non-accessible to the
accessible compartment [min−1 ]; VG = distribution volume of the accessible com-
partment [Litre]; y and G is the (measurable) glucose concentration [mmol/l],
and EGP0 = endogenous glucose production extrapolated to the zero insulin
concentration [mmol/min].
Fig. 4.4. Hovorka’s model (Adapted from Hovorka et al [211])
c
F01 is the total insulin-independent glucose flux, corrected for the ambient
glucose concentration [mmol/min]:

c F01 if G ≥ 4.5 mmol/l
F01 =
F01 G/4.5 otherwise
FR is the renal glucose clearance above the glucose threshold of 9 mmol/l

0.003(G − 9)VG if G ≥ 9 mmol/l
FR =
0 otherwise
The gut absorption rate UG is represented by a two-compartment chain with

identical transfer rates of 1/tmax,G
DG AG · t exp(−t/tmax,G)
UG (t) =
t2max,G
where tmax,G = the time-maximum appearance rate of glucose in the accessible

glucose compartment; DG = amount of carbohydrate digested; AG = carbohy-
drate bio-availability; UG in [mmol/min].
Insulin Subsystem
Insulin absorption is described as
dS1 (t) S1 (t)

= u(t) −
dt tmax,I
dS2 (t) S1 (t) S2 (t)
= −
dt tmax,I tmax,I
where S1 , S2 are a two-compartment chain representing absorption of subcuta-

neously administered short-acting insulin (e.g. Lispro); u(t) = administration of
insulin (bolus and infusion); tmax,I = time-to-maximum insulin absorption. The
insulin absorption rate (i.e. appearance of insulin in plasma) can be obtained as
S2 (t)
U1 =
tmax,I
The plasma insulin concentration I(t) [mU/l] is described by
dI(t) UI (t)
= − ke I(t)
dt VI
where ke = fractional elimination rate [min−1 ], and VI is the distribution volume

[L].
Insulin Action Subsystem
The three insulin actions on glucose kinetics are represented by

dx1 kb1
= −ka1 x1 (t) + kb1 I(t), x1 (0) = Ib
dt ka1
dx2 kb2
= −ka2 x2 (t) + kb2 I(t), x2 (0) = Ib
dt ka2
dx3 kb3
= −ka3 x3 (t) + kb3 I(t), x3 (0) = Ib
dt ka3
where x1 , x2 , x3 = (remote) effects of insulin on glucose distribution/ trans-
port, glucose disposal and endogenous glucose production, respectively [min−1 ];
ka1 , ka2 , ka3 = deactivation rate constants [min−1 ]; kb1 , kb2 , kb3 = activation rate
constants [min−2 per mU/l]; Ib = basal plasma insulin [mU/l].
In conjunction with this model, a non-linear Model Predictive Controller with
on-line parameter estimation has been developed (see Section 4.5.4). The follow-
ing parameterisation has been made for the controller:
• Insulin sensitivity of distribution/transport
f kb1
SIT =
ka1
• Insulin sensitivity of disposal
f kb2
SID =
ka2
• Insulin sensitivity of EGP
f kb3
SIE =
ka3
The model quantities were divided into model constants and model parameters,
mainly to reduce the number of parameters to be calculated, and at the same
time, to be able to represent a wider range of glucose excursion observed in
diabetic patients [211]. Some values for these model quantities can be found in
Appendix B.
4.4.8 Sorensen’s Model
The model presented here is mainly adapted from [209], with some parameters
from [207] and [208], which are originally based on Sorensen’s work [217, 222].
Using the compartmental modelling technique, the patient model is represented
in Fig 4.5. Individual compartment models were obtained by performing mass
balances around tissues important to glucose or insulin dynamics. There are
six compartments, i.e. brain, heart/lungs, gut (combined effects of stomach and
intestine), liver, kidney, and periphery (combined effects of muscle and adipose
tissue). Blood transported glucose or insulin to the various compartments. It
was assumed that the glucose or insulin concentration in a compartment was in
equilibrium with the blood leaving the given compartment.
The nomenclature and parameter values can be found in [209] and are re-
peated in Appendix B.
Mass Balance for Glucose
• Brain:
dGBV QG VBI
= GB (GH − GBV ) − G (GBV − GBI )
dt VBV VBV TB
dGBI VBI ΓBGU
= (GBV − GBI ) −
dt VBI TB VBI
• Heart and lungs:
dGH 1
= G · QG G G G G
B GBV + QL GL + QK GK + QP GP V − QH GH − ΓRBCU
dt VH
• Gut:
dGG QG 1
= G · (GH − GG ) + G · (Γmeal − ΓGGU )
dt VGG VG
Fig. 4.5. Comprehensive glucose-insulin modelling (Adapted from [207, 208])
• Liver:
dGL 1
= G · QG G G
A GH + QG GG − QL GL + ΓHGP − ΓHGU
dt VL
• Kidney:
dGK QG ΓKGE
= K · (GH − GK ) −
dt VKG VKG
• Periphery:
dGP V QG VP I
= GP · (GH − GP V ) − G G · (GP V − GP I )
dt VP V TP V P V
dGP I VP I ΓP GU
= G · (GP V − GP I ) −
dt TP V P I VP I
Metabolic Sources and Sinks (Glucose Compartment)
• Peripheral glucose uptake:
ΓP GU = ΓPBGU · GN N
P I 7.03 + 6.52 tanh 0.388 IP I − 5.82
• Hepatic glucose production:

B I
ΓHGP = ΓHGP MHGP 2.7 tanh 0.39χN − f2 1.42 − 1.41 tanh 0.62 GN
L − 0.497
d I 1
(MHGP )= 1.21 − 1.14 tanh 1.66 ILN − 0.89 I
− MHGP
dt τI
df2 1 2.7 tanh(0.39χN ) − 1
= − f2
dt τχ 2
• Hepatic glucose uptake:

B I
ΓHGU = ΓHGU MHGU 5.66 − 5.66 tanh 2.44 GN
L − 1.48
d I 1
(MHGP )= 2.0 tanh 0.55ILN − MHGU
I
dt τI
• Kidney glucose excretion:
71 + 71 tanh [0.11 (GK − 460)] , 0 ≤ GK < 460 mg/dl

ΓKGE =
−330 + 0.872GK , GK ≥ 460 mg/dl
Mass Balance for Insulin
• Brain:
dIB QI
= BI (IH − IB )
dt VB
• Heart and lungs:

dIH 1
= I · QIB IB + QIL IL + QIK IK + QIP IP V − QIH IH + i(t)
dt VH
• Gut:
dIG QI
= G · (IH − IG )
dt VGI
• Liver:
dIL 1
= I · QIA IH + QIG IG − QIL IL + ΓP IR − ΓLIC
dt VL
• Kidney:
dIK QI ΓKIC
= K · (IH − IK ) −
dt VKI VKI
• Periphery:
dIP V QI VP I
= IP · (IH − IP V ) − I I · (IP V − IP I )
dt VP V TP V P V
dIP I VP I ΓP IC
= I I · (IP V − IP I ) −
dt TP V P I VP I
Metabolic Sources and Sinks (Insulin Compartment)

• Liver insulin clearance:

ΓLIC = FLIC QIA IH + QIG IG + ΓP IR
where FLIC = 0.40
• Peripheral insulin clearance:
ΓP IR = 0.0 (no pancreatic insulin release)
• Kidney insulin clearance:

ΓKIC = FKIC QIK IK
where FKIC = 0.30
• Peripheral insulin clearance:
I
ΓP IC = PI
TPI
1−FP IC
FP IC
1
QIP
− VP I
where FP IC = 0.15
Glucagon Model
dχN 1
= ΓMχ C · ΓPNχ R − ΓMχ C · χN
dt Vχ
• Pancreatic glucagon release:
ΓPNχ R = 2.93 − 2.10 tanh 4.18 GN

H − 0.61 ·
N
1.31 − 0.61 tanh 1.06 IH − 0.47
where ΓMχ C = 0.91 L/min, and Vχ = 9.94 L.

4.4.9 Sub-models That Complements the Comprehensive Models
• Subcutaneously injected insulin:

A first order model can be added to describe the transfer of the insulin mass
from the subcutaneous (s.c.) depot to the systemic circulation [199]:
dM (t)
= −ksc · M (t) + RIsc (t)
dt
IR(t) = ksc · M (t)
where M (t) = insulin mass in the s.c. depot [μU ]; RIsc (t) = insulin rate of
appearance in the s.c. depot [μU/min]; ksc (t) = fractional transfer rate from
s.c. depot toward the systemic circulation [min−1 ].
• Rate of glucose absorption via the gut wall:

Lynch et al [223] cited the equations for the rate of glucose absorption from
the gut wall [in mg/min]:
RGabs = Kgabs Ggut
where Ggut = the amount (mg) of glucose in the gut following ingestion of a
meal and is defined by the following differential equation:
dGgut
= RGempty − Kgabs Ggut
dt
where RGempty = the rate of gastric emptying which is described by a trape-
zoidal function, saturating at a certain maximal rate. This model was evalu-
ated by Yates et al [224].
• Subcutaneous glucose infusion:

A linear, first-order system in the frequency domain was introduced in [225]
to be used in conjunction with Cobelli et al’s model:
Ksc
Gsc (s) = · exp(−tm s)
1 + tsc s
where Ksc = the gain; tsc = time delay between blood glucose and s.c. glucose
(5 min); tm = time delay due to the tubing in an ex-vivo monitoring system
(0–20 min).
• Subcutaneous glucose sensor:

Rebrin et al [25] presented a model describing ISF glucose-sensor signal dy-
namics, characterized to allow for a gradient between plasma and interstitial
glucose:
dC2 V1
= − (k02 + k12 ) C2 + k21 C1
dt V2
where C1 and C2 are plasma and ISF glucose kinetics respectively, and V1 and
V2 are plasma and ISF volume respectively. By substituting sensor current
4.5 Mathematics of the Model-Based Control Algorithms 87
Fig. 4.6. ISF glucose-sensor signal dynamics (Adapted from [25])
s = αC2 , the sensor signal can be described by

ds
= −p2 s + p3 C1
dt
where p2 = k02 + k12 and p3 = k21 V1 /V2 α which can be identified from plasma
glucose and sensor current data.
Quotes
Quote from [168], “It is clear that metabolic processes exhibit complex dynamics
and involve many levels of control action ranging from autoregulatory effects
occurring within the cell to hormonal control of metabolic systems at the global
level. The approach that should be adopted in modelling metabolic system is
conditioned, along with purpose and theory, the nature of the data available for
system identification and model validation.”
4.5 Mathematics of the Model-Based Control Algorithms

As previously stated, the availability of glucose-insulin model allows glucose-
control problem to be treated as mathematical problems, and solutions (control
algorithms) can be derived using various mathematical techniques. For linear
glucose-insulin model, analytical results are obtainable in most cases. For non-
linear models, the models are either linearised to reduce the number of pa-
rameters to be determined (and take advantage of the available mathematical
techniques for linear models), or the control problem are solved numerically.
4.5.1 Pole Placement
This is a standard technique used in control systems design. It can be shown

that if the system considered is completely state controllable, then the poles of
the closed-loop system may be placed at any desired locations by means of state
Table 4.6. Definitions from Ricker [226]
• Plant:
This represents the real “patient”. The states of a plant are never fully
measureable. One would not know the “parameters” or even the “order”
of the plant, since it is generally a time-varying, non-linear, distributed-
parameter system. The values of the signals leaving the controller are
known, and the plant outputs can be measured, but nothing else in the
plant is accessible. The plant model is not part of the control system. Its
is only used to represent the plant in simulations and analytical work.
• Internal model:
The internal model provides a prediction of future plant outputs as a
function of contemplated adjustments in the manipulated variables and
estimated disturbances, The controller chooses the values of control-
inputs to send to the plant such that the predicted plant outputs are
optimal according to a specified criterion. The internal model is part of
the control system, and its states are all known exactly. Furthermore,
the structure and parameters of the internal model are known.
feedback through an appropriate state feedback gain matrix [213]. Consider a

control system
ẋ(t) = Ax + Bu
where x= state vector (n-vector); u= scalar control signal; A = n × n constant
matrix; B = n × 1 constant matrix. If we choose u = −Kx, then we have
ẋ(t) = (A − BK) x(t)
If matrix K is chosen properly, the matrix A − BK can be made asymptotically

stable, and for all x(0) = 0, it is possible to make x(t) approach 0 as t approaches
infinity.
4.5.2 Optimal Control
In general, the optimal control problem attempts to find a feedback controller

u(t) such that a linear system of the form (4.1) is satisfied, and a quadratic
performance criterion ∞
2
J(u) = (x(t) − xd ) dt
0
is minimised. Minimising the criterion is equivalent to making x(t) as close to

xd as possible. xd is a fixed constant from biological considerations.
Mathematically, the optimal control problem can be stated as
∞
2
min (x(t) − xd ) dt
u(t) 0
The application of optimal control theory in glucose control was reported in

several papers (e.g. [227, 228]). In these manuscript, a quadratic performance
criterion of the form
∞
2
J(u) = x1 (t) + ρu2 (t) dt
0
was generally used on linear internal models (e.g. Ackerman’s), where:

• x1 (t) = glucose level above a basal value (i.e. x(t) − xd )
• ρ > 0 = a scalar weighting factor included to adjust the sensitivity of the
control to be achieved. Large values of ρ penalizes the use of large values of
u(t).
• both the glucose level and insulin usage are included in the performance index
to be minimized.
A minimizing controller can generally be found for the linear system above (see
Table 4.7):
Implementation in Literature
• Swan [228], and Fisher & Teo [227]: Both consider Ackerman’s linear model
as internal model,
x˙1 (t) =−m1 x1 (t) − m2 x2 (t) + p(t)

x˙2 (t) = m4 x1 (t) − m3 x2 (t) + u(t) (4.9)
where x1 = G and x2 = H.
However, Fisher & Teo lumped the two equations into one single equation
g̈1 + (m1 + m3 )ġ1 + m1 m3 x1 = −m2 u(t)
with initial condition g1 (0) = x10 and ġ1 (0) = B − m1 x10 and B = meal size
[mg/dl per min]. The control problem is then solved as per usual method.
One observation from the solution to the glucose-control problem is that it is
possible for u(t) to become negative (i.e. the infusion of “negative” insulin). If
we seek to infuse insulin only (and not be supplemented by glucose infusion to
counter the previously given insulin), then the solution to the linear quadratic
optimal control problem would not be suitable. Fisher & Teo suggested a “sub-
optimal” control regime, where an extra constraint u(t) ≥ 0, ∀t ≥ 0 to address
this issue. Alternatively, the performance criterion can be re-formulated
∞
J(u) = x21 (t) dt
0
subject to 0 ≤ u(t) ≤ umax , ∀t ≥ 0 (which could not be solved analytically

[227]).
Table 4.7. Solving classical Optimal control problem
Solving Classical Optimal Control Problem
Given the system equation ẋ = Ax + Bu and performance index

∞
J= [x∗ Nx + u∗ Ru] dt
0
where N, R are both positive-definite Hermitian or real symmetric matrix.

Let u(t) = −Kx(t) where matrix K contains the optimal control vector,
then ẋ = Ax + Bu = (A − BK) x and
∞ ∞
J= [x∗ Qx + u∗ Ru] dt = x∗ [Q + K∗ RK] xdt
0 0
Choose a possible Liapunov function V (x) = x∗ Px where P is a positive

definite Hermitian matrix. Since V(x) is chosen to be a positive definite,
then V̇ (x) would be negative definite (for stability), V̇ (x) = −x∗ Nx. Note
that:
V̇ (x) = ẋ∗ Px + x∗ Pẋ

= ((A − BK) x)∗ Px + x∗ P (A − BK) x
= x∗ [(A − BK)∗ P + P (A − BK)] x
= −x∗ Nx
Applying this to the performance index J, we obtained
(A − BK)∗ P + P (A − BK) = − [Q + K∗ RK] (4.8)
Since R is assumed to be a positive-definite Hermitian or real symmetric

matrix, we can write R = T∗ T and re-write equation 4.8 to be
A∗ P + PA − B∗ K∗ P − PBK + [Q + K∗ T∗ TK] = 0
∗
1 ∗ 1 ∗
A∗ P + PA + TK − B P TK − B P − PBR−1 B∗ P + Q = 0
T∗ T∗
For the equation above to be the minimum requires TK − 1

T∗
B∗ P = 0, or
1 1 ∗
K= B P
T T∗
1
= B∗ P
R
Thus, the optimal control law is u(t) = −Kx(t) = −R−1 B∗ Px(t).
The optimal control problem can also be solved using a Hamiltonian (See
Gopal [229]).
4.5.3 Adaptive Control
In adaptive control of glucose level, a glucose-insulin model is used as internal

model, and the parameters of the model are first estimated using adaptive al-
gorithms. Then other mathematical techniques (including optimal control) are
applied onto this model that has the estimated parameters, for the eventual
calculation of the insulin infusion rate. Fig 4.7 shows the general arrangement.
Fig. 4.7. Adaptive control general arrangement
The adaptive method usually considered in the literature is the (recursive)

least square estimation algorithm.
Least Square Estimation Algorithm (from [230])
Consider an input samples from a time series

T
x(k) = [x(k), x(k − 1), · · · , x(k − p)] , and the coefficients of a time-varying
T
Finite Impulse Response filter w(n) = [w0 (n), w1 (n), · · · , wp (n)] . Together,
they represent the output of a system as y(k) = w (n) · x(k) at time k.
T
If d(k) is the desired output at time k, then the error signals is
e(n, k) = d(k) − y(k) = d(k) − wT (n) · x(k)

By defining a forgetting factor β(n, k) to reduce the influence of old data
0 < β(n, k) ≤ 1, k = 1, 2, . . . , n
the error can be redefined as
ξ(n) = ξ (w0 (n), w1 (n), · · · , wp (n))

n
2
= β(n, k) |e(n, k)|
k=0
n
2
= β(n, k) d(k) − wT (n) · x(k)
k=0
In least-square estimation, the goal is to find the optimal filter coefficients that
minimises the error ξ(n). This is done by setting the derivative of ξ(n) with
∗
respect to wm (n) to zero:
∂ξ(n) n ∗
= k=0 β(n, k)e(n, k) ∂e∂w(n,k)
∗ =0
∂w∗
It is common to use β(n, k) that has the form
β(n, k) = λn−k , k = 1, 2, . . . , n; 0<λ<1
It can be shown (see Appendix A) that the optimal filter coefficients can be
found using a recursion technique, which begins by
1. setting wT (n − 1) = 0 and P(0) = δ −1 I where δ is a small positive constant
2. computing
λ−1 P(n − 1)x∗ (n)

k(n) =
1 + λ−1 xT (n)P(n − 1)x∗

w(n) = w(n − 1) + d(n) − xT (n)w(n − 1) k(n)
P(n) = λ−1 P(n − 1) − λ−1 k(n)xT (n)P(n − 1) (4.10)
3. Repeat until w(n) converges or continue indefinitely.
• Pagurek et al [231] used a generalised least square algorithm to estimate

parameters in Ackerman’s linear model (i.e. m1 to m4 ) on-line, with signal pre-
filtering. The recursive estimation of the parameters followed that of system
4.10, with the filter coefficients derived from a difference equation y(n) +
a1 y(n − 1) + a2 y(n − 2) = b1 u(n − k − 1) + b2 u(n − k − 2) + e(n) relating
to Ackerman’s model. The model, and a disturbance model (assumed to be
samples of stationary independent Gaussian zero-mean sequences) are then
used in a minimum variance control strategy.
• Kikuchi et al [232] used recursive least square estimation to estimate the pa-
rameters in the Ackerman’s model (i.e. m1 to m4 ). The recursive estimation
followed that of system 4.10 but with modification to include an extra term
“−(1 − m1 T)X1 (n − 2)” in the term d(n) − xT (n)w(n − 1) of system 4.10
above. X1 was defined to be the equivalent of glucose concentration x1 in
the Ackerman’s equations. The filter coefficients were derived from a differ-
ence equation relating to Ackerman’s model. A state observer was then used
to estimate the insulin state variable before calculating the insulin infusion
rate using optimal control strategy. The state observer is used because only
the glucose (first state variable) and not the insulin (the second variable)
in Ackerman’s model is directly measured in real-time. Due to the delay in
BG measurement from the glucose analyser, a state-predictor was included to
deduce the present time BG (based on the BG taken some minutes ago).
• Sarti et al [233] devised a self-tuning adaptive controller based on a discrete-
time linear glucose-insulin model P described by:
Gk = P (Gk−1 , . . . , Gk−m , IDk−1 , . . . , IDk−n , Θ)
where Gk = BG level at time k; IDk = insulin dose at time k ; Θ = a set

of unknown parameters; and m, n = known time delays. The parameter set
is recursively estimated at each time k based on Gk measurements to pro-
duce Gk+1 (a one-step-ahead prediction). An optimal insulin dose, IDk+1 is
calculated by minimizing a cost function J, which in the case of a minimum
variance controller, is
2
Jk = (Gk+1 − Gdesired ) − r · ID2k+1
with r being a weighting factor that penalises insulin dosage.

• Fischer et al [214] devised an adaptive controller with system identification.
System identification was based upon a linear second-order difference equation
model describing the actual state of the controlled plant of the glucose-insulin
system (in terms of blood glucose concentration BG, and exogenous insulin
dose ID) at discrete intervals k. A recursive least square regression analysis
was applied to estimate the model parameters of this internal model. The
performance index was

Σ e2 (k) − 0.003ID2 (k)
where e(k) = BG(k) − BGsetpoint . The insulin dose is then given by
ID(k + 1) = a∗1 · e(k) + a∗2 · e(k − 1) − a∗3 · ID(k − 1)
where a∗i = estimated control constants.

• Candas & Radziuk [199] used an adaptive control based on a non-linear
insulin-glucose model, and instead of infusing insulin only, glucose infusion
is also considered. The approach counters the glucose removal (due to exter-
nally infused insulin) with an exogenous infusion of glucose. The externally
infused insulin is kept at a constant rate, and blood glucose concentration is

kept at a desired level by means of controlling an external infusion of glucose
(glucose-clamp experiment). The process parameters on the non-linear model
are recursively adjusted (via generalized-least-square method) after each new
observation of the glucose concentration. Control is then achieved by using the
nonlinear glucose kinetics model equation in an inverse fashion to calculate
the glucose infusion required to maintain a preset profile for glucose.
4.5.4 Model Predictive Control
The MPC control law can be understood by referring to Fig 4.8. For any assumed
set of present and future control moves Δu(k), Δu(k +1), . . . , Δu(k +m−1) the
future behaviour of the process outputs y(k + 1|k), y(k + 2|k), . . ., y(k + p|k) can
Fig. 4.8. Model Predictive control illustration (Adapted from [234])
be predicted over a horizon p. The m present and future control moves (m ≤ p)

are calculated to minimize a quadratic objective of the form

p
m
min Γ y [y(k + n|k) − r(k + n)]2 + Γ u [Δu(k + n − 1)]2
Δu(k)...Δu(k+m−1)
n=1 n=1
where y(k + n|k) = predicted value of y at time k+n based on information

available at time k ; Δu(k + n) = u(k + n) − u(k + n − 1); Γ y and Γ u = weighting
matrices to penalise certain components of y or u at certain future time intervals;
r(k + n) = vector of future reference values (setpoints).
Although m control moves are calculated, only the first one (i.e. Δu(k)) is
implemented. At the next sampling interval, new values of the measured output
are obtained, the control horizon is shifted forward by one step, and the same
calculations are repeated. The resulting control law is referred to as “moving
horizon” or “receding horizon”. The predicted process outputs y(k + 1|k), . . .,

y(k + p|k) depend on
• the current measurement ŷ(k), and
• the assumptions made about the unmeasured disturbances and measurement
noise affecting the outputs (see [234]).
In MPC, the goal is to find the best value for the manipulated variable u(t).
More details about MPC and the inclusion of disturbance into the calculation
can be found in [235] and [234].
• Parker et al [207] used linear Model Predictive Control with state estimation
and Kalman filtering. The internal model is
x(k + 1) = Φx(k) + Γ u(k)

y(k) = Cx(k)
The state x̂ and output ŷ of the plant can be estimated using
x̂(k + 1) = Φx̂(k) + Γ u(k) + κ[y(k) − C x̂(k)]

ŷ(k) = C x̂(k)
where the Kalman filter κ is calculated iteratively off-line:

1
κ(i) = ΦP (i)C T
CP (i)C T + R2
P (i + 1) = ΦP (i)ΦT + R1 − κ(i)CP (i)ΦT
Tuning of the Kalman filter is accomplished using the matrices R1 , R2 and P(0).
The internal model is obtained from the continuous nonlinear diabetic pa-
tient model by first linearising the nonlinear model analytically to produce a
linear continuous-time model. The linear model is then converted to a discrete-
time minimum-phase model for use in simulation, yielding Φ, Γ and C.
• Lynch et al [223] used a similar concept except that Bergman’s model is
discretized (instead of linearising the comprehensive model), and that subcu-
taneous glucose measurement is used (rather than the arterial glucose mea-
surement in Parker’s case).
• Trajanoski et al [225] devised a “Neural Predictive Controller”, which is based
on off-line identification of the glucoregulatory system using neural network,
amalgamated with a nonlinear model predictive controller (nMPC) design using
multiple step-ahead predictions of the previously identified model. The nMPC
uses moving horizon approach with a whole set of predictions (instead of the
classical approach where one prediction is used). The performance index was:
min J = Σ pj=N1 eTj Γe ej + Σj=0

m−1
ΔuT (t + j) · Γu · Δu(t + j)
u(t),u(t+1)...u(t+m−1)
where ej = deviation of the predicted output ŷ(t+j) from the setpoints r(t+j);
Δu = u(t)−u(t−1) is the difference in the control moves; m = control horizon;
p = prediction horizon; N1 = minimum costing horizon; Γe , Γu = prediction
and control weighting respectively. Bellazi et al [236] explained this approach
from a different perspective.
• Hovorka et al [211] have developed a non-linear model predictive controller
with on-line parameter estimation using Bayesian approach. The model pa-
f f f
rameters SIT , SID , SIE , F01 and EGP0 (see Section 4.4.7) are tuned using
2
the objective function
Nw 6
arg min wt−i [ŷ (t − i|p1 · · · p6 ) − y (t − i)]2 + p2k
−2.5≤p1 ...p6 ≤2.5
i=1 k=1
where wi = weight reciprocal to the square of the measurement error (co-

efficient of variation of 2–9% depending on the source of glucose sample);
ŷ (t − i|p1 · · · p6 ) = glucose concentration predicted by the model at time t
given the standardised parameters p1 · · · p6 ; Nw = length of the learning win-
dow; y (t − i) = previously measured glucose concentration.
The parameters p1 · · · p6 are independent normal distributions with a zero
mean and a unit standard deviation (i.e. p1 ∼ N (0, 1), i = 1, . . . , 5), and they
are related to the glucoregulatory model by
f
ln SIT = a11 p1 + b1
f
ln SID = a12 p1 + a22 p2 + b2
f
ln SIE = a13 p1 + a23 p2 + a33 p3 + b3
ln (F01 ) = a14 p1 + a24 p2 + a34 p3 + a44 p4 + b4
ln (EGP0 ) = a15 p1 + a25 p2 + a35 p3 + a45 p4 + a55 p5 + b5
with the coefficients aij and bi derived from Random Variable Transformation
technique. The insulin infusion rates are calculated by minimising an objective
function over a prediction horizon N :

N
2
arg min [ŷ (t + i|t) − y (t + i)]
0≤u(t+1)···u(t+N )≤4
i=1

1
N
2
+ [u (t + i) − u (t + i − 1)]
kagr i=1
where ŷ (t + i|t) = predicted glucose; y (t + i) = target (or desired) trajec-

tory of glucose associated with high and low glucose concentration; kagr =
an “aggressive” constant balancing the contribution of the two summation
terms above. An example of the target trajectory that the system would fol-
low would be a BG curve that linearly decreases at pre-defined rates per
2
Notation: arg minx f (x) is the value of x for which f (x) has the minimum value.
hour (see Fig 4 in [211]). A sequence of insulin rates u(t + 1), . . . , u(t + N )
are generated, but only the first one, u(t + 1), is delivered to the patient.
The insulin rate was limited to 4 U/hr due to the limitation of the infusion
pump used (see [211]). Note that in clinical trials involving human subjects,
blood glucose is measured intravenously (i.e. accurate BG values were ob-
tained and fed into the controller) while insulin are given via a subcutaneous
route.
4.5.5 H∞ Control
H∞ refers to the space of all bounded functions that are analytic and stable
(i.e. poles are in the Right-hand-plane), and have proper transfer functions (i.e.
degree of the denominator ≥ the degree of the numerator) (see [237]). Consider
a stable single-input-single-output (SISO) linear system with transfer function
G(s), the H∞ norm is defined as
G∞ = sup |G(jω)|

ω
The interest is to optimise over the space of transfer function. Graphically, it is

simply the peak in the Bode magnitude plot of the transfer function. H∞ norm
is a measure of a worst-case system gain.
Table 4.8. L2 -norm and L∞ -norm
Definition: L2 -norm & L∞ -Norm
For a scalar-valued signal v(t) defined for t ≥ 0 the L2 -norm is defined as

the square root of the integral of v(t)2 :
∞ 1/2
v2 = v(t)2 dt
0
The L2 -norm is a special case of Lp -norm, defined as

∞ 1/p
vp = v(t)p dt , p≥1
0
As p → ∞, the Lp -norm tends to the so-called ∞-norm, or L∞ -norm, which

is the least upper bound (or supremum) of the absolute value, i.e..
v∞ = sup |v(t)|

t
The supremum is used in place of maximum is because there is no guarantee

that a maximum exists for v(t) in general.
Table 4.9. Packed-Matrix Notation
Definition: Pack-Matrix Notation (See [238])
The transfer function of a system with state-space matrices [A, B, C, D] is

given by
G(s) = C(sI − A)−1 + D
This transfer function in packed-matrix form is written as
AB
G(s) =
C D
Note that the packed-matrix form is the frequency domain transfer function
expressed like a time-domain matrix, i.e.
AB
G(s) =
C D
= C(sI − A)−1 + D
= s − domain transfer function
Consider the standard configuration shown in Fig 4.9. Plant P has:

• Two inputs:
1. w = the exogenous input, which includes the reference signal and distur-
bances;
2. u = manipulated variables.
Fig. 4.9. H∞ standard configuration (Adapted from [239, 240])

• Two outputs:
1. z = error signals (which is what we wanted to minimise);
2. y = measured variables (which is used in the controller K to calculate the
manipulated variable u, so as to control the system).
P and K are matrices while the input & output signals are generally vectors.
The system can be described by:
⎡ ⎤
reference signal
tracking error
= P ⎣ disturbances ⎦
control output
control input
Let the SISO model P has a state-space system of the form
dx(t)
= Ax(t) + B1 w(t) + B2 u(t)
dt
z(t) = C1 x(t) + D11 w(t) + D12 u(t)
y(t) = C2 x(t) + D21 w(t) + D22 u(t)
Some algebraic manipulation:
sx = Ax+B1 w + B2 u ⇒(sI − A)x = B1 w + B2 u

z = C1 x + D11 w + D12 u
y = C2 x + D21 w + D22 u
1 1
and substituting x = sI−A B1 w + sI−A B2 u into the second and third equation,
we obtain

1 1
z = C1 B1 + D11 w + C1 B2 + D12 u
(sI − A) (sI − A)

1 1
y = C2 B1 + D21 w + C2 B2 + D22 u
(sI − A) (sI − A)
or

z P11 (s) P12 (s) w
= (4.11)
y P21 (s) P22 (s) u

D11 D12 C1 −1 w
= + (sI − A) B1 B2
D21 D22 C2 u
⎡ ⎤
A B1 B2
w
= ⎣ C1 D11 D12 ⎦ by definition of packed matrix notation
u
C2 D21 D22
If a feedback controller of the form u = K(s)y is introduced, the equations can

be transformed as follows
z = P11 w + P12 u and y = P21 w + P22 u

= P11 w + P12 Ky = P21 w + P22 Ky
(I − P22 K) y = P21 w
−1
⇒ y = (I − P22 K) P21 w
We can express the transfer function of w-to-z as
z = F (P, K) · w
where
1
F (P, K) = P11 + P21 K P21
(I − P22 K)
The expression for the closed-loop transfer function F (P, K) is called Linear
Fractional Transformation. The objective of H∞ control design is to find a con-
troller K that minimises the cost
J∞ (K) = F (P, K)∞

= sup σ̄ (F (P, K) (jω))
ω

z
= sup σ̄
ω w
The direct minimisation of the cost J∞ (K) turns out to be difficult. It is
much easier (and practical) to construct conditions which state whether there
exists a stabilising controller such that the H∞ norm is bounded
J∞ (K) < γ
for a given γ > 0.

One interesting properties of H∞ norm is that it gives the maximum factor
by which a system magnifies the L2 norm of any input [241],
!
z2
J∞ (K) = sup : w = 0 < γ (4.12)
ω w2
2 2
Let L(w, u) = z2 − γ 2 w2 < 0, ∀w = 0, then by Parseval’s theorem, the
time-domain expression of the L2 norm becomes
∞

L(w, u) = z(t)T z(t) − γ 2 w(t)T w(t) dt < 0, ∀w = 0
0
The problem of finding a controller u = Ky can be stated in terms of a max-min

problem !
sup inf L(w, u) < 0
w=0 u=Ky
Put simply, w (player 1) will try to make the cost L(u, w) as large as possible,
while u (player 2) attempts to find the smallest L(u, w) < 0.
Certain assumptions must be satisfied for the H∞ controller to exist [237,238]:

1. The pair (A, B2 ) be controllable and (C2 , A) be detectable
2. Rank D
12 = dimension of u, and rank D21 = dimension of y
A − jωI B2
3. Rank = dimension of x + dimension of u
C1 D12
A − jωI B1
4. Rank = dimension of x + dimension of y
C2 D21
A procedure called γ-iteration is used to find a stabilising controller: Start with
a value for γ and reduce it until the problem failed to have a solution [237].
The design of linear H∞ controller requires a linear system model. From
literature survey, controller design used:
1. linear glucose-insulin models, or
2. linearised models, where
• linearisation techniques were used to transform a non-linear model into a
linear one, or
• model reduction technique was used to generate a low-order process model
for use in controller synthesis.
• Kienitz et al [242] designed a full-observer/state-feedback-type H∞ controller

by first using a slightly extended linear Ackerman’s model:
dx
= Ag x(t) + B1g p(t) + B2g u(t)
dt
y(t) = Cg x(t)
where p(t) = the absorption of glucose from blood; u(t) = infused insulin;
T
x= GI C is the state vector that has an extra component C ; G, I and C
represent the glucose concentration, insulin concentration and the aggregate
of glucagon, cortisol and catecholamine in the blood, respectively. The system
above may also be written as

y(s) = Cg (sI − A)−1 g B2g u(s) + Cg (sI − A)
−1
B1g p(s)
= G1 (s)u(s) + G2 (s)p(s)
= P22 (s)u(s) + P21 (s)w(s) by analogy to equation 4.11
We can see that u(s) and p(s) is analogous to u and w, just as G1 and G2
being analogous to P21 and P22 in equation 4.11. G1 can be seen as the patient
model with the insulin infusion, and G2 being the meal disturbance model.
Parameter uncertainties in Ag were then modelled by additive uncertainties
G1 (s) = G10 (s) + G1 (s)

G2 (s) = G20 (s) + G2 (s)
Table 4.10. Robustness and Uncertainty modelling
Robustness and Uncertainty Modelling
A robust control system is one that can withstand variations in real envi-
ronment and operate properly in realistic situations [240]. Control systems
are usually designed using simplified models of the system and environment,
and thus may not work on the real plant in real environment. A robust con-
troller must perform satisfactorily not just for one plant, but for a family
of plants.
Model uncertainty is generally divided into structured uncertainty and un-
structured uncertainty. Structured uncertainty assumes that the uncertainty
is modelled and there are bounds for uncertain parameters in the system.
Unstructured uncertainties assume less knowledge of the system, and as-
sumed only that the frequency response of the system lies between two
bounds. Two most common unstructured uncertainties are:
• Additive, where
actual system = modelled system + model error

G(s) = G(s) + Δa (s)
The model error, or the additive uncertainty, is given by
Δa (s) = G(s) − G(s)
• multiplicative, where
G(s) = G(s) + Δm (s)G(s)

= [1 + Δm (s)] G(s)
The model error, or the multiplicative uncertainty, is given by
G(s) − G(s)
Δm (s) =
G(s)
The multiplicative model is used more often because it represents the rel-
ative error in the model, rather than the absolute error as represented by
the additive model.
For convenience, the characterisation of an uncertain plant is usually stated

in an equivalent form obtained by defining a quantity “Δ” such that
Δa = Wa Δ for additive uncertainty, and Δm = Wm Δ for multiplicative
uncertainty [240]. The uncertainty block Δ is assumed to be uniformly
bounded at all frequencies, and the varying amount of plant uncertainty
at various frequencies is completely incorporated into the filter Wn . The
filter Wn is commonly called “uncertainty weighting filter”.
The design was refined by adding a first-order low pass filter and weights to
the exogenous input p, and weights to the uncertainty input and output. The
solution algorithm was based directly on those in [238].
Fig. 4.10. H∞ example – Adapted from Kienitz et al [242]
• Parker et al [207] used a detailed nonlinear patient model in the development

of a continuous-time H∞ controller, in contrast to Kienitz et al who synthesize
the controller from an empirical low-order linear patient model. A non-linear
pharmacokinetic/pharmacodynamic model (Sorensen) was first linearised via
Jacobian linearisation and then reduced to a third-order linear form by bal-
anced truncation. The reduced linear model was given in state-space form:
ẋred = Ared xred + Bred u + Bm md

y = Cred xred
and could be split into a process model (G), and a disturbance model (Gm )
representing the effects of the meal disturbance (md ), in similar fashion to
those in Kienitz et al (above).
Quoted from [207], ”Uncertainty due to differences between an actual patient
and the diabetic patient model was thought to be related to variations in
model parameters.” Sets of three parameters from a possible of eight physi-
ological parameters – five parameters from liver, and three parameters from
the peripheral (muscle/fat and tissues) – were chosen for uncertainty charac-
terization, yielding 56 combinations. With each parameter set varied in five
permutations (+max, +1/2 max, no change, -1/2 max, -max) about the nom-
inal value, there are a total of 125 variations possible in each three-parameter
set (35 ), or 7000 possible perturbed models (56 × 35 ). The nonlinear model
was linearised around each of the 125 parameter variations for each set. The
most sensitive parameter set was identified by summing the multiplicative

uncertainty
Gperturbed (ω) − G(ω)

Urel (ω) =
G(ω)

in the frequency range of interest ω = 0.002 0.2 (rad/min) for each pertur-
bation, and then summing over each parameter set. A similar analysis was
performed for the uncertainty in the meal disturbance model (Gm ).
The controller design was completed by constructing the weighting functions:
Wu (input weight), Wm (meal disturbance weight), Wn (noise weight), Wp (per-
formance weight), Wi (input multiplicative uncertainty for perturbed patient
model), and Wim (input multiplicative uncertainty for perturbed meal distur-
bance model). The weights Wi and Wim were calculated as the least upper
bound on the relative uncertainty of the model respectively.
Fig. 4.11. H∞ example – Adapted from Parker et al [207]
• Ruiz-Velázquez et al [208] had focused on the synthesis of a controller to

“compute the intravenous infusion from the subcutaneous measurements of
BG concentration such that the glucose level tracks the curve of non-diabetic
subjects by accounting”. A non-linear pharmacokinetic/ pharmacodynamic
compartmental model (Sorensen) was linearised around euglycaemic condition
(at 80 mg/dl, with insulin at a constant rate of 22 mU/min). The linearisation
of the nonlinear system provided a nominal model around nominal parameters
values. The resulting generalized plant P (s) is then obtained by balanced
truncation:
z1 d1 Wp Gm Wm 0 Wp G d1
z2 = P (s) d2 = 0 0 Wu d2
y u −Wm Gm −Wn −G u
where G = transfer function of the reduced linear model; Gm = transfer func-

tion of the reduced linear model for the meal; Wp = performance weight,
chosen such that the frequency content of Gm is captured; Wm = the effect of
the meal model; Wn = sensor noise effect; Wu = weight for the control input;
d1 = meal input; and d2 = sensor noise input.
In designing a controller robust for BG regulation, the uncertainty descrip-
tion was computed when parameter variations were incorporated into the
nonlinear model. A family of plants was obtained by adding parametric vari-
ations into the non-linear model, and then every one of the non-linear models
was linearised at the same euglycaemic condition. A multiplicative uncertainty
was used:
" = [I + Wip Δ] G
G
where Δ∞ ≤ 1, G is the nominal plant, and Wip is a weighting function.
The transfer function Wip was identified via the maximum in the frequency
response of the set of multiplicative uncertainty on the nominal plant G:

G G(jω) − G(jω)
" − G "
Wip Δ = ⇒ Wip (jω) = max
G G(jω)
The same procedure was followed to identify a multiplicative uncertainty for

the input disturbance (meal), Wim . The generalised plant, P (s) became a
fifth-order model:
z1 d1 Wm Gm Wp 0 Wn G Gm Wp Wp G d1
z2 d2 0 0 0 0 Wu d2
z3 = P (s) d3 = 0 0 0 0 Wip d3
z4 d4 Wm Wim 0 0 0 0 d4
y u −Wm Gm −Wn −G −Gm −G u
The synthesis problem is then to find a controller that minimises z =

[z1 , z2 , z3 , z4 , y]T in the H∞ sense by the classical technique and LMI tech-
nique.
• Chee et al [243] designed a “switching controller” based on a non-standard
H∞ control theory [244–246]. The internal model is based on the Minimal
model for diabetic state, but linearised at a nominal glucose value. The inter-
nal model has the general structure:
x(t) = A(t)x(t) + B1 (t)ξ(t) + B2 (t)u(t)

zc (t) = Kc (t)x(t), t = iT
zd (iT ) = Kd (iT )x(iT ), i = 0, 1, 2, . . .
y(iT ) = Ci x(iT ) + vi , i = 0, 1, 2, . . .
where x(t) = state vector; ξ(t) = external glucose disturbance input; u(t) =
{0, U0 , U1 , . . . } = a sequence of (pre-defined) insulin infusion rate; zc (t) = con-
tinuous controlled output; zd (iT ) = discrete controlled output; y(iT ) = dis-
crete BG measurement; vi = noise. The control strategy is a rule for switching
Fig. 4.12. H∞ example – Adapted from Ruiz-Velázquez et al [208]
from one sequences of insulin rates to another. A performance index similar to

that of equation 4.12 is used, and the related Riccati equation is derived using
the standard H∞ method initially. Bertsekas et al’s method [247] is then used
to transformed the performance index into a form that does not depend on
ξ(t). To find the optimum insulin infusion sequence, dynamic programming is
used to calculate the performance index associated with each possibilities of
u(t) sequence with arbitrary y(iT ).
4.5.6 Note
It should be noted that most control algorithms described above use BG level
measured at i.v. catheter or subcutaneous regions. Insulin are delivered either
intravenously or subcutaneously, and not directly at the portal circulation. This
is partly because it is safer to access blood vessels at the peripheral and subcu-
taneous site, than at the portal site, despite the fact that insulin delivered at
peripheral site usually causes hyperinsulinemia, and adverse effects have been
reported (e.g. an increased risk of developing Alzheimer disease [154, 155]).
4.6 Conclusion
The design of effective control algorithms require a knowledge of the characteris-
tics of the individual components that constitute the loop (i.e. the sensor, insulin
dynamics etc). Furthermore, practical implementation also needs to address the
issue of noise in measurement and uncertainties in the model. The use of full
and comprehensive glucose-insulin description requires a knowledge of all the
interdependent physiological processes involved, and also on the accessibility of
the parameters involved. Some parameters are not easily obtainable with simple
tests, and may not be measurable due to technical barriers, ethical issues, or
4.7 Summary 107
Dynamics of different
Dynamics of different
patch ?
Intra−peritoneal
Infusion Routes
(e.g. slow−acting
Subcutaneous
infusion route
Intravenous
Transdermal
insulin, etc)
insulin type
Algorithm
Control
resistance, sensitivity etc)

to insulin (type, route
Clouds of uncertainty
Patient’s response
Measurement
frequency,
delay
BG measurement
BG measurement
site dynamics
Non−invasive
Minimally−
Invasive
invasive
Fig. 4.13. Map of inter-relationship in a closed-loop insulin delivery system for BG

control
cost-benefit issues. Having said, there has been effort in developing algorithms
that uses “average” model parameter values as a start, and then “personalized”
as data are gathered (e.g. Hovorka et al).
The application of mathematical modelling and control theory certainly
helped in the understanding of the glucose-insulin kinetic and the subsequent
control problem, bringing us one-step closer to achieving better glucose regula-
tion. What remains missing is a reliable, accurate and minimally/non-invasive
glucose sensor.
4.7 Summary
This chapter examines the two approaches to glucose controller design. Model-
less approach relies heavily on rules or empirical method to devise the control
algorithm. In contrast, model-based approach attempts to understand, and

mathematically “capture” the underlying system, before devising the control
strategy. When the underlying system is sufficiently captured in the form of
mathematical equations, glucose-control problem can be viewed as mathematical
problem, and solved using standard/state-of-the-art mathematical techniques.
Mathematical models of the glucose-insulin system ranges from simpler linear
models, through to non-linear and comprehensive models. Many control strate-
gies have been proposed, in various degrees of complexity. This chapter then
briefly described some of the well-known implementations as found in the lit-
erature. To-date, knowledge of the glucose-insulin system continues to evolve,
and new control strategies are being proposed constantly. The unavailability of
a reliable and accurate glucose sensor remains a hindrance to the solving of
glucose-control problem.
5
Closed-Loop Control Apparatus Example
MiniMed CGMS is, perhaps, the first commercially-available subcutaneous

glucose sensor for use by the masses. CGMS consists of a disposable subcuta-
neous glucose sensor assembly connected by a cable to a pager-sized glucose
monitor. CGMS takes a glucose measurement every 10 seconds and stores a
smoothed and filtered average of these values every 5 minutes in the memory
bank on-board.
To view the BG level readings, the sensor signal collected by CGMS must
first be downloaded to a PC using MiniMed Com-Station via RS232-C serial
communication (Fig 5.1).
In this chapter, we show how to:
• Construct a bedside closed-loop glucose control apparatus (with a basic con-
trol algorithm) using CGMS and other off-the-shelf components;
• Create a portable closed-loop system by pairing the subcutaneous glucose
sensor assembly with a syringe pump and microprocessor.
5.1 CGMS Integration

With PC interface included in CGMS, it is possible to connect CGMS to a
personal computer (PC), and adding a bedside infusion pump to effect a closed-
loop glucose control apparatus. In this section, a “basic” bedside closed-loop
insulin infusion system is described. Although the system uses intravenous insulin
infusion, it is not difficult to expand the work to use subcutaneous infusion pump.
5.1.1 Hardware Integration
The three components of the closed-loop insulin delivery system are:

• personal computer, running an operating system of your choice;
• MiniMed-Medtronic CGMS connected to the PC through RS232-C serial ca-
ble;
• Computer-controllable medical infusion pump.
F. Chee & T. Fernando: Closed-Loop Control of Blood Glucose, LNCIS 368, pp. 109–126, 2007.
springerlink.com c Springer-Verlag Berlin Heidelberg 2007
110 5 Closed-Loop Control Apparatus Example
CGMS cable
CGMS
monitor
CGMS
Sensor assembly
Com-Station
Fig. 5.1. CGMS hardware
As a real-life example, we used a PC with Intel Pentium III 800MHz running

Microsoft Windows 98SE (which was popular in year 2000), and an IMED
PC-1 peristaltic pump (Alaris Medical System, San Diego CA) that can be
controlled from a PC through RS232-C serial null-modem cable (Fig 5.2).
5.1.2 Software Integration
Once the hardware is in place, the major functions of the closed-loop system are
managed by software. The software performs four logical functions:
• Sensor data collection
• Pump driving
• Control algorithm computing
• Alert reporting (alarming)
Sensor Data Collection
CGMS is approved for investigational use by the FDA (Food and Drug Admin-
istration, USA), but is not designed to give real-time glucose information (at the
5.1 CGMS Integration 111
time of writing). Each sensor is meant to be worn for up to three days, with the
patient performing SMBG (i.e. Self-Monitoring of Blood Glucose) at least four
times a day and entering the SMBG result into the monitor. When initiated,
MiniMed Com-Station Software V1.6 downloads the “raw data” (i.e. current,
voltage from the sensor assembly) that are stored onboard CGMS, translating it
and then storing it as plain text data file on the PC. MiniMed’s Solution Software
V1.1A then formats this text data file, and applies Regression Calibration to
estimate BG readings obtained during the three days retrospectively.
Insulin dosage signal Computer calculates insulin dose

sent to infusion pump based on measured BSL
Automated closed-loop control system
Insulin infused
to patient Patient’s BSL
Fig. 5.2. Closed-loop insulin delivery system components
To use CGMS in real-time for closed-loop control purposes would required:

• sensor ”raw data” readings to be (automatically) downloaded every 5 min,
and processed without using MiniMed’s Solution Software V1.1A;
This can be achieved by simulating the mouse-clicks and keyboard key presses
using software.
• a modification to the regression calibration method to allow real-time estima-
tion of BG.
Regression calibration was implemented in MiniMed’s Solution Software
V1.1A as an off-line method, whereby BG readings were analysed in a retro-
spective manner (see Section 2.5.2). To obtain BG readings on-line, a quick
way would be to use multi-point in-vivo calibration, but with the calibration
conducted in real-time.
Regression calibration was applied whenever the sensor readings deviated

from the reference glucometer BG reading by a pre-determined amount (e.g.
3 mmol/l). The glucometer value was taken as the calibration BG entry.
Pump Driving
The IMED Gemini PC-1 peristaltic pump uses a C2 Communication Protocol

(Alaris Medical System, San Diego CA). C2 Protocol defines how communi-
cations can be achieved between a computer and a pump through a RS232-C
serial interface by means of pre-defined command set. The software needs to
generates/interprets C2 protocol in order to control the pump.
Control Algorithm
The control algorithm translates BG levels readings into matching insulin de-
livery rates to be given to the patients. As a basic control apparatus, a sliding
table control algorithm is used to serve as an example. The basic sliding table
has the form shown in Table 5.1, and features a commonly used starting scale
for treating hyperglycaemia in critically ill patients at Sir Charles Gairdner Hos-
pital, Perth (where a real-life clinical study was performed using the apparatus).
Insulin delivery is adjusted at the turn of each hour, and maintained at the new
Table 5.1. Basic sliding table, with a region assigned to each BG ranges
BG ranges Insulin infusion rate BG Region

(mmol/l) (U/hr)
>20.0 4 4
15.1 – 20.0 3 3
10.1 – 15.0 2 2
6.1 – 10.0 1 1
0 – 6.0 0 0
delivery rate during the hour. In conjunction with the basic sliding table control
algorithm, the following extra steps were included:
1. Ability to start at an insulin rate higher or lower than those set-out in Table
5.1, by adding a constant Uoffset that adds to the basic set of infusion rates.
This is to address the issues that not all patients have the same insulin re-
quirement as exemplified in Table 5.1. With any selected “starting” infusion
rate, the initial Uoffset would be automatically calculated by
+
Uoffset = (first prescribed rate − BG region)
where the operation (·)+ would return a “-1” if the enclosed term is negative-
valued. Any subsequent insulin will be given according to
Ugiven = BG region + Uoffset
2. Automatic scale-shift every 6 hours.

This is to address the issues that any static sliding table may not be appropri-
ate if a patient’s medical condition changed during the course of treatment.
The scale-shift procedure alters the value of Uoffset in the following manner:
• When there was an upward region transition:

Uoffset (n − 1) + 1, All other upward transition
Uoffset (n) =
Uoffset (n − 1) , From Region 0 to Region 1
• When there ⎧ was no change in BG region:
⎨ Uoffset (n − 1) + 1, Region 2 and above
Uoffset (n) = Uoffset (n − 1) , Region 1
⎩
Uoffset (n − 1) − 1, Region 0
• When there was a downward region transition:

Uoffset (n − 1) , All other downward transition
Uoffset (n) =
Uoffset (n − 1) − 1, From Region 1 to Region 0
3. Rules to immediately reduce, or turn-off insulin delivery when BG fell below
a pre-defined threshold (rather than waiting until the start of the next hour).
This is to safeguard against unwanted hypoglycaemic episodes and protect

the patients. The rules are:
• On detection of BG rise or fall into 6–7 mmol/l range, adjust insulin rate
as per Ugiven immediately;
• On detection of BG falling below 5.5 mmol/l, set Ugiven = 0 and Uoffset = 0
immediately.
Alert Reporting
The software also needs to report any errors or alarms to the user. This includes
any errors in sensor reading, pump driving or other conditions.
5.1.3 Clinical Trial Example
To show that such a control apparatus actually works, we included a few clinical
results obtained during a clinical study on critically-ill patients (see [167]). Note
that the study was approved by the Institutional Ethics Committee of Sir Charles
Gairdner Hospital, Perth, Western Australia, and the Therapeutics Goods Ad-
ministration (TGA), Canberra, Australia. Informed consent were obtained from
the patients or their legal surrogates prior to conducting the study.
Fig 5.3 shows the best performing glucose sensor from the clinical study. The
sensor BG reading tracked the MediSense 2 glucometer reading with 2.9±9.6%
error (mean±2SD). With such excellent BG readings, the action of the sliding
table control algorithm can be examined. One can see the sliding table slowly
increased the insulin rate to bring BG to near 10 mmol/l at the end of the 24-
hour period. This result showed the classical control performed by sliding table
Patient L00 (Clinical study profile)
Date & Time
Fig. 5.3. Profile for Patient L00. Shown are the glucometer reading (◦), sensor reading
(•), insulin infusion rate (×), Total Parenteral Nutrition (+, in ml/hr), naso-gastric
feed (“−”, Jevity in ml/hr), and sensor calibration (). Jevity is a nasogastric feed
from Abbott Laboratories Inc, USA, and contains 36.5 g of carbohydrate per 237 ml
(8FL Oz).
method. The average BG attained during the trial period was 12.0±3.2 mmol/l
(mean±2SD). The control algorithm was configured to bring BG to the range
6–10 mmol/l, taking a conservative control approach. Fig 5.4 shows the complete
closed-loop infusion system used in the clinical study.
It was observed that the glucose sensor did not always perform as expected.
Some observed sensor performances that are undesirable for use in closed-loop
control included below:
1. Sensor signals (Isig in nA) decreases with time (Fig 5.5):
2. Sensor signals (Isig in nA) delayed by 1 hour with respect to the glucometer
readings (Fig 5.6):
Fig. 5.4. Photographs of the complete system used during clinical studies (top). CGMS
glucose sensor on patient’s upperarm area (bottom).
Fig. 5.5. Sensor Isig decreases with time, when compared to the glucometer read-
ings (◦)
Fig. 5.6. Sensor Isig delayed by 1 hour, with respect to the glucometer readings (◦)
3. Sensor signals (Isig in nA) fluctuating around the glucometer readings, while
decreasing in strength with time (Fig 5.7):
Fig. 5.7. Undulating sensor Isig when compared to the glucometer readings (◦), while
decreasing in strength with time
5.2 Miniaturisation – Glucose Sensing Circuit 117
Considering the many factors that can influence the sensor-glucose dynamics in
the subcutaneous space, it may not be surprising that the sensor signal fluctuated
in all the results obtained. In addition, the clinical study was conducted on the
critically-ill patient population, which has a different physiological condition
compared to healthy individuals. The estimation of BG from the sensor current
signal is dependent on three factors:
1. the linearity between sensor signal and ISF glucose concentration (interac-
tion factor).
2. the linearity between the ISF glucose concentration and blood glucose con-
centration.
3. the calibration factors that translate the current signals into BG estimation.
Among the five clinical studies, Patient L00 (Fig 5.3) exhibited the least changes
in sensor linearity, with the highest change at 0.17 mg/dl per nA, which
amounted to 1.7–17 mg/dl (or 0.09–0.9 mmol/l) for a current range of 10–100 nA.
The example of the bedside system shown in this section showed that a bed-
side closed-loop system can be constructed fairly easily, due to the commercial-
availability of the components – especially the glucose sensor.
5.2 Miniaturisation – Glucose Sensing Circuit
In this section, we show how to integrate the disposable subcutaneous glucose

sensor assembly with a “glucose sensing circuit”. The glucose sensing circuit
interfaces physically with the CGMS glucose sensor assembly (Fig 5.1) through
the CGMS cable, and provides all the necessary electrical environments required
for glucose sensing. To create a portable closed-loop system, the glucose-sensing
circuit is further paired up with a syringe pump and a microprocessor.
This integration is done purely to illustrate how a commercial glucose sensor
can be used for research and educational purposes. We (as well as the FDA
(Food and Drug Administration, USA)) are, in no way, approving this modified
arrangement for use in human subjects.
5.2.1 Background
CGMS uses amperometry as its main measuring technique (see Section 2.1). The
commonly used sensor configurations are:
1. two-electrode system — a platinum (Pt) working-indicating electrode with a
Ag/AgCl reference-counter electrode, such as those described in [16,248,249].
2. three-electrode system — a platinum (Pt) working electrode, an Ag/AgCl
reference electrode, and a platinum (Pt) counter electrode such as those
described in [33, 248, 250, 251].
CGMS uses a three-electrode system. A circuit that can be used with CGMS is
presented here, and it consists of three components:
1. a constant potential application circuit.

2. a nano-ampere measuring circuit.
3. a supply splitting circuit.
5.2.2 Constant Potential Application Circuit
The constant potential application circuit (formed by op-amps 2 and 3 in Fig

5.8) functions to impose a constant potential between two electrodes, for the
hydrolysis of hydrogen peroxide (see Table 2.1 in Chapter 2).
Rbody1 Rbody2
WRK REF CNT

Vout
C1
R=5M
− −
IN4148 1 2
+ +
−V
+V
VCC VEE
(VCC, VEE connected to all op−amps) VR1

150k
+ C2
+ +
5V
− − 4 3 −
LM385
200k
Vgnd
−V
Fig. 5.8. Glucose sensing circuit for integration with MiniMed CGMS glucose sensor
assembly. Figure shows a constant potential application circuit (right), a nano-ampere
measuring circuit (upper left) and a supply potential splitting circuit (bottom left).
The resistances, Rbody1 and Rbody2 represent the flow of the current generated from
the hydrolysis of H2 O2 , across the constant potential applied to the electrodes.
In this design, a negative potential is applied at the reference electrode, REF,

with respect to the working electrode, WRK. The design used two op-amps (i.e.
op-amps 2 and 3) arranged in voltage-follower configuration, and the variable
resistor VR1 tuned to provide the negative potential. Since there should be no
currents flowing into the input terminals of op-amps 2 or 3 (i.e. due to the high
resistance of the terminals), any current generated from the hydrolysis of H2 O2
5.2 Miniaturisation – Glucose Sensing Circuit 119
would flow to the counter electrode, CNT, and then sink to VEE via the output
terminal of op-amp 3. This configuration allows the voltage at REF to remain
constant.
CNT
WRK
REF
Fig. 5.9. Connection to the CGMS sensor assembly (Source: MiniMed’s patent
WO00/74753A1, US6,360,888B1)
CGMS sensor
assembly
Slotted wheel
Motor
Receiver
Emitter
Circuit board + − + −
Fig. 5.10. Illustrative diagram of pump assembly
5.2.3 Nano-ampere Measuring Circuit
The generated current, due to its small magnitude (in nano-ampere range), is
measured by a nano-ampere measuring circuit, which is formed around op-amp
1. The circuit employs a transimpedance topology, and converts the out-flowing
current (with respect to the op-amp’s “-” input) into a positive output voltage.
The output voltage is proportional to the current, and is thus proportional to
the amount of glucose present at the sensor insertion site. The two diodes that
are connected to the op-amps inputs function to limit excessive current from
damaging the op-amp. The capacitor in series&with the 5MΩ resistor forms a
low-pass filter with a cut-off frequency of 1/ 2π · (5M Ω × C1) to eliminate
noise in the measurement.
5.2.4 Supply Potential Splitting Circuit
To permit the use of a single 5V battery, and yet simultaneously allowing the ex-
istence of a negative potential (for the reference electrode), and a positive voltage
output (at the nano-ampere measuring circuit), a supply potential splitting cir-
cuit is used. Formed by op-amp 4, the supply splitting circuit selects a potential
within the source potential range (by means of a LM385 adjustable micropower
voltage reference diode feeding into op-amp 4), and maps it for use as “ground”
reference by the other circuits. This effectively splits the source potential into
two separate potential sources, giving a positive and negative voltage reference.
This eliminated the need for a separate generation of a negative power supply.
The use of op-amps with very low power consumption minimised the effect of
limited current sourcing/sinking ability of the op-amps configuration employed
in the supply splitting circuit.
Used in conjunction with the CGMS sensor assembly and cable, this glucose
sensing circuit allows minute-by-minute measurement of subcutaneous glucose
concentration, to be used by a control algorithm in the insulin delivery device.
5.3 Miniaturisation – Syringe Pump and Microprocessor

The glucose-sensing circuit can be interfaced with a microprocessor and syringe
pump to effect a closed-loop control system.
Vcc
*
R3 R1
Microprocessor
TR1
LCD
PA1
R5 R2
PA2 TR3 R4
TR2 Motor
Buttons’ array M connects to
the shaft of
the slotted
PA3 wheel.
Vout from
glucose sensing ATD
circuit
* indicates slotted wheel
Fig. 5.11. Syringe pump circuit schematic, showing the interface between the motor
driving circuit, the motor rotation detector circuit, with the microprocessor
The use of a microprocessor allowed flexibility in changing functionality (in

the control algorithm). Fig 5.11 shows the circuit design of the microprocessor-
controlled syringe pump, which consists of
1. a motor driving circuit
2. a motor rotation detector circuit
3. a liquid crystal display (LCD) and hardware buttons
4. a microprocessor
5.3.1 Motor Driving Circuit
The motor driving circuit consists of a voltage source with current limiter. The
motor is turned on by a logic high placed on the I/O pins of the microprocessor
that connects to transistor TR1 (i.e. PA3 in Fig 5.11). When the motor drew
5.4 Signal Processing and Control Algorithm Programming 121
more than a pre-defined current determined by the value of resistor R2 (e.g. when
there is a flow resistance in the delivery catheter), TR2 would turn on, limiting
the current to the motor, and protecting the motor windings from overcurrent.
5.3.2 Motor Rotation Detector Circuit
The motor rotation detector circuit consists of a light emitting diode, and a
phototransistor, which can also be replaced by a slotted optical switch assembly.
A slotted wheel, which is mounted on the motor shaft, is located between the
slot of the optical switch assembly. Feedback on the degree of motor rotation
was provided by the motion of the wheel interrupting the light beam between
the slot (see Fig 5.10 and Fig 5.11).
As the optical switch is mainly used to detect the degree of rotation made by
the motor, the microprocessor can be programmed to save power by switching
on the optical switch only when the motor is turned on. The motor and the
optical switch can then be switched off when the motor has rotated a pre-selected
amount, or when a time-out has occurred. The values of resistor R1 to R4 vary
with different types of motor and optical switch desired.
5.3.3 LCD and Hardware Buttons
The LCD and hardware button are provided for sensor value readout and infusion
rate programming or viewing. The designer can also choose to provide dials in
place of the LCD and hardware buttons.
5.3.4 Microprocessor
A microprocessor is used to coordinate the on-off sequence of the syringe pump

motor mechanism, to digitise and perform signal processing on the sensor read-
ings, and to allow the user interaction with the system (e.g. checking the sensor
readings, or changing the infusion rates via the LCD and hardware buttons).
Sensor signals are digitized using the on-board Analog-To-Digital (ATD) chan-
nel. As described, the bulk functionality of the system are software-oriented.
5.4 Signal Processing and Control Algorithm

Programming
Noise will be inevitable when sensor signals are in nano-current magnitude, and
hence filtering of noise was required. Since the sensor signal is fed into the mi-
croprocessor, most digital filters could be implemented.
In this design, a simple moving average filtering was employed to obtain a rel-
atively stable current readout on the microprocessor. A fast filtering mechanism
was required to permit real-time reporting of sensor readings. Table 5.2 describes
a fast moving average algorithm employed in the system. The algorithm used a
first-in first-out buffer, with 1000 taps set aside for the averaging process.
Table 5.2. Fast moving average filter
The moving average filter has the following transfer function:
Y (z −1 )
G(z −1 ) =
X(z −1 )
1 + z −1 + · · · + z −M +1
=
M
where X(.) and Y (.) represent the input and output of the filter respectively,
and M is the length of the averaging window.
The above filter can be implemented in software as follows:
x[n] + · · · + x[n − M + 1]
y[n] = (5.1)
M
or
x[n] x[n − M ]
y[n] = y[n − 1] + − (5.2)
M M
The number of operations required to compute y[n] in equation (5.2) is less

than the operations required in equation (5.1).
5.5 Circuit Performance

To examine the performance of the nano-ampere measuring circuit, used in con-
junction with the fast moving averaging algorithm, the linearity of the mea-
suring circuit was tested by injecting test input currents into the circuit and
measuring the output of the circuit. Test currents in the nano-ampere (nA)
range were generated, by means of using different resistor values for Rbody1 and
Rbody2 to effect different current values (see Fig 5.8). Since Rbody1 and Rbody2
reflect the generation of current from hydrogen peroxide and its flow across a
constant potential, this current generation can be simulated by using different
resistors.
For the purpose of the test, Rbody2 was set equal to Rbody1 , and the glucose
sensing circuit was constructed using the National Semiconductors LMC6464
op-amp. This micro-power quad CMOS operational amplifier features a 20μA
supply current per amplifier, >10TΩ input resistance, 150fA input current, and
+5V DC single supply operation.
Fig 5.12 shows a comparison of the output measured by the nano-ampere
measuring circuit (i.e. Vout /R in Fig 5.8) with the test currents applied at the
input of the circuit (i.e. at WRK electrode). Both the input current and Vout
were measured using MetraHit 30M (Gossen-Metawatt GmbH, Germany) digi-
tal multimeter which has internal noise-filtering circuit for stable readings. The
result showed that the output measurement is linear with respect to the input
current.
5.5 Circuit Performance 123
Input current versus measured current

3
/R (in nA)
2
out
V
1
1 2 3
Input current (in nA)
Fig. 5.12. The output measured by the nano-ampere measuring circuit is linear with
respect to the applied test current, over the tested range of 10 nA to 400 nA
V from circuit versus V from ATD

out out
2
1.8
1.6
ATD measurement (in V)
1.4
1.2
0.8
0.6
0.4
0.2
0
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
Vout (in V)
Fig. 5.13. Vout is linear with respect to the ATD’s measured output (10-bit ATD)
To examine the performance of the moving averaging filter in combination

with the ATD, the output of the nano-ampere measuring circuit was compared
to the those sampled by a 10-bit ATD channel and subsequently filtered by the
moving average algorithm. Fig 5.13 showed that Vout is linear with respect to
the ATD’s output measurement.
5.6 Basic Safety Issues

The microprocessor-controlled closed-loop system discussed in this chapter is in
its basic functional form and is not suitable for use in any testing involving live
human subjects. The design here provides a ready reference for the construction
of a compact and portable closed-loop system. The system hardware design
would still need to undergo consideration of the issue of safety. Some possible
safety issues (and guidelines to overcome them) are:
• Occlusion – obstruction in the tubings preventing the flow of infusate.
Occlusion can happen at the “patient” side (i.e. connection between the tip
of the tubing and the patient’s vessels or skin) or at the “pump” side (i.e.
connection between the syringe and the tubing). Occlusion can be sensed
from the lack of rotation of the pump motor, and the software can be
made to detect this event. However, for added safety, a dedicated pressure
sensor should be attached to the tubings to increase the detection
sensitivity.
• Circuit failure, e.g. motor rotation detector circuit failure.
The software should be written such that any failure in the motor rotation
detector or motor driving circuit would result in an immediate shut-down of
the infusion, and an alert to be sent to the user.
• Low battery alarm.
The fitting of a low battery alarm is useful. This can either be implemented
in software (for certain types of microprocessors) or by installing an external
battery monitoring circuit.
• Air-in-line detection.
In the case of intravenous delivery, the presence of air bubbles in the tubing
would be undesirable (and sometimes disastrous). Care should be taken to
emptied the bubbles from the tubing before connecting to the patient. Al-
ternatively, an “air-in-line” detector could be installed. The microprocessor-
based syringe pump designed in this did not include an “air-in-line” detector,
giving other researchers the flexibility to chose their detection method (e.g.
ultrasonic detector).
Fig 5.14 shows a working example of the system built around an existing
commercially-available syringe pump (SIMS Graseby MS-26 Syringe Driver,
Graseby Medical Ltd, Watford, Hertfordshire, UK), as an alternative to a
custom-made pump.
5.6 Basic Safety Issues 125
Fig. 5.14. Working example of a portable closed-loop insulin delivery apparatus
Fig. 5.15. The internal of the example portable closed-loop insulin delivery apparatus
5.7 Conclusion
With the commercial-availability of subcutaneous glucose sensor, a closed-loop

insulin delivery apparatus that uses subcutaneous sensor readings could be con-
structed by adding a few off-the-shelf components. However, a real-time control
of BG level using MiniMed CGMS glucose sensor would be difficult unless
there is a method to ensure that the real-time subcutaneous sensor readings
could correlate well with reference glucometer readings in a predictable way.
MiniMed CGMS subcutaneous glucose sensor assembly can be interfaced
with a glucose-sensing circuit, allowing a closed-loop control apparatus to be
pocketable. The design and integration of the apparatus has been presented as
a basic “working” framework to provide a ready reference to other research in
the field, allowing them to construct a portable and wearable closed-loop control
system in a minimal timeframe. With the design having a software component,
other researchers can conveniently tailor aspects of the system according to their
needs.
5.8 Summary
This chapter presented the design of a bedside closed-loop insulin delivery sys-
tem, constructed using off-the-shelf components. Such an apparatus was shown
to function in a clinical settings under constant supervision. As a natural step
towards miniaturisation, this chapter detailed the design of a portable and wear-
able closed-loop delivery system that uses CGMS as its glucose sensor. Cir-
cuit diagrams explaining the design of the glucose sensor interface, and the
microprocessor-driven syringe pump were presented. The chapter then described
a noise filtering algorithm that worked well with the glucose sensing circuit in-
troduced in the chapter. A working example of the design was presented in the
later part of the chapter, together with discussions on the basic safety issues
related to the pump.
6
Conclusions
Researches have found that a well-controlled BG greatly reduces the mortality

and morbidity associated with diabetes. A closed-loop control system could be
used to achieve a tight BG control. It can be seen (in Chapter 4) that designing a
closed-loop system for BG control is not an easy task. The design of the closed-
loop system needs to consider the characteristics of each components in the
system, the design of the control algorithm, and how the combination (of the
sensor, infuser and algorithm) can achieve the desired outcome. The design of
the control algorithm depends on the selection of the BG measurement method
and insulin infusion routes. The selection of the BG measurement method and
insulin infusion route depends on the available technology and issue of safety
associated with the selected method/routes.
It is not possible to describe all the mathematical models and control algo-
rithms reported in the literature in one book. This book gives the frequently
cited algorithms & models as examples. The readers are encouraged to read
other books in parallel, to get a different perspective.
In this book, our aims were:
• to show that blood glucose control is important not only in ambulatory per-
sons, but also in critically-ill patients;
• to introduce the reader to the basics of closed-loop blood glucose control,
which included the functional description of the gluco-regulatory system, and
the various components making up the complete system;
• to show, as an example, how a commercially available glucose sensor can be
used to build a bedside closed-loop system, and how can it be incorporated
to form a pocketable microprocessor-controlled closed-loop system;
• to show that glucose control is a vast and complicated field, simply because
our body’s glucose-regulatory system is complex and inter-related with various
parts of the body function.
This book does not discuss biotechnology, gene therapy, stem cells or
pancreas/beta-cell transplantation, which also played a part in helping to cure
diabetes.
springerlink.com
128 6 Conclusions
The ultimate goal in closed-loop control of BG is to have a system that reg-

ulates BG using the non-invasive glucose sensing and insulin delivery, and is
also capable of being “set and forget”. This goal may not be too far away, as
progress is constantly being made both in the field of non-invasive glucose sensing
and insulin delivery. Perhaps, one day, the current landscape may dramatically
change, as new understanding of the glucose-regulatory system emerge, bringing
new insight into the way blood glucose will be controlled.
A
Mathematical Derivation
A.1 Solving Recursive Least Square Estimation

Recall in Chapter 4 Section 4.5.3 that the optimal filter coefficients can be found
∗
by setting the derivative of ξ(n) with respect to wm (n) to zero to obtain:
n
∂ξ(n) ∂e∗ (n, k)
= λn−k e(n, k)
∂w∗ k=0
∂w∗
n ∂ d(k) − {w∗ }T x∗
n−k T
⇒ 0= λ d(k) − w (n)x(k)
k=0
∂w∗
n
λn−k d(k) − wT (n)x(k) {x∗ }
T
0=
k=0
n n
⇒ λn−k x∗ (k)xT (k) w(n) = λn−k d(k)x∗
k=0 k=0
i.e.
Φ(n)w(n) = θ(n) (A.1)
where
n
Φ(n) = λn−k x∗ (k)xT (k)
k=0
n−1
= λn−k x∗ (k)xT (k) + λn−n x∗ (n)xT (n)
k=0
n−1
= λ(n−1)−k x∗ (k)xT (k) + x∗ (n)xT (n)
k=0
= λΦ(n − 1) + x∗ (n)xT (n)
is of recursive relationship, and
130 A Mathematical Derivation
n
θ(n) = λn−k d(k)x∗
k=0
= λθ(n − 1) + d(n)x∗ (n)
where x∗ denotes a complex conjugate.

To solve equation A.1, it is necessary to find the inverse of Φ(n). This can be
achieved by using Matrix Inversion Lemma:
If A and B are M × M positive definite matrices, D is a N × N matrix, C is
a M × N matrix which are related by
A = B −1 + CD−1 C T
then
A−1 = B + BC(D + C T BC)−1 C T B
Let A = Φ(n), B −1 = λΦ(n − 1), C = x(n), D = 1, we can write
λ−1 Φ−1 (n − 1)x∗ (n)

Φ−1 (n) = λ−1 Φ−1 (n − 1) − λ−1 xT (n)Φ−1 (n − 1)
1 + λ−1 xT (n)Φ−1 (n − 1)x∗
or
P(n) = λ−1 P(n − 1) − λ−1 k(n)xT (n)P(n − 1) (A.2)
It is now possible to develop the recursive expression for the filter coefficients
w(n) = Φ−1 (n)θ(n)
= Φ−1 (n) (λθ(n − 1) + d(n)x∗ (n))
= λ λ−1 Φ−1 (n − 1) − λ−1 k(n)xT (n)Φ−1 (n − 1) θ(n − 1) + d(n)Φ−1 (n)x∗ (n)
= Φ−1 (n − 1)θ(n − 1) − k(n)xT (n)Φ−1 (n − 1)θ(n − 1) + d(n)Φ−1 (n)x∗ (n)

= w(n − 1) − k(n)x∗ (n)w(n − 1) + d(n)k(n)
= w(n − 1) + d(n) − xT (n)w(n − 1) k(n)
Equation A.2 is a form of Riccati equation, and we can relate k(n) = P(n)x∗ (n)
with some algebraic manipulation (Adapted from [230]).
A.2 Linearisation of Non-linear Model

In linearising a nonlinear system, we assume that the variables deviate only
slightly from some operating condition. Consider a nonlinear system whose out-
put y is a function of two inputs x1 and x2 :
y = f (x1 , x2 ) (A.3)
To obtain a linear approximation to this nonlinear system, equation A.3 may be

expanded into a Taylor series about the normal operating point x̄1 ,x̄2 :
A.2 Linearisation of Non-linear Model 131

∂f ∂f
y =f (x̄1 , x̄2 ) + (x1 − x̄1 ) + (x2 − x̄2 )
∂x1 ∂x2
2 2

1 ∂ f 2 ∂ f ∂2f 2
+ (x1 − x̄1 ) + 2 (x1 − x̄1 ) (x2 − x̄2 ) + (x2 − x̄2 )
2! ∂x21 ∂x1 ∂x2 ∂x22
+ ···
where the partial derivatives are evaluated at x1 = x̄1 , x2 = x̄2 . The higher
order terms may be neglected. The linear model of this nonlinear system around
the normal operating point can be given by
y − ȳ = K1 (x1 − x̄1 ) + K2 (x2 − x̄2 )

∂f ∂f
where ȳ = f (x̄1 , x̄2 ), K1 = and K2 =
∂x1 x1 =x̄1 ,x2 =x̄2 ∂x2 x1 =x̄1 ,x2 =x̄2
The linearisation above is valid in the vicinity of the operating condition. It

would not be valid if the operating conditions vary widely (Adapted from [213]).
Example
Linearisation of the “Minimal model for glucose disappearance”:

dG
= − (p1 + X) G − p1 Gb
dt
Let f = dG
dt = − (p1 + X) G − p1 Gb , then
∂f ∂
= [− (p1 + X) G + p1 Gb ] = − (p1 + X)|X=X̄ = − p1 + X̄
∂G G=Ḡ,X=X̄
∂G
∂f ∂
= [−p1 G − XG + p1 Gb ] = −G|G=Ḡ = −Ḡ
∂X G=Ḡ,X=X̄ ∂X
f − f¯ = − p1 + X̄ G − Ḡ + −Ḡ X − X̄
⇒ f = − p1 + X̄ G − Ḡ + −Ḡ X − X̄ + − p1 + X̄ Ḡ + p1 Gb
= − p1 + X̄ G − Ḡ X − X̄ + p1 Gb
dG
The linearised model is = − p1 + X̄ G − Ḡ X − X̄ + p1 Gb
dt
B
Model Parameters
B.1 Ackerman’s Model
Ackerman’s model parameters, as tabulated by Yipintsoi et al [252], is partially

reproduced here.
Table B.1. Ackerman’s model parameters
Subject EG EH m1 m2 m3 m4
N6 0.09 2.5 0.0351 0.0262 0.0540 0.0262
N4 0.22 6.5 0.0273 0.0271 0.0540 0.0136
N3 0.23 1.3 0.0617 0.0438 0.0590 0.0277
N1 0.41 6.5 0.0251 0.0703 0.0980 0
N2 0.65 11.3 0.0536 0.0858 0.0108 0.0423
N5 0.70 2.6 0.0574 0.1575 0.0729 0
D4a 0.15 8.8 0.0020 0.0014 0.0220 0
D4b 0.08 3.9 0.0009 0.0031 0.0415 0
D2 0.11 4.4 0.0016 0.0143 0.0293 0
D8 0.32 13.5 0.0035 0.0006 0.0418 0
D7a 0.33 364 0.0027 0.0012 0.0260 0
D6 0.33 939 0.0081 0.0857 0.0011 0.0006
D11 0.36 6.9 0.0088 0.0001 0.0338 0
D1 1.10 121 0.0063 0.0011 0.0072 0
D7b 1.60 442 0.0056 0.0011 0.0097 0
• Nn = Normal patient; Dn = diabetic patient.

2
N ni − ni
• EG and EH are error values defined as where ni = measured
i=1
N

data, ni = best-fitted data, and N = total number of data points.
134 B Model Parameters
B.2 Minimal Model

Minimal model’s parameters, as reported by Bergman et al [197] and Furler et
al [176].
Table B.2. Human subject physiological values at time of experiment
Subject Gender IBW G0 I0 VG

Lean % mg/dl μU/ml dl
1 M 101 94 17 92.6
2 M 100 91 9 125.4
3 M 105 85 15 132.1
4 M 95 99 8 108.4
5 M 100 93 11 127.6
6 M 97 98 4 101.5
7 M 88 97 3 126.5
8 M 98 93 6 118
Obese % mg/dl μU/ml dl
9 F 153 100 8 92.9
10 F 154 94 15 110.9
11 F 147 96 17 81
12 F 142 110 26 161
13 F 148 99 21 168.4
14 F 130 92 20 135.4
15 F 138 102 81 116.7
16 F 172 109 37 152.8
17 F 206 104 68 195.2
18 F 153 103 16 146
Furler (all) – 70 81 15 12 L
• IBW = Ideal Body Weight; G0 = basal glucose value; I0 = basal insulin value;
VG = glucose distribution space;
• G(0) = BG concentration at the start of clinical experiment;
• “Furler (all)” = values for all three “hypothetical subjects”, as presented in
Furler et al [176].
1. “Furler1” = normal patient;
2. “Furler2” = insulin-resistant patient;
3. “Furler3” = glucose-resistant patient.
B.3 Cobelli’s Model

The parameter values for Cobelli’s model considered a normal individual of
70 kg body weight, with plasma blood glucose y1 = 91.5 mg/dl, plasma insulin
concentration y2 =11 μU/ml, and blood glucagon concentration y5 =75 pg/ml.
u1p (0)=4.9 U and u2p (0)=0.49 U.
B.4 Hovorka’s Model 135
Table B.3. Minimal Model parameters
Subject G(0) p1 p2 p3 (×106 ) I(0) γ(×103 ) h n

Lean mg/dl min−1 min−1 min−2 /(μU/ml) μU/ml μU/ml min2 mg/dl min−1
1 298 0.0296 0.0186 6.51 333 5.36 90.9 0.23
2 276 0.0192 0.0262 14.7 69 1.40 100.0 0.18
3 250 0.0374 0.0478 8.73 33 2.93 87.5 0.30
4 337 0.0363 0.0081 4.01 192 2.40 93 0.23
5 329 0.0464 0.0038 3.61 73 1.69 119 0.13
6 296 0.0136 0.0341 17.3 50 0.89 90.9 0.22
7 248 0.0151 0.0313 9.7 15 0.69 82.6 0.22
8 271 0.0217 0.0292 19.1 14 0.28 87.3 0.11
Obese
9 297 0.0400 0.0042 2.56 209 3.72 154 0.22
10 242 0.0323 0.0034 2.67 104 3.46 145 0.19
11 348 0.0466 0.0676 16.1 264 7.47 186 0.09
12 256 0.0180 0.0108 2.29 99 3.42 153 0.13
13 217 0.0113 0.0034 0.97 225 3.30 122 0.14
14 224 0.0113 0.0240 4.89 185 1.36 123 0.11
15 258 0.0246 0.0069 0.55 337 2.70 175 0.13
16 267 0.0100 0.0166 2.02 248 3.40 108 0.13
17 242 0.0071 0.0125 1.58 20 6.11 143 0.13
18 254 0.0093 0.0200 7.78 16 1.19 132 0.13
Furler1 – 0.028 0.0250 13 – – – 0.09
Furler2 – 0.028 0.0250 5 – – – 0.09
Furler3 – 0 0.0250 13 – – – 0.09
Table B.4. Parameter values for pathological states and other conditions
Overt diabetes Latent diabetes Obesity

F2 =0.037 constant F2 =0.037 constant H4 =0.0012 constant
H4 =0.0012 constant H4 =0.0012 constant b42 = 7 × 10−3
b42 = 7 × 10−3 b42 = 7 × 10−3 c42 =40.47
c42 =−40.47 c42 =−40.47 m02 =0.13
bw =4.5×10−3 bw =4.5×10−3 b6 =0.5
b6 =5×10−3 b6 =1×10−2 c6 =−3.64
cw =−306.25 cw =−306.25
c6 =−363.55 c6 =−176.68
B.4 Hovorka’s Model

Hovorka’s model constants and parameters (primarily from reference [211])
136 B Model Parameters
Table B.5. Hovorka’s model constants
Symbol Quantity Value

k12 Transfer rate 0.066 min−1
ka1 Deactivation rate 0.006 min−1
ke Insulin elimination from plasma 0.138 min−1
VI Insulin distribution volume 0.12 L kg−1
VG Glucose distribution volume 0.16 L kg−1
AG Carbohydrate bio-availability 0.8 (unitless)
tmax,G Time-to-maximum of carbohydrate 40 min
absorption
Table B.6. Hovorka’s model parameters for the purpose of Bayesian parameter esti-
mation
Symbol Quantity Mean Value

f
SIT Insulin sensitivity of distribution/transport 51.2 × 10−4 min−1 per mU/l
f
SID Insulin sensitivity of disposal 8.2 × 10−4 min−1 per mU/l
f
SIE Insulin sensitivity of EGP 520 × 10−4 per mU/l
EGP0 EGP extrapolated to zero insulin concentration 0.0161 mmol · kg−1 · min−1
F01 Insulin-independent glucose flux 0.0097 mmol · kg−1 · min−1
tmax,I Time-to-maximum of absorption of subcuta- 55 min
neously injected short-acting insulin
B.5 Sorensen’s Model

The nomenclature used in Sorensen’s model parameter, as according to [209]:
i(t) = insulin rate entering patient [mU/min];
G = glucose concentration [mg/dl];
I = insulin concentration [mU/dl];
χ = glucagon concentration [pg/ml].
V = volume [dl or L];
Q = vascular blood water flow rate [dl/min or L/min];
Γ = metabolic source or sink rate [mg/min];
T = trans-capillary diffusion rate [min];
t = time [min];
• The first subscript represents physiological compartment, i.e., B (brain), G

(gut), K (kidney), L (Liver), H (heart), PV (Periphery), A (hepatic artery)
• Second subscript are defined as: I = interstitial fluid space; V = vascular blood
water space;
B.5 Sorensen’s Model 137
• Metabolic rate subscripts:

PGU = peripheral glucose uptake;
RBCU = red blood cell glucose uptake;
PIR = peripheral insulin release;
LIC = liver insulin clearance;
KIC = kidney insulin clearance;
PIC = peripheral insulin clearance;
KGE = kidney glucose excretion;
HGU = hepatic glucose uptake;
HGP = hepatic glucose production;
GGU = gut glucose utilisation;
BGU = brain glucose uptake;
Pχ C = plasma glucagon clearance;
Mχ C = metabolic glucagon clearance;
Pχ R = pancreatic glucagon release;
• First superscript are defined as: G = glucose; I = insulin; χ = glucagon; B =
basal value; N = normalised value (divided by basal value)
Table B.7. Values for metabolic sources/sinks
ΓBGU = 70 mg/min
ΓRBGU = 10 mg/min
ΓGGU = 20 mg/min
ΓPBGU = 35 mg/min
B
ΓHGP = 155 mg/min
B
ΓHGU = 20 mg/min
GP I
GNPI = basal value
(e.g. basal value = 86.81)
N IP I
IP I = basal value (e.g. basal value = 5.304)
GL
GNL = basal value (e.g. basal value = 101)
N IL
IL = basal value (e.g. basal value = 21.43)
Table B.8. Sorensen’s model parameter values for the diabetic patient
G
VBV = 3.5 dL QG
B = 5.9 dL/min VBI = 0.26 L QIB = 0.45 L/min τI = 25 min
G
VH = 13.8 dL QG
H = 43.7 dL/min VHI = 0.99 dL QIH = 3.12 L/min τχ = 65 min
VGG = 11.2 dL QG
G = 10.1 dL/min VGI = 0.94 L QIG = 0.72 L/min TB = 2.1 min
VLG = 25.1 dL QG
L = 12.6 dL/min VLI = 1.14 L QIL = 0.9 L/min TPG = 5.0 min
VKG = 6.6 dL QG
K = 10.1 dL/min VKI = 0.51 L QIK = 0.72 L/min TPI = 20 min
VPGV = 10.4 dL QG
P = 15.1 dL/min VPI V = 0.74 L QIP = 1.05 L/min VBI = 4.5 dL
QG
A = 2.5 dL/min QIA = 0.18 L/min VP I = 63 dL
References
1. J. Desemore, J.P. Mordes, and A.A. Rossini. Hypoglycemia. In J.M. Rippe,

R.S. Irwin, J.S. Alpert, and M.P. Fink, editors, Intensive Care Medicine 2nd Ed.,
chapter 99, pages 1000–1009. Little, Brown and Company, 1991.
2. I.B. Hirsch and J.B. McGill. Role of insulin in management of surgical patients
with diabetes melliltus. Diabetes Care, 13(9):980–991, September 1990.
3. W.F. Ganong. Review of Medical Physiology. Appleton & Lange, Stanford, Con-
necticut, 1997.
4. A.C. Guyton and J.E. Hall. Textbook of Medical Physiology. W.B. Saunders
Company, 9 edition, 1996.
5. I.A. Macfarlane. The Millenia Before Insulin, volume 1, chapter 1, pages 3–9.
Blackwell Scientific Publications, 1991.
6. M. Bliss. The Discovery of Insulin, volume 1, chapter 2, pages 10–14. Blackwell
Scientific Publications, 1991.
7. B.A. Mizock. Alterations in carbohydrate metabolism during stress: A review of
the literature. The American Journal of Medicine, 98:75–84, January 1995.
8. N.B. Watts, S.S.P. Gebhart, R.V. Clark, and L.S. Phillips. Postoperative man-
agement of diabetes mellitus: Steady-state glucose control with bedside algorithm
for insulin adjustment. Diabetes Care, 10(6):722–728, Nov - Dec 1987.
9. G. Van den Berghe, P. Wouters, F. Weekers, C. Verwaest, F. Bruyninckx,
M. Schetz, D. Vlasselaers, P. Ferdinande, P. Lauwers, and R. Bouillon. Intensive
insulin therapy in critically ill patients. N Engl J Med, 345(19):1359–1367, 2001.
10. A.M. Albisser and B.S. Leibel. Artificial beta cell, 1981. US Patent 4,245,643.
11. The Diabetes Control and Complications Trial Research Group. The effect of
intensive treatment of diabetes on the development and progression of long-term
complications in insulin-dependent diabetes mellitus. N Engl J Med, 329(14):977–
986, 30 September 1993.
12. UK Prospective Diabetes Study (UKPDS) Group. Intensive blood-glucose control
with sulphonylureas or insulin compared with conventional treatment and risk of
complications in patients with type 2 diabetes (UKPDS 33). The Lancet, 352:837–
853, 12 September 1998.
13. Y. Ohkubo, H. Kishikawa, E. Araki, T. Miyata, S. Isami, S. Motoyoshi, Y. Kojima,
N. Furuyoshi, and M. Shichiri. Intensive insulin therapy prevents the progression
of diabetic microvascular complications in Japanese patients with non-insulin-
dependent diabetes mellitus: a randomised prospective 6-year study. Diabetes
Research and Clinical practice, 28:103–117, May 1995.
140 References
14. S.H. Golden, C. Peart-Vigilance, W.H. L. Kao, and F.L. Brancati. Perioperative
glycemic control and the risk of infectious complications in a cohort of adults with
diabetes. Diabetes Care, 22(9):1408–1414, September 1999.
15. L.C. Gatewood, E. Ackerman, J.W. Rosevear, and G.D. Molnar. Simulation stud-
ies of blood-glucose regulation: Effect of intestinal glucose absorption. Computers
and Biomedical Research, 2:15–27, 1968.
16. A.H. Clemens, P.H. Chang, and R.W. Myers. The development of Biostator,
a glucose-controlled insulin infusion systems (GCIIS). Horm Metab Res Suppl,
7:23–33, March 1977.
17. A.M. Albisser, B.S. Leibel, T.G. Ewart, Z. Davidovac, C.K. Botz, and W. Zingg.
Clinical control of diabetes by the artificial pancreas. Diabetes, 23(5):397–404,
May 1974.
18. D.J. Chisolm, E.W. Kraegen, D.J. Bell, and D.R. Chipps. A semi-closed loop
computer-assisted insulin infusion system: Hospital use for control of diabetes in
patients. Med J Aust, 141:784–788, 1984.
19. L. Heinemann and F. J. Ampudia-Blasco. Glucose clamps with the Biostator: A
critical reappraisal. Horm Metab Res, 26:579–583, 1994.
20. U. Fischer, K. Rebrin, T.v. Woedtke, and P. Abel. Clinical usefulness of the
glucose concentration in the subcutaneous tissue - properties and pitfalls of elec-
trochemical biosensors. Horm Metab Res, 26:515–522, 1994.
21. U. Fischer. Continuous in vivo monitoring in diabetes: the subcutaneous glucose
concentration. Acta Anaesthesiol Scand Suppl, 39(104):21–29, 1995.
22. E. F. Pfeiffer. Review. on the way to the automated (blood) glucose regulation
in diabetes: the dark past, the gray present and the rosy future. Diabetologia,
30:51–65, 1987.
23. A.M. Albisser. Insulin delivery systems: Do they need a glucose sensor? Diabetes
Care, 5(3):166–173, May-June 1982.
24. B.H. Ginsberg. An overview of minimally invasive technologies. Clin. Chem.,
38(9):1596–1600, 1992.
25. K. Rebrin, G. M. Steil, William P. Van Antwerp, and John J. Mastrototaro. Sub-
cutaneous glucose predicts plasma glucose independent of insulin: Implications
for continuous monitoring. Am J Physiol Endocrinol. Metab. 40, 277:E561–E571,
1999.
26. C. Meyerhoff, E. Bischf, F. Sternberg, H. Zier, and E.F. Pfeiffer. On line con-
tinuous monitoring of subcutaneous tissue glucose in men by combining portable
glucosensor with microdialysis. Diabetologia, 35:1087–1092, 1992.
27. A. Heller. Implanted electrochemical glucose sensor for the management of dia-
betes. Annu. Rev. Biomed. Eng., 1:153–175, 1999.
28. B.H. Ginsberg. The role of technology in diabetes therapy. Diabetes Care,
17 Suppl. 1:50–55, June 1994.
29. Editorial. Problems associated with subcutaneously implanted glucose sensors.
Diabetes Care, 22(2):143–148, February 1999.
30. M. Shichiri, R. Kawamori, Y. Goriya, Y. Yamasaki, M. Nomura, N. Hakui, and
H. Abe. Glycaemic control in pancreatectomized dogs with a wearable artificial
endocrine pancreas. Diabetologia, 24:179–184, 1983.
31. M. Shichiri, R. Kawamori, Y. Goriya, Y. Yamasaki, M. Nomura, N. Hakui, and
H. Abe. Closed-loop glycaemic control with a wearable artificial endocrine pan-
creas - variations in daily insulin requirements to glycemic response. Diabetes,
33:1200–1202, December 1984.
References 141
32. Tran Minh Canh. Biosensors, volume 1 of Sensor Physics and Technology Series,
chapter 2, pages 7–19. Chapman & Hall and Masson, 2-6 Boundary Row, London
SE1 8HN, 1993.
33. Gbor Harsnyi. Sensors in Biomedical Applications, - Fundamentals, Technology
& Applications. Technomic Publishing Company, Inc., Pennsylvania USA, 2000.
34. MiniMed. MiniMed glucose sensor - instructions for use (REF MMT-7002).
Northridge, CA, May 2001.
35. V. Poitout, D. Moatti-Sirat, G. Reach, Y. Zhang, G.S. Wilson, F. Lemonnier,
and J. C. Klein. A glucose monitoring system for on-line estimation in man of
blood glucose concentration using a miniaturized glucose sensor implanted in the
subcutaneous tissue and a wearable control unit. Diabetologia, 36:658–663, 1993.
36. J.S. Schultz, S. Mansouri, and I.J. Goldstein. Affinity sensor: A new technique
for developing implantable sensors for glucose and other metabolites. Diabetes
Care, 5(3):245–253, 1982.
37. J.I. Peterson and G.G. Vurek. Fiber-optic sensors for biomedical applications.
Science, 224(4645):123–127, 1984.
38. R.J. McNichols and G.L. Coté. Optical glucose sensing in biological fluids: an
overview. Journal of Biomedical Optics, 5(1):5–16, January 2000.
39. L.J. McCartney, J.C. Pickup, O.J. Rolinski, and D.J.S. Birch. Near-infrared
fluorescence lifetime assay for serum glucose based on Allophycocyanin-Labeled
Concanavalin A. Analytical Biochemistry, 292:216–221, 2001.
40. J.C. Pickup, F. Hussain, N.D. Evans, O.J. Rolinski, and D.J.S. Birch.
Fluorescence-based glucose sensors. Biosensors and Bioelectronics, 20:2555–2565,
2005.
41. J. Pickup, L. McCartney, O. Rolinski, and D. Birch. In vivo glucose sensing for
diabetes management: Progress towards non-invasive monitoring. BMJ, 319:1–4,
13 November 1999.
42. R. Ballerstadt and J.S. Schultz. A fluorescence affinity hollow fiber sensor for
continuous transdermal glucose monitoring. Analytical Chemistry, 72:4185–4192,
2000.
43. J.S. Kristensen, S. Aasmul, J.K. Nielsen, and J.S. Christiansen. Trans-
cutaneous fluorescence lifetime based continuous glucose reading for long term
interrogation, 2005. Poster at 65th annual American Diabetes Association
meeting 2005, Download from http://www.precisense.com/graphics/Synkron-
Library/Products/ presicense handout higress.pdf.
44. W.F. March, K. Ochsner, and J. Horna. Intraocular lens glucose sensor. Diabetes
Technology & Therapeutics, 2(1):27–30, 2000.
45. L. Li and D.R. Walt. Dual-analyte fiber-optic sensor for the simultaneous and
continuous measurement of glucose and oxygen. Analytical Chemistry, 67:3746–
3752, 1995.
46. H. Gunasingham, Chin-Huat Tan, and J.K.L. Seow. Fiber-optic glucose sen-
sor with electrochemical generation of indicator reagent. Analytical Chemistry,
62:755–759, 1990.
47. M.S. Abdel-Latif and G.G. Guilbault. Fiber-optic sensor for the determination of
glucose using micellar enhanced chemiluminescence of the peroxyoxalate reaction.
Analytical Chemistry, 60:2671–2674, 1988.
48. N. Wisniewski, F. Moussy, and W.M. Reichert. Characterization of implantable
biosensor membrane biofouling. Fresenius J Anal Chem, 366:611–621, 2000.
49. M.J. McShane. Potential for glucose monitoring with nanoengineered fluorescent
biosensors. Diabetes Technology & Therapeutics, 4(4):533–538, 2002.
142 References
50. E.F. Pfeiffer. The ”Ulm Zucker Uhr System” and its consequences. Horm Metab
Res, 26:510–514, 1994.
51. J.A. Tamada, M. Lesho, and M.J. Tierney. Keeping watch on glucose - New
monitors help fight the long-term complications of diabetes. IEEE Spectrum,
39(4):52–57, 2002.
52. S. Mansouri and J.S. Schultz. A miniature optical glucose sensor based on affinity
binding. Biotechnology, 2:885–890, October 1984.
53. J.C. Pickup. In vivo glucose monitoring: Sense and sensorbility. Diabetes Care,
16(2):535–539, 1993.
54. D.C. Klonoff. Noninvasive blood glucose monitoring. Diabetes Care, 20(3):433–
437, March 1997.
55. J.A. Tamada, S. Garg, L. Jovanovic, K.R. Pitzer, S. Fermi, R.O. Potts, and the
Cygnus Research Team. Noninvasive glucose monitoring - Comprehensive clinical
results. JAMA, 282(19):1839–1844, November 1999.
56. M.J. Tierney, Y. Jayalakshmi, N.A. Parris, M.P. Reidy, C. Uhegbu, and P. Vi-
jayakumar. Design of a biosensor for continual, transdermal glucose monitoring.
Clinical Chemistry, 45(9):1681–1683, 1999.
57. S.K. Garg, R.O. Potts, N.R. Ackerman, S.J. Fermi, J.A. Tamada, and H.P Chase.
Correlation of Fingerstick blood glucose measurements with GlucoWatch biog-
rapher glucose results in young subjects with type 1 diabetes. Diabetes Care,
22(10):1708–1712, October 1999.
58. J. Kost. Ultrasound-assisted insulin delivery and noninvasive glucose sensing.
Diabetes Technology & Therapeutics, 4(4):489–497, 2002.
59. S. Mitragotri, M. Coleman, J. Kost, and R. Langer. Transdermal extraction of
analytes using low-frequency ultrasound. Pharmaceutical Research, 17(4):466–
470, 2000.
60. J. Kost, S. Mitragotri, R.A. Gabbay, M. Pishko, and R. Langer. Transdermal
monitoring of glucose and other analytes using ultrasound. Nature Medicine,
6(3):347–350, March 2000.
61. S. Gebhart, M. Faupel, R. Fowler, C.Kapsner, D. Lincoln, V. McGee, J. Pasqua,
L. Steed, M. Wangsness, F. Xu, and M. Vanstory. Glucose sensing in transdermal
body fluid collected under continuous vacuum pressure via micropores in the
stratum corneum. Diabetes Technology & Therapeutics, 5(2):159–166, 2003.
62. J.A. Eppstein, M.R. Hatch, and D. Yang. Microporation of human skin for drug
delivery and monitoring applications, 2000. United States Patent US 6,142,939.
63. L. Heinemann and G. Schmelzeisen-Redeker on behalf of the Non-invasive task
force N IT F . Non-invasive continuous glucose monitoring in type I diabetic pa-
tients with optical glucose sensors. Diabetologia, 41:848–854, 1998.
64. J.W. Hall and A. Pollard. Near-infrared spectrophotometry: A new dimension in
clinical chemistry. Clinical Chemistry, 38(9):1623–1631, 1992.
65. John A Dean. Dean’s Analytical Chemistry Handbook. McGraw-Hill, New York,
2nd edition, 2004.
66. Y. Mendelson, A.C. Clermont, R.A. Peura, and Been-Chyuan Lin. Blood glu-
cose measurement by multiple attenuated total reflection and infrared absorption
spectroscopy. IEEE Trans. Biomed. Eng., 37(5):458–465, May 1990.
67. D.A. Skoog and J.J. Leary. Principles of Instrumental Analysis. Saunders College
Publishing, Florida, USA, 4th edition, 1992.
68. K. Danzer, Ch. Fischbacher, K.U. Jagemann, and K.J. Reicheit. Near-infrared
diffuse reflection spectroscopy for non-invasive blood glucose monitoring. IEEE-
LEOS Newsletter (Special issue), pages 9–11, April 1998.
References 143
69. Yoen-Joo Kim and G. Yoon. Prediction of glucose in whole blood by near-infrared
spectroscopy: Influence of wavelength region, preprocessing, and hemoglobin con-
centration. Journal of Biomedical Optics, 11(4):041128, July/August 2006.
70. J.J. Burmeister and M.A. Arnold. Spectroscopic considerations for noninva-
sive blood glucose measurements with near infrared spectroscopy. IEEE-LEOS
Newsletter (Special issue), pages 6–9, April 1998.
71. G. Yoon, A.K. Amerov, K.J. Jeon, J.B. Kim, and Y-J Kim. Optical measurement
of glucose levels in scattering media. volume 20, pages 1897–1899. IEEE, 1998.
72. R. Abbink and C. Gardner. Getting under the skin. spie’s oemagazine, pages
18–20, September 2003.
73. N. Kaiser. Laser absorption spectroscopy with an ATR prism - noninvasive in
vivo determination of glucose. Horm Metab Res Suppl, 8:30–33, 1979.
74. P. Zheng, C.E. Kramer, C.W. Barnes, J.R. Braig, and B.B. Sterling. Noninva-
sive glucose determination by oscillating thermal gradient spectrometry. Diabetes
Technology & Therapeutics, 2(1):17–26, 2000.
75. G.B. Christison and H.A. MacKenzie. Laser photoacoustic determination of phys-
iological glucose concentrations in human whole blood. Medical & Biological En-
gineering & Computing, 31(3):284–290, 1993.
76. H.A. MacKenzie, H.S. Ashton, S. Spiers, Y. Shen, S.S Freeborn, J. Hannigan,
J. Lindberg, and P. Rae. Advances in photoacoustic noninvasive glucose testing.
77. F.M. Ham, G.M. Cohen, K. Patel, and B.R. Gooch. Multivariate determination
of glucose using NIR spectra of human blood serum. In Proceedings of the 16th
Annual International Conference of the IEEE on Engineering Advances: New
Opportunities for Biomedical Engineers, volume 2, pages 818–819, Baltimore, MD,
USA, 03 November-06 November 1994. IEEE, IEEE Engineering in Medicine and
Biology Society, 1994.
78. Ch. Fischbacher, K.-U. Jagemann, K. Danzer, U.A. Müller, L. Papenkordt, and
J. Schüler. Enhancing calibration models for non-invasive near-infrared spectro-
scopical blood glucose determination. Fresenius J Anal Chem, 359:78–82, 1997.
79. M.R. Robinson, R.P. Eaton, D.M. Haaland, G.W. Koepp, and E.V. Thomas.
Noninvasive glucose monitoring in diabetic patients: A preliminary evaluation.
80. G.W. Small, M.A. Arnold, and L.A. Marquardt. Strategies for coupling digital
filtering with partial least-squares regression: Application to the determination of
glucose in plasma by Fourier Transform Near Infrared spectroscopy. Analytical
Chemistry, 65:3279–3289, 1993.
81. H.M. Heise and R. Marbach. Multivariate determination of glucose in whole
blood by attenuated total reflection infrared spectroscopy. Analytical Chemistry,
61:2009–2015, 1989.
82. J. Tenhunen, H. Kopola, and R. Myllylä. Non-invasive glucose measurement
based on selective near infrared absorption requirements on instrumentation and
spectral range. Measurement, 24:173–177, 1998.
83. PerkinElmer Inc. FT-IR spectroscopy: Attenuated total re-
flectance (ATR) - Technical Note, 2007. [Online] Available at
http://las.perkinelmer.com/content/TechnicalInfo/TCH FTIRATR.pdf.
84. H.M. Heise. Non-invasive monitoring of metabolites using near infrared spec-
troscopy: State of the art. HormMetabRes, 28:527–534, 1996.
85. G.M. Oosta, Tayy-Wen Jeng, J.M. Lindberg, M.L. McGlashen, and J.L. Pezzaniti.
Multiplex sensor and method of use, 2003. United States Patent US 6,567,678
B1.
144 References
86. C.D. Malchoff, K. Shoukri, J.I. Landau, and J.M. Buchert. A novel noninvasive
blood glucose monitor. Diabetes Care, 25(12):2268–2275, December 2002.
87. D.C. Klonoff and J. Braig. Mid-infrared spectroscopy for noninvasive blood glu-
cose monitoring. IEEE-LEOS Newsletter (Special issue), pages 13–14, April 1998.
88. J.R. Braig, C.E. Kramer, B.B. Sterling, D.S. Goldberger, P. Zheng, A.M. Shulen-
berger, R. Trebino, R.A. King, and C.W. Barnes. Method for determining analyte
concentration using periodic temperature modulation and phase detection, 2006.
United States Patent US 7,006,857 B2.
89. L. Heinemann, U. Krämer, Hans-Martin Klötzer, M. Hein, D. Volz, M. Hermann,
T. Heise, and K. Rave the Non-invasive task force N IT F . Non-invasive glucose
measurement by monitoring of scattering coefficient during oral glucose tolerance
tests. Diabetes Technology & Therapeutics, 2(2):211–220, 2000.
90. S.L. Jacques and S.A. Prahl. ECE532 Biomedical Optics, Oregon Graduate In-
stitute, 1998. [Online] Available at http://www.phas.ubc.ca/∼mackay/phys404/
documents/BiomedicalOptics/ece532/class3/mie.html.
91. J.T. Bruulsema, J.E. Hayward, T.J. Farrell, and M.S. Patterson. Correlation
between blood glucose concentration in diabetics and noninvasively measured
tissue optical scattering coefficient. Optics Letters, 22(3):190–192, February 1997.
92. M. Kohl, M. Essenpreis, and M. Cope. The influence of glucose concentration
upon the transport of light in tissue-simulating phantoms. Phy. Med. Biol,
40:1267–1287, 1995.
93. J.S. Maier, S.A. Walker, S. Fantini, M.A. Franceschini, and E. Gratton. Possible
correlation between blood glucose concentration and the reduced scattering coef-
ficient of tissues in the near infrared. Optics Letters, 19(24):2062–2064, December
1994.
94. O.S. Khalil. Spectroscopic and clinical aspects of noninvasive glucose measure-
ments. Clin. Chem., 45(2):165–177, 1999.
95. M. Kinnunen, R. Myllylä, T. Jokela, and S. Vainio. In vitro studies toward non-
invasive glucose monitoring with optical coherence tomography. Applied Optics,
45(10):2251–2260, 1 April 2006.
96. H.M. Heise. Diffuse reflectance near-infrared spectrometry for non-invasive blood
glucose monitoring. IEEE-LEOS Newsletter (Special issue), pages 11–13, April
1998.
97. A. Kienle, L. Lilge, M.S. Patterson, R. Hibst, and R. Steiner. Spatially resolved
absolute diffuse reflectance measurements for noninvasive determination of the
optical scattering and absorption coefficients of biological tissue. Applied Optics,
35(13):2304–2314, 1 May 1996.
98. K.V. Larin, M.S. Eledrisi, M. Motamedi, and R.O. Esenaliev. Noninvasive
blood glucose monitoring with optical coherence tomography. Diabetes Care,
25(12):2263–2267, December 2002.
99. E.M.C. Hillman. Optical Tomography, volume 4, pages 2203–2235. Wiley-VCH
Verlag GmBH & Co. KGaA, 2003.
100. D. Huang, E.A. Swanson, C.P. Lin, J.S. Schuman, W.G. Stinson, W. Chang, M.R.
Hee, T. Flotte, K. Gregory, C.A. Puliafito, and J.G. Fujimoto. Optical coherence
tomography. Science, 254(5035):1178–1181, 22 November 1991.
101. Ilya Fine. Non-invasive method and system of optical measurements for deter-
mining the concentration of a substance in blood, 2002. United States Patent US
6,400,972 B1.
102. I. Fine, B. Fikhte, and L.D. Shvartsman. Occlusion spectroscopy as a new
paradigm for non-invasive blood measurements. Proceedings of SPIE, 4263:122–
130, 2001.
References 145
103. O. Cohen, I. Fine, E. Monashkin, and A. Karasik. Glucose correlation with

light scattering patterns - novel method for non-invasive glucose measurements.
Diabetes Technology & Therapeutics, 5(1):11–17, 2003.
104. L.D. Shvartsman and I. Fine. Light scattering changes caused by RBC aggrega-
tion: Physical basis for new approach to non-invasive blood count. Proceedings of
SPIE, 4263:131–142, 2001.
105. K. Yamakoshi and Y. Yamakoshi. Pulse glucometry: a new approach for
noninvasive blood glucose measurement using instantaneous differential near-
infrared spectrophotometry. Journal of Biomedical Optics, 11(5):054028, Septem-
ber/October 2006.
106. (NJ, USA) Princeton Instruments Inc and (MA, USA) Acton Re-
search Corp. Raman spectroscopy basics, 2006. [Online] Avail-
able at http://architect.wwwcomm.com/Uploads/Princeton/Documents/ Li-
brary/App Wht Papers/Raman Basics.pdf.
107. USA Kaiser Optical Systems Inc, MI. Raman Techni-
cal Resources - Raman Tutorial, 2006. [Online] Available at
http://www.kosi.com/raman/resources/tutorial/index.html.
108. Tae-Wong Koo, A.J. Berger, I. Itzkan, G. Horowitz, and M.S. Feld. Measure-
ment of glucose in human blood serum using Raman spectroscopy. IEEE-LEOS
Newsletter (Special issue), pages 18–19, April 1998.
109. R.J. McNichols, B.D. Cameron, and G.L. Coté. Development of a non-invasive
polarimetric glucose sensor. IEEE-LEOS Newsletter (Special issue), pages 30–31,
April 1998.
110. J. Lambert, M. Storrie-Lombardi, and M. Borchert. Measurement of physiologic
glucose levels using Raman spectroscopy in a rabbit aqueous humor model. IEEE-
LEOS Newsletter (Special issue), pages 19–22, April 1998.
111. A.J. Berger, Tae-Woong Koo, I. Itzkan, G. Horowitz, and M.S. Feld. Multi-
component blood analysis by near-infrared Raman spectroscopy. Applied Optics,
38(13):2916–2926, 1 May 1999.
112. L.D. Barron. Molecular Light Scattering and Optical Activity. Cambridge Uni-
versity Press, Cambridge CB2 1RP, 1982.
113. Gordan G. Hammes. Spectroscopy for the biological sciences. John Wiley & Sons
Inc, New Jersey, USA, 1sh edition, 2005.
114. B.D. Cameron and G.L. Coté. Noninvasive glucose sensing utilizing a digital
closed-loop polarimetric approach. IEEE Trans. Biomed. Eng., 44(12):1221–1227,
December 1997.
115. G.L. Cote, M.D. Fox, and R.B. Northrop. Optical polarimetric sensor for blood
glucose measurement. Proceedings of the 1990 Sixteenth Annual Northeast Bio-
engineering Conference, 1990, pages 101–102, 26-27 March 1990.
116. G.L. Cote, M.D. Fox, and R.B. Northrop. Laser polarimetric sensor for glucose
monitoring. Proceedings of the Twelfth Annual International Conference of the
IEEE Engineering in Medicine and Biology Society, pages 476–477, 1-4 November
1990.
117. H. Anumula, A. Nezhuvingal, Y. Li, , and B.D. Cameron. Development of a nonin-
vasive corneal birefringence compensated glucose sensing polarimeter. SPIE–The
International Society for Optical Engineering, 4958:303–312, February 2003.
118. G.L. Cote, M.D. Fox, and R.B. Northrop. Noninvasive optical polarimetric glucose
sensing using a true phase measurement technique. IEEE Trans. Biomed. Eng.,
39(7):752–756, July 1992.
119. G.L. Cote, M.D. Fox, and R.B. Northrop. Optical glucose sensing apparatus and
method, 1993. United States Patent US 5,209,231.
146 References
120. A.F. Browne, T.R. Nelson, and R.B. Northrop. Microdegree polarimetric mea-
surement of glucose concentrations for biotechnology applications. Proceedings
of the IEEE 1997 23rd Northeast Bioengineering Conference, 1997, pages 9–10,
21-22 May 1997.
121. C. Chou, Chien-Yuan Han, Wen-Chuan Kuo, Yen-Chuen Huang, Ching-Mei Feng,
and Jenn-Chyang Shyu. Noninvasive glucose monitoring in vivo with an optical
heterodyne polarimeter. Applied Optics, 37(16):3553–3557, 1998.
122. C. Chou and Po-Kang Lin. Noninvasive glucose monitoring with optical hetero-
dyne technique. Diabetes Technology & Therapeutics, 2(1):45–47, 2000.
123. S. Jang and M.D. Fox. Optical sensor using the magnetic optical rotatory effect
of glucose. IEEE-LEOS Newsletter (Special issue), pages 28–30, April 1998.
124. R.W. Waynant and V.M. Chenault. Overview of non-invasive fluid glucose mea-
surement using optical techniques to maintain glucose control in diabetes mellitus.
IEEE-LEOS Newsletter (Special issue), pages 3–6, April 1998.
125. Andrew C. Tam. Applications of photoacoustic sensing techniques. Reviews of
Modern Physics, 58(2):381–431, April 1986.
126. R. Nagar, B. Pesach, and U. Ben-Ami. Photoacoustic assay and imaging system,
2005. United States Patent US 6,846,288 B2.
127. C. Ruff, S. Lohmann, H. Lubatschowski, and W. Ertner. Photoacoustic deter-
mination of optical parameters of biological tissue. pages 325–325. IEEE, 8-
13 September 1996.
128. Zuomin Zhao. Pulsed photoacoustic techniques and glucose determination in hu-
man blood and tissue. ISBN 951-42-6689-7. Oulu University Press, Oulu, 2002.
[Online] Available from http://herkules.oulu.fi/isbn9514266900/.
129. Z. Zhao and R.A. Myllylä. Photoacoustic blood glucose and tissue measurements
based on optical scattering effect. Proceedings of the SPIE, 4707:153–157, July
2002.
130. K.M. Quan, E.M. Johnston, and H.A. MacKenzie. Noninvasive glucose monitoring
using nearinfrared photoacoustic spectroscopy. pages 264–264. IEEE, 28 August-
2 September 1994.
131. G. Spanner and R. Nießner. Noninvasive determination of blood constituents us-
ing an array of modulated laser diodes and a photoacoustic sensor head. Fresenius
J Anal Chem, 355:327–328, 1996.
132. G. Spanner and R. Nießner. New concept for the non-invasive determination of
physiological glucose concentrations using modulated laser diodes. Fresenius J
Anal Chem, 354:306–310, 1996.
133. R. Weiss, Y. Yegorchikov, A. Shusterman, and I. Raz. Noninvasive continuous
glucose monitoring using photoacoustic technology - Results from the first 62
subjects. Diabetes Technology & Therapeutics, 9(1):68–74, 2007.
134. S.A. Asher and J.H. Holtz. Glucose sensing intelligent polymerized crystalline
colloidal arrays. IEEE-LEOS Newsletter (Special issue), pages 32–34, April 1998.
135. A.J. Marshall, D.S. Young, J. Blyth, S. Kabilan, and C.R. Lowe. Metabolite-
sensitive holographic biosensors. Analytical Chemistry, 76:1518–1523, 2004.
136. S. Kabilan, A.J. Marshall an F.K. Sartain, M. Lee, A. Hussain, X. Yang, J. Blyth,
N. Karangu, K. James, J. Zeng, D. Smith, A. Domschke, and C.R. Lowe. Holo-
graphic glucose sensors. Biosensors and Bioelectronics, 20:1602–1610, 2005.
137. A. Domschke, W.F. March, S. Kabilan, and C. Lowe. Initial clinical testing of
a holographic non-invasive contact lens glucose sensor. Diabetes Technology &
Therapeutics, 8(1):89–93, 2006.
References 147
138. Y. Hayashi, L. Livshits, A. Caduff, and Y. Feldman. Dielectric spectroscopy study

of specific glucose influence on human erythrocyte membranes. Journal of Physics
D: Applied Physics, 36:369–374, 2003.
139. A. Sieg, R.H. Guy, and M.B. Delgado-Charro. Noninvasive and minimally invasive
methods for transdermal glucose monitoring. Diabetes Technology & Therapeutics,
7(1):174–197, 2005.
140. A. Caduff, E. Hirt, Y. Feldman, Z. Ali, and L. Heinemann. First human ex-
periments with a novel non-invasive, non-optical continuous glucose monitoring
system. Biosensors and Bioelectronics, 19:209–217, 2003.
141. G. Velho, PH. Froguel, D.R.Thevenot, and G. Reach. Strategies for calibrating a
subcutaneous glucose sensor. Biomed. Biochim Acta, 48(11/12):957–964, 1989.
142. U. Fischer, R. Ertle, K. Rebrin, E. Brunstein, H. Hahn von Dorsche, and E.J.
Freyse. Assessment of subcutaneous glucose concentration: Validation of the wick
in normal technique as a reference for implanted electrochemical sensors in normal
and diabetic dogs. Diabetologia, 30:940–945, 1987.
143. MiniMed Inc., Sylmar CA. MiniMed CGMStm continuous glucose monitoring
system User’s Guide, March 2000.
144. M.J. Tierney, J.A. Tamada, R.O. Potts, L. Jovanovic, S. Gard, and Cygnus Re-
search Team. Clinical evaluation of the glucowatch(r) biographer: a continual,
non-invasive glucose monitor for patients with diabetes. Biosensors & Bioelec-
tronics, 16:621–629, 1999.
145. John Pickup. Sensitive glucose sensing in diabetes. The Lancet, 355:426–427,
5 February 2000.
146. B. Berner, T.C. Dunn, K.C. Farinas, M.D. Garrison, R.T. Kurnik, M.J. Lesho,
R.O. Potts, J.A. Tamada, and M.J. Tierney. Signal processing for measurement
of physiological analysis, 2001. United States Patent US 6,233,471 B1.
147. R.T. Kurnik, J.J. Oliver, S.R. Waterhouse, T. Dunn, Y. Jayalakshmi, M. Lesho,
M. Lopatin, J. Tamada, C. Wei, and R.O. Potts. Application of the mixtures
of experts algorithm for signal processing in a noninvasive glucose monitoring
system. Sensors and Actuators B, 60:19–26, May 1999.
148. G.M. Steil, K. Rebrin, P.V. Goode, J.J. Mastrototaro, R.E. Purvis, W.P. van
Antwerp, and J.J. Shin. Closed loop system for controlling insulin infusion, 2000.
World Intellectual Property Organization WO00/74753 A1.
149. William L. Clarke, Daniel Cox, Linda A. Gonder-Frederick, William Carter, and
Stephen L. Pohl. Evaluating clinical accuracy of systems for self-monitoring of
blood glucose. Diabetes Care, 10(5):622–628, 1987.
150. P.D. Home. Alternative Methods, Systems and Routes of Insulin Delivery, vol-
ume 1, chapter 46, page 446. Blackwell Scientific Publications, 1991.
151. L.L. Young and M.A. Koda-Kimble. Applied Therapeutics: The clinical use of
drugs., chapter 48. Applied Therapeutics, Inc., Vancouver WA, 7th edition, 2000.
152. N. Pørksen. The in vivo regulation of pulsatile insulin secretion. Diabetologia,
45:3–20, 2002.
153. W.C. Duckworth, C.D. Saudek, and R.R. Henry. Why intraperitoneal delivery of
insulin with implantable pumps in NIDDM? Diabetes, 41:657–661, June 1992.
154. M.A. Fishel, G.S. Watson, T.J. Montine, Q. Wang, P.S. Green, J.J. Kulstad,
D.G.Cook, E.R. Peskind, L.D. Baker, D.Goldgaber, W.Nie, S. Asthana, S.R. Ply-
mate, M.W. Schwartz, and S. Craft. Hyperinsulinemia provokes synchronous
increases in central inflammation and β-amyloid in normal adults. Archives of
Neurology, 62:1539–1544, 2005.
155. J.A. Luchsinger, Ming-Xin Tang, S. Shea, and R Mayeux. Hyperinsulinemia and
risk of alzheimer disease. Neurology, 63:1187–1192, 2004.
148 References
156. I.A. Macfarlane. The Pharmacokinetics of Insulin, chapter 38, pages 371–383.
Blackwell Scientific Publications, 1991.
157. K.S. Polonsky, M.M. Byrne, and J. Sturis. Alternative insulin delivery systems:
How demanding should the patient be? Diabetologia, 40:S97–S101, 1997.
158. J.P. Mordes and A.A. Rossini. Management of diabetes in the critically ill patient.
In J.M. Rippe, R.S. Irwin, J.S. Alpert, and M.P. Fink, editors, Intensive Care
Medicine 2nd Ed., chapter 92, pages 955–962. Little, Brown and Company, 1991.
159. N.M. O’Meara and K.S. Polonsky. Insulin secretion in vivo. In C.R. Kahn and
G.C. Weir, editors, Joslin’s Diabetes Mellitus 13th Ed, chapter 5, pages 81–96.
Lea & Febiger, Malvern, Pennsylvania, 1994.
160. D. Elahi. In praise of the hyperglycemic clamp. Diabetes Care, 19(3):278–286,
1996.
161. L.C. Ramirez and P. Raskin. Pancreatic Abnormalities in Non-Insulin-Dependent
Diabetes Mellitus, volume 1, chapter 22, pages 198–204. Blackwell Scientific Pub-
lications, 1991.
162. D.S. Bell. Importance of postprandial glucose control. South Med. J., 94(8):804–
809, 2001.
163. E.W. Kraegen, D.J. Chisholm, and M.E. McNamara. Timing of insulin delivery
with meals. Horm Metab Res, 13:365–367, 1981.
164. A. Motluk and L.Geddes. Breakthrough sheds light on cause of di-
abetes, 15 December 2006. New Scientist Tech [Online] Available at
http://www.newscientisttech.com/article.ns?id=dn10812.
165. B.C. Borlase, T.J. Babineau, R.A. Forse, S.J. Bell, and G.L. Blackburn. Enteral
nutritional support. In J.M. Rippe, R.S. Irwin, J.S. Alpert, and M.P. Fink, editors,
Intensive Care Medicine 2nd Ed., chapter 174, pages 1669–1674. Little, Brown
and Company, 1991.
166. D. Dazzi, F. Taddei, A. Gavarini, E. Uggeri, R. Negro, and A. Pezzarossa. The
control of blood glucose in the critical diabetic patient: A neuro-fuzzy method.
Journal of Diabetes and Complications, 15:80–87, 2001.
167. F. Chee, T. Fernando, and P.V. van Heerden. Closed-loop glucose control in
critically-ill patients using continuous glucose monitoring system (CGMS) in real-
time. IEEE Trans. Inf. Tech. Biomed., 7(1):43–53, March 2003.
168. E. R. Carson, C. Cobelli, and L. Finkelstein. The Mathematical Modelling of
Metabolic and Endocrine Systems - Model Formulation, Identification, and Vali-
dation, volume 2 of Biomedical Engineering and Health Systems. John-Wiley and
Sons, 1983.
169. Y. Cherruault. Mathematical Modelling in Biomedicine. D. Reidel Publishing
Company, Dordrecht, Holland, 1986.
170. J.C. Henquin. Cell biology of insulin secretion. In C. R. Kahn and G.C. Weir,
editors, Joslin’s Diabetes Mellitus 13th Ed, pages 56–80. Lea & Febiger, Malvern,
Pennsylvania, 1994.
171. R.A. Rizza, L.J. Mandarino, and J.E. Gerich. Dose-response characteristics for
effects of insulin on production and utilization of glucose in man. Am. J. Physiol.
(Endocrinol. Metab), 240:E630–E639, 1981.
172. A.M. Albisser, B.S. Leibel, T.G. Ewart, Z. Davidovac, C.K. Botz, and W. Zingg.
An artificial endocrine pancreas. Diabetes, 23(5):389–396, May 1974.
173. E.W. Kraegen, R. Whiteside, D. Bell, Y.O. Chia, and L. Lazarus. Development
of a closed-loop artificial pancreas. Horm Metab Res Suppl, 8:39–42, 1979.
174. J.D. Kruse-Jarres, G. Braun, R. Naegeie, M. Bresch, and U. Lehmann. Blood
glucose monitoring and computer regulation by means of an artificial endocrine
pancreas. Horm Metab Res Suppl, 8:43–45, 1979.
References 149
175. N.H. White, D. Skor, and J.V. Santiago. Practical closed-loop insulin delivery: A
system for the maintenance of overnight euglycemia and the calculation of basal
insulin requirements in insulin-dependent diabetes. Ann Intern Med, 97:210–213,
1982.
176. S.M. Furler, E.W. Kraegen, R.H. Smallwood, and D.J. Chisolm. Blood glucose
control by intermittent loop closure in the basal mode: Computer simulation
studies with a diabetic model. Diabetes Care, 8(6):553–561, 1985.
177. R.L. Ollerton. Application of optimal control theory to diabetes mellitus. Int. J.
Control, 50(6):2503–2522, 1989.
178. P. Palazzo and V. Viti. A new glucose-clamp algorithm theoretical consideration
and computer simulations. IEEE Trans. Biomed. Eng., 37(5):535–543, 1990.
179. F. Chee, T.L. Fernando, A.V. Savkin, and P.V. van Heerden. Expert PID control
system for blood glucose control in critically-ill patients. IEEE Trans. Inf. Tech.
Biomed., 7(4):419–425, December 2003.
180. G.M. Steil, A.E. Panteleon, and K. Rebrin. Closed-loop insulin delivery - the
path to physiological glucose control. Advanced Drug Delivery Reviews, 56:125–
144, 2004.
181. E. Renard, G. Costalat, H. Chevassus, and J. Bringer. Closed loop insulin delivery
using implanted insulin pumps and sensors in type 1 diabetic patients. Diabetes
Research and Clinical Practice, 74:S173–S177, 2006.
182. B. Gopakumaran, H.M. Duman, D.P. Overholser, I.F. Federiuk, M.J. Quinn, M.D.
Wood, and W.K. Ward. A novel insulin delivery algorithm in rats with type 1
diabetes: The fading memory proportional-derivative method. Artifical Organs,
29(8):599–607, 2005.
183. G. Marchetti, M. Barolo, L. Jovanovic, and H. Zisser. An improved PID switching
control strategy for Type 1 diabetes. In Proceedings of the 28th IEEE EMBS An-
nual International Conference, pages 5041–5044, New York City, USA, 30 August-
3 September 2006. IEEE.
184. V.W. Bolie. Coefficients of normal blood glucose regulation. Journal of Applied
Physiology, 16(5):783–788, 1961.
185. E. Ackerman, J.W. Rosevear, and W.F. McGuckin. A mathematical model of the
glucose tolerance test. Physics in Medicine and Biology, 9(2):203–213, 1964.
186. E. Ackerman, Laël.C. Gatewood, J.W. Rosevear, and G.D. Molnar. Model studies
of blood-glucose regulation. Bulletin of Mathematical Biophysics, 27:21–37, 1965.
Special Issue.
187. L.C. Gatewood, E. Ackerman, J.W. Rosevear, and G.D. Molnar. Test of a mathe-
matical model of the blood-glucose regulatory system. Computers and Biomedical
Research, 2:1–14, 1968.
188. F. Ceresa, F. Ghermi, P. F. Martini, P. Martino, G. Segre, and A. Vitelli. Control
of blood glucose in normal and in diabetic subjects - studies by compartmental
analysis and digital computer technics. Diabetes, 17(9):571–578, 1968.
189. T. Chorbajian. An in numero study of glucose-insulin interaction. Bulletin of
Mathematical Biophysics, 33:451–462, 1971.
190. E. Cerasi, G. Fick, and M. Rudemo. A mathematical model for the glucose
induced insulin release in man. European Journal of Clinical Investigation, 4:269–
278, 1974.
191. M. Nomura, M. Shichiri, R. Kawamori, Y. Yamasaki, N. Iwama, and H. Abe. A
mathematical insulin-secretion model and its validation in isolated rat pancreatic
islets perifusion. Computers and Biomedical Research, 17:570–579, 1984.
150 References
192. J. S. Bajaj, G. Subba Rao, J. Subba Rao, and R. Khardori. A mathematical

model for insulin kinetics and its application to protein-deficient (malnutrition-
related) diabetes mellitus (PDDM). Journal of Theoretical Biology, 126:491–503,
1987.
193. G. Subba Rao, J. S. Bajaj, and J. Subba Rao. A mathematical model for insulin
kinetics II. extension of the model to include response to oral glucose administra-
tion and application to insulin-dependent diabetes mellitus (IDDM). Journal of
Theoretical Biology, 142:473–483, 1990.
194. L. Jansson, L. Lindskog, N. E. Norden, S. Carlstrom, and B. Schersten. Diagnostic
value of the oral glucose tolerance test evaluated with a mathematical model.
Computers and Biomedical Research, 13:512–521, 1980.
195. E. Salzsieder, G. Albrecht, U. Fischer, and E.-J Freyse. Kinetic modeling of
glucoregulatory system to improve insulin therapy. IEEE Trans. Biomed. Eng.,
32(10):846–855, October 1985.
196. R.N. Bergman and J. Urquhart. The pilot gland approach to the study of insulin
secretory dynamics. Recent Prog Horm Res, 27:583–605, 1971.
197. R. N. Bergman, L. S. Phillips, and C. Cobelli. Physiological evaluation of factors
controlling glucose tolerance in man. Journal of Clinical Investigation, 68:1456–
1467, December 1981.
198. G. Toffolo, R. N. Bergman, D. T. Finegood, C. R. Bowden, and C. Cobelli. Quan-
titative estimation of beta cell sensitivity to glucose in the intact organism - a
minimal model of insulin kinetics in the dog. Diabetes, 29:979–990, December
1980.
199. B. Candas and J. Radziuk. An adaptive plasma glucose controller based on
a nonlinear insulin/glucose model. IEEE Trans. Biomed. Eng., 41(2):116–124,
February 1994.
200. F.J. Doyle III, C. Dorski, J. Harting, and N.A. Peppas. Control and modelling of
drug delivery devices for the treatment of diabetes. In Proceedings of the American
Conference, volume 1, pages 776–780, Seattle, Washington, USA, 21-23 June 1995.
IEEE.
201. A. De Gaetano and O. Arino. Mathematical modeling of the intravenous glucose
tolerance test. Journal of Mathematical Biology, 40:136–168, 2000.
202. C. Cobelli, G. Federspil, G. Pacini, A. Salvan, and C. Scandellari. An integrated
mathematical model of the dynamics of blood glucose and its hormonal control.
Mathematical Biosciences, 58:27–60, 1982.
203. C. Cobelli and A. Mari. Control of diabetes with artificial systems for insulin
delivery - algorithm independent limitations revealed by a modeling study. IEEE
Trans. Biomed. Eng., 32(10):840–845, October 1985.
204. C. Cobelli, G. Pacini, G. Toffolo, and L. Sacca. Estimation of insulin sensitivity
and glucose clearance from minimal model: new insights from labeled IVGTT.
American Journal of Physiology, 250:E591–E598, 1986.
205. D. K. Parrish and D. B. Ridgely. Control of artificial human pancreas using
the SDRE method. In Proceedings of the American Control Conference, pages
1059–1060, Albquerque, New Mexico, June 1997. Academic Press.
206. R. S. Parker, F. J. Doyle III, and N. A Peppas. A model-based algorithm for
blood glucose control in type I diabetic patients. IEEE Trans. Biomed. Eng.,
46(2):148–157, February 1999.
207. R.S. Parker, F.J. Doyle III, J.H. Ward, and N.A. Peppas. Robust H ∞ glucose
control in diabetes using a physiological model. AIChE Journal, 46(12):2537–
2549, December 2000.
References 151
208. E. RuizVelázquez, R. Femat, and D.U. CamposDelgado. Blood glucose control

for type I diabetes mellitus: A robust H ∞ tracking problem. Control Engineering
Practice, 12:1179–1195, 2004.
209. G.M. Steil, B. Clark, S. Kanderian, and K. Rebrin. Modelling insulin action for
development of close-loop artificial pancreas. Diabetes Technology & Therapeutics,
7(1):94–108, 2005.
210. R. Hovorka, F. Shojaee-Moradie, P.V. Carroll, L.J. Chassin, I.J. Gowrie, N.C.
Jackson, R.S. Tudor, A.M. Umpleby, and R.H. Jones. Partitioning glucose dis-
tribution/transport, disposal and endogenous production during IVGTT. Am J
Physiol Endocrinol Metab, 282:992–1007, 2002.
211. R. Hovorka, V. Canonico, L.J. Chassin, U. Haueter, M. Massi-Benedetti, M.O.
Federici, T.R. Pieber, H.C. Schaller, L. Schaupp, T. Vering, and M.E. Wilinska.
Nonlinear model predictive control of glucose concentration in subjects with type
1 diabetes. Physiological Measurement, 25:905–920, 2004.
212. M. Giugliano, M. Bove, and M. Grattarola. Insulin release at the molecular level:
Metabolic-electrophysiological modeling of the pancreatic beta-cells. IEEE Eng.
Med. Biol. Mag., 47(5):611–623, May 2000.
213. K. Ogata. Modern Control Engineering. Prentice-Hall International, Inc, Upper
Saddle River, New Jersey, 3rd edition, 1997.
214. U. Fischer, W. Schenk, E. Salzsieder, G. Albrecht, P. Abel, and E.-J Freyse. Does
physiological blood glucose control require an adaptive control strategy? IEEE
Trans. Biomed. Eng., 34(8):575–582, Aug 1987.
215. C. Dalla Man, G. Toffolo, and C. Cobelli. Gluocse kinetics during meal:one vs two
compartment minimal model. In Proceedings of the Second Joint EMBS/BMES
Conference, pages 2222–2223, Houston, TX, USA, 23 -26 October 2002. IEEE.
216. C. Dalla Man, R.A. Rizza, and C. Cobelli. Mixed meal simulation model of
glucose-insulin system. In Proceedings of the 28th IEEE EMBS Annual Interna-
tional Conference, pages 307–310, New York City, USA, 30 August-3 September
2006. IEEE.
217. J.T. Sorensen, C.K. Colton, R.S. Hillman, and J.S. Soeldner. Use of a physio-
logic pharmacokinetic model of glucose homeostasis for assessment of performance
requirements for improved insulin therapies. Diabetes Care, 5(3):148–157, 1982.
218. M. Ader, G. Pacini, Y.J. Yang, and R. Bergman. Importance of glucose per se to
intravenous glucose tolerance. Diabetes, 34:1093–1103, 1985.
219. F. J. Vega-Catalan. A program to estimate insulin sensitivity and pancreatic
responsivity from an IVGTT using the minimal modeling technique. Computers
and Biomedical Research, 23:1–9, 1990.
220. T. Van Herpe, B. Pluymers, M. Espinoza, G. Van den Berghe, and B. De Moor.
A minimal model for glycemia control in critically ill patients. pages 5432–5435.
IEEE, 30 August-3 September 2006.
221. R. Hovorka, L.J. Chassin, M.E. Wilinska, V. Canonico, J.A. Akwi, M.O. Federici,
M. Massi-Benedetti, I. Hutzli, C. Zaugg, H. Kaufmann, M. Both, T. Vering, H.C.
Schaller, L. Schaupp, M. Bodenlenz, and T.R. Pieber. Closing the loop: The
adicol experience. Diabetes Technology & Therapeutics, 6(3):307–318, 2004.
222. J.T. Sorensen. A Physiological Model of Glucose Metabolism in Man and it Use to
Design and Access Improved Insulin Therapies for Diabetes. PhD thesis, Dept.of
Chemical Engineering, MIT, 1985.
223. S.M. Lynch and B.W. Bequette. Model predictive control of blood glucose in
type i diabetes using subcutaneous glucose measurements. In Proceedings of the
American Control Conference, Anchorage, AK, 8 -10 May 2002.
152 References
224. T.L Yates and L.R. Fletcher. Prediction of a glucose appearance function from
food using deconvolution. IMA Journal of Mathematics Applied in Medicine and
Biology, 17:169–184, 2000.
225. Z. Trajanoski and P. Wach. Neural predictive controller for insulin delivery using
the subcutaneous route. IEEE Trans. Biomed. Eng., 45(9):1122–1134, 1998.
226. N.L. Ricker. Model predictive control with state estimation. Ind. Eng. Chem.
Res, 29:374–382, 1990.
227. M. E. Fisher and K. L. Teo. Optimal insulin infusion resulting from a mathemat-
ical model of blood glucose dynamics. IEEE Trans. Biomed. Eng., 36(4):479–486,
April 1989.
228. G.W. Swan. An optimal control model of diabetes mellitus. Bulletin of Mathe-
matical Biology, 44(6):793–808, 1982.
229. M. Gopal. Modern Control System Theory. Wiley Eastern Ltd, New Delhi, 1984.
230. Anthony Zaknich. Principles of Adaptive Filters and Self-learning Systems. ISBN
1-85233-984-5. Springer-Verlag, 2005. Series on Advanced Textbooks in Control
and Signal Processing.
231. B. Pagurek, J.S. Riordon, and S. Mahmoud. Adaptive control of the human
glucose-regulatory system. Med. & Biol. Eng, 10:753–761, 1972.
232. M. Kikuchi, E. Machiyama, N. Kabei, and A. Yamada. Adaptive control system
of blood glucose regulation. In Lindberg/Kaihara, editor, MEDINFO 80, pages
96–100. IFIP, North-Holland Publishing Company, 1980.
233. E. Sarti and P. Cruciani. Self-tuning control algorithm for wearable artificial
pancreas. In Vol. 14. Proceedings of the Annual International Conference of the
IEEE, volume 6, pages 2267–2269. Engineering in Medicine and Biology Society,
29 October-1 November 1992.
234. M. Morari and N.L. Ricker. Model Predictive Control Toolbox User’s Guide. The
Mathworks, Inc, Natick, MA, 1998.
235. C.E. Garcı́a, D.M. Prett, and M. Morari. Model predictive control: Theory and
practice - a survey. Automatica, 25(3):335–348, 1989.
236. R. Bellazzi, G. Nucci, and C. Cobelli. The subcutaneous route to insulin-
dependent diabetes therapy. IEEE Eng. Med. Biol. Mag., 20(1):54–64, January-
February 2001.
237. R.T. Stefani, B. Shahian, C.J. Savant, and G.H. Hostetter. Design of Feedback
Control Systems. Oxford University Press, New York, 4th edition, 2002.
238. K. Glover and J.C. Doyle. State-space formulae for all stabilizing controllers that
satisfy and H∞ -norm bound and relations to risk sensitivity. System & Control
Letters, 11:167–172, 1988.
239. W.S. Levine and R.T. Reichert. An introduction to H∞ control system design. In
Proceedings of the 29th Conference on Decision and Control, Honolulu, Hawaii,
December 1990.
240. J.C. Doyle, K. Glover, P.P. Khargonekar, and B.A. Francis. State-space solutions
to standard H2 and H∞ control problems. IEEE Transactions on Automatic
Control, 34(8):831–847, 1989.
241. H. Toivonen. Robust control method, AS-74.330 Postgraduate Course in Control
Engineering, 2002. Process Control Laboratory, Abo Akademi [Online] Available
at http://www.abo.fi/∼htoivone/courses/AS74330.html.
242. K. H. Kienitz and T. Yoneyama. A robust controller for insulin pump based on
H ∞ theory. IEEE Trans. Biomed. Eng., 40(11):1133–1137, December 1993.
243. F. Chee, A.V. Savkin, T.L. Fernando, and S. Nahavandi. Optimal H ∞ insulin
injection control for blood glucose regulation in diabetic patients. IEEE Trans.
Biomed. Eng., 52(10):1625–1631, October 2005.
References 153
244. A.V. Savkin and R.J. Evans. A new approach to robust control of hybrid systems
over infinite time. IEEE Transactions on Automatic Control, 43(9):1292–1296,
1998.
245. A.V. Savkin and I.R. Petersen. Robust filtering and model validation for uncer-
tain continuous-time systems with discrete and continuous measurements. Inter-
national Journal of Control, 69(1):163–174, 1998.
246. A.V. Savkin and R.J. Evans. Hybrid Dynamical Systems. Controller and Sensor
Switching Problems. Birkhäuser, Boston, Cambridge MA, 1st edition, 2002.
247. D.P. Bertsekas and I.B. Rhodes. Recursive state estimation for a set-membership
description of uncertainty. IEEE transactions on automatic control, AC-
16(2):117–128, April 1971.
248. D.A. Gough, S. Aisenberg, C.K. Colton, J.Giner, and J.S. Soeldner. The status of
electrochemical sensors for in vivo glucose monitoring. Horm Metab Res Suppl,
7:10–22, 1977.
249. R.D. Beach, F. V. Kuster, and F. Moussy. Subminiature implantable potentiostat
and modified commercial telemetry device for remote glucose monitoring. IEEE
Transaction in Instrumentation and Measurement, 48(6):1239–1245, 1999.
250. K.C. McIvor, J.L. Cabernoch, N.M. Van Antwerp, E.C. Halili, and J.J. Mastro-
totaro. Glucose sensor package system, 2002. US Patent 6,360,888 B1.
251. S.J. Updike, M.C. Shults, B.J. Gilligan, and R.K. Rhodes. A subcutaneous glu-
cose sensor with improved longevity, dynamic range, and stability of calibration.
Diabetes Care, 23(2):208–214, 2000.
252. T. Yipintsoi, L.C. Gatewood, E. Ackerman, P.L. Spivak, G.D. Molnar, J.W. Ro-
sevear, and F.J. Service. Mathematical analysis of blood glucose and plasma
insulin responses to insulin infusion in healthy and diabetic subjects. Computers
in biology & medicine, 3:71–78, 1973.
Index
absorption coefficient, 33 cessation of blood flow, 26

acetaminophen, 12 compartmental analysis, 72
ADICOL, 81 Concanavalin A, 10
affinity binding, 9 control
air-in-line, 128 adaptive, 93
Albisser et al, 62 H infinity, 104
amperometry, 8, 121 lookup table, 63
anticipatory phase, 52, 53 model predictive, 83, 96
arteriosclerosis, 46 neural predictive, 97
Attenuated Total Reflection, 19, 21 neuro-fuzzy, 66
optimal, 90
back-scattered, 24, 25 P-D, 62, 68
back-scattering, 19 P-I-D, 66, 68
baseline background, 40 Position, 67
Bayesian estimation, 98 Velocity, 67
Beer-Lambert law, 16 titration, 64
beta cells, 51, 54 cortisol, 56
biofouling, 11, 13 crystalline colloidal array, 35
Biostator, 63
dextran
biphasic, 52
definition, 10
birefringence, 32
dextro-rotatory, 30
blackbody radiation, 19
diastole, 26
blank subtraction, 41
dielectric spectroscopy, 35
Bovine Serum Albumin, 9
diffuse reflectance, 23
Bragg’s law, 34
diffusion, 12, 14, 23, 44, 46
dimers, 44
calibration dose-response curve, 63
multi-point, 39
multivariate, 15 enantiomer, 30
one-point, 36 epinephrine, 56
Regression, 38 evanescent wave, 21
two-point, 36
Wick, 38 finger-print region, 18
cavitation, 14 FITC-dextran, 12
156 Index
bound, 11 light diffraction, 34

unbound, 10 light scattering, 15, 17, 22, 32
fluorophore, 10, 11 Linear Fractional Transformation, 103
free fatty acids, 51 LTSS, 68
FRET, 10
FTIR-spectrometer, 19 Matrix Inversion Lemma, 134
mitochondria, 23
gluconeogenesis, 1, 51, 52, 56 model
gluconolactone, 8, 11 Ackerman, 73, 91
glucose tolerance test, 75 Bolie, 74
glucose-clamp, 53, 96 Cobelli, 77
glucose-insulin response curve, 62 Hovorka, 81
glucose-oxidase, 8, 9, 11, 13, 35 Minimal, 75
glutaraldehyde, 9 Sorensen, 84
glycogenesis, 2 model linearisation, 104
glycogenolysis, 1, 51 model reduction, 104
glycosuria, 54 molar absorptivity, 16
molar extinction coefficient, 16
hexamers, 44 molecular vibrations, 16
holographic, 35 monomers, 44
hyperinsulinemia, 46, 47 moving average filtering, 125
hypermetabolic, 55 moving horizon, 97
hypoglycaemic coma, 3 multiple scattering, 22, 25
IMED, 114 net hepatic glucose balance, 77

immobilised, 6, 9 Neural network, 18
impedance spectroscopy, 35 Back Error Propagation, 66
InGaAs, 24 occlusion, 128
InSb, 19, 24 Occlusion spectroscopy, 26
insulin optical
Lispro, 44, 83 absorbance, 16, 19
negative, 91 activity, 28
insulin-dependent modulator, 31
fractional removal rate, 81 rotatory dispersion, ORD, 30
glucose utilization, 77 spectroscopy, 15
insulin-independent window, 16
fractional removal rate, 81 Optical Coherence Tomography, 24, 25
glucose flux, 82 osmolarity, 23
glucose uptake, 79 overtone, 16, 17
glucose utilization, 77 oxidation, 11, 12
interference
water, 16 Parseval’s theorem, 103
interferogram, 19 peristaltic, 57
internal model, 90 photoacoustic, 33
intra-ocular lens, 11 photothermal, 33
intubation, 57 plant, 90, 101
Princple of Superposition, 71
Kalman filtering, 97 Pulse glucometry, 26
Langerhans, 51 Raman shift , 27
least square estimation, 93 Raman spectroscopy, 26
Index 157
Random Variable Transformation, 98 sliding scale, 63, 66, 116

redox mediator, 11 Sodium D line, 30
refractive index , 22 specific rotation, 29
reverse iontophoresis, 13, 39 stress, 55, 57
Riccati equation, 109, 134 suction, 15
sweat, 14, 35
saline, 59 systole, 26
scattering
coefficient, 23 temperature gradient, 21
elastic, 26 thermal expansion coefficient, 33
inelastic, 27 total parenteral nutrition, 57, 118
Mie, 22 transflectance, 19
Raman, 27 transimpedance, 123
anti-Strokes shift, 27 tympanic membrane, 21
Strokes shift, 27
Rayleigh, 26, 27 ultrasound, 14
scattering centers, 22
silver halide, 35 vasopressors, 57
Printing: Mercedes-Druck, Berlin
Binding: Stein + Lehmann, Berlin
Lecture Notes in Control and Information Sciences
Edited by M. Thoma, M. Morari
Further volumes of this series can be found on our homepage:
springer.com
Vol. 368: Chee F.; Fernando T. Vol. 356: Karatkevich A.:

Closed-Loop Control of Blood Glucose Dynamic Analysis of Petri Net-Based Discrete
157 p. 2007 [978-3-540-74030-8] Systems
Vol. 367: Turner M.C.; Bates D.G. (Eds.) 166 p. 2007 [978-3-540-71464-4]
Mathematical Methods for Robust and Nonlinear Vol. 355: Zhang H.; Xie L.:
Control Control and Estimation of Systems with
444 p. 2007 [978-1-84800-024-7] Input/Output Delays
Vol. 366: Bullo F.; Fujimoto K. (Eds.) 213 p. 2007 [978-3-540-71118-6]
Lagrangian and Hamiltonian Methods For Non- Vol. 354: Witczak M.:
linear Control Modelling and Estimation Strategies for Fault
398 p. 2007 [978-3-540-73889-3] Diagnosis of Non-Linear Systems
Vol. 365: Bates D.; Hagström M. (Eds.) 215 p. 2007 [978-3-540-71114-8]
Nonlinear Analysis and Synthesis Techniques for Vol. 353: Bonivento C.; Isidori A.; Marconi L.;
Aircraft Control Rossi C. (Eds.)
360 p. 2007 [978-3-540-73718-6] Advances in Control Theory and Applications
Vol. 364: Chiuso A.; Ferrante A.; 305 p. 2007 [978-3-540-70700-4]
Pinzoni S. (Eds.) Vol. 352: Chiasson, J.; Loiseau, J.J. (Eds.)
Modeling, Estimation and Control Applications of Time Delay Systems
356 p. 2007 [978-3-540-73569-4] 358 p. 2007 [978-3-540-49555-0]
Vol. 363: Besançon G. (Ed.) Vol. 351: Lin, C.; Wang, Q.-G.; Lee, T.H., He, Y.
Nonlinear Observers and Applications LMI Approach to Analysis and Control of
224 p. 2007 [978-3-540-73502-1] Takagi-Sugeno Fuzzy Systems with Time Delay
204 p. 2007 [978-3-540-49552-9]
Vol. 362: Tarn T.-J.; Chen S.-B.;
Zhou C. (Eds.) Vol. 350: Bandyopadhyay, B.; Manjunath, T.C.;
Robotic Welding, Intelligence and Automation Umapathy, M.
562 p. 2007 [978-3-540-73373-7] Modeling, Control and Implementation of Smart
Structures 250 p. 2007 [978-3-540-48393-9]
Vol. 361: Méndez-Acosta H.O.; Femat R.;
González-Álvarez V. (Eds.): Vol. 349: Rogers, E.T.A.; Galkowski, K.;
Selected Topics in Dynamics and Control of Owens, D.H.
Chemical and Biological Processes Control Systems Theory
320 p. 2007 [978-3-540-73187-0] and Applications for Linear
Repetitive Processes 482
Vol. 360: Kozlowski K. (Ed.)
p. 2007 [978-3-540-42663-9]
Robot Motion and Control 2007
452 p. 2007 [978-1-84628-973-6] Vol. 347: Assawinchaichote, W.; Nguang, K.S.;
Vol. 359: Christophersen F.J. Shi P.
Optimal Control of Constrained Fuzzy Control and Filter Design
Piecewise Affine Systems for Uncertain Fuzzy Systems
190 p. 2007 [978-3-540-72700-2] 188 p. 2006 [978-3-540-37011-6]
Vol. 358: Findeisen R.; Allgöwer Vol. 346: Tarbouriech, S.; Garcia, G.; Glattfelder,
F.; Biegler L.T. (Eds.): Assessment and Future Di- A.H. (Eds.)
rections of Nonlinear Advanced Strategies in Control Systems
Model Predictive Control with Input and Output Constraints
642 p. 2007 [978-3-540-72698-2] 480 p. 2006 [978-3-540-37009-3]
Vol. 357: Queinnec I.; Tarbouriech Vol. 345: Huang, D.-S.; Li, K.; Irwin, G.W. (Eds.)
S.; Garcia G.; Niculescu S.-I. (Eds.): Intelligent Computing in Signal Processing
Biology and Control Theory: Current Challenges and Pattern Recognition
589 p. 2007 [978-3-540-71987-8] 1179 p. 2006 [978-3-540-37257-8]
Vol. 327: Fournier, J.-D.; Grimm, J.; Leblond, J.;
Vol. 344: Huang, D.-S.; Li, K.; Irwin, G.W. (Eds.) Partington, J.R. (Eds.)
Intelligent Control and Automation Harmonic Analysis and Rational Approximation
1121 p. 2006 [978-3-540-37255-4] 301 p. 2006 [978-3-540-30922-2]
Vol. 341: Commault, C.; Marchand, N. (Eds.) Vol. 326: Wang, H.-S.; Yung, C.-F.; Chang, F.-R.
Positive Systems H∞ Control for Nonlinear Descriptor Systems
448 p. 2006 [978-3-540-34771-2] 164 p. 2006 [978-1-84628-289-8]
Vol. 340: Diehl, M.; Mombaur, K. (Eds.) Vol. 325: Amato, F.
Fast Motions in Biomechanics and Robotics Robust Control of Linear Systems Subject to
500 p. 2006 [978-3-540-36118-3] Uncertain
Vol. 339: Alamir, M. Time-Varying Parameters
Stabilization of Nonlinear Systems Using 180 p. 2006 [978-3-540-23950-5]
Receding-horizon Control Schemes Vol. 324: Christofides, P.; El-Farra, N.
325 p. 2006 [978-1-84628-470-0] Control of Nonlinear and Hybrid Process Systems
Vol. 338: Tokarzewski, J. 446 p. 2005 [978-3-540-28456-7]
Finite Zeros in Discrete Time Control Systems Vol. 323: Bandyopadhyay, B.; Janardhanan, S.
325 p. 2006 [978-3-540-33464-4] Discrete-time Sliding Mode Control
Vol. 337: Blom, H.; Lygeros, J. (Eds.) 147 p. 2005 [978-3-540-28140-5]
Stochastic Hybrid Systems Vol. 322: Meurer, T.; Graichen, K.; Gilles, E.D.
395 p. 2006 [978-3-540-33466-8] (Eds.)
Vol. 336: Pettersen, K.Y.; Gravdahl, J.T.; Control and Observer Design for Nonlinear Finite
Nijmeijer, H. (Eds.) and Infinite Dimensional Systems
Group Coordination and Cooperative Control 422 p. 2005 [978-3-540-27938-9]
310 p. 2006 [978-3-540-33468-2] Vol. 321: Dayawansa, W.P.; Lindquist, A.;
Vol. 335: Kozłowski, K. (Ed.) Zhou, Y. (Eds.)
Robot Motion and Control New Directions and Applications in Control
424 p. 2006 [978-1-84628-404-5] Theory
400 p. 2005 [978-3-540-23953-6]
Vol. 334: Edwards, C.; Fossas Colet, E.;
Fridman, L. (Eds.) Vol. 320: Steffen, T.
Advances in Variable Structure and Sliding Mode Control Reconfiguration of Dynamical Systems
Control 290 p. 2005 [978-3-540-25730-1]
504 p. 2006 [978-3-540-32800-1] Vol. 319: Hofbaur, M.W.
Vol. 333: Banavar, R.N.; Sankaranarayanan, V. Hybrid Estimation of Complex Systems
Switched Finite Time Control of a Class of 148 p. 2005 [978-3-540-25727-1]
Underactuated Systems Vol. 318: Gershon, E.; Shaked, U.; Yaesh, I.
99 p. 2006 [978-3-540-32799-8] H∞ Control and Estimation of State-multiplicative
Vol. 332: Xu, S.; Lam, J. Linear Systems
Robust Control and Filtering of Singular Systems 256 p. 2005 [978-1-85233-997-5]
234 p. 2006 [978-3-540-32797-4] Vol. 317: Ma, C.; Wonham, M.
Vol. 331: Antsaklis, P.J.; Tabuada, P. (Eds.) Nonblocking Supervisory Control of State Tree
Networked Embedded Sensing and Control Structures
367 p. 2006 [978-3-540-32794-3] 208 p. 2005 [978-3-540-25069-2]
Vol. 330: Koumoutsakos, P.; Mezic, I. (Eds.) Vol. 316: Patel, R.V.; Shadpey, F.
Control of Fluid Flow Control of Redundant Robot Manipulators
200 p. 2006 [978-3-540-25140-8] 224 p. 2005 [978-3-540-25071-5]
Vol. 329: Francis, B.A.; Smith, M.C.; Willems, Vol. 315: Herbordt, W.
J.C. (Eds.) Sound Capture for Human/Machine Interfaces:
Control of Uncertain Systems: Modelling, Practical Aspects of Microphone Array
Approximation, and Design Signal Processing
429 p. 2006 [978-3-540-31754-8] 286 p. 2005 [978-3-540-23954-3]
Vol. 328: Loría, A.; Lamnabhi-Lagarrigue, F.; Vol. 314: Gil’, M.I.
Panteley, E. (Eds.) Explicit Stability Conditions for
Advanced Topics in Control Systems Theory Continuous Systems
305 p. 2006 [978-1-84628-313-0] 193 p. 2005 [978-3-540-23984-0]

978 3 540 74031 5

Uploaded by

Copyright:

Available Formats

978 3 540 74031 5

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

978 3 540 74031 5

Uploaded by

Copyright:

Available Formats

Lecture Notes

in Control and Information Sciences 368

Closed-Loop Control of Blood

Library of Congress Control Number: 2007932674

ISSN print edition: 0170-8643

To my wife Yasmin and two daughters Melissa and Rochelle

This would help in understanding the nature of the glucose-regulatory prob-

Perth Frederick Chee

2 Glucose Control: Input and Output . . . . . . . . . . . . . . . . . . . . . . . . . . 5

2.5.3 GlucoWatch Calibration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38

3 Glucose Control: Patient Dynamics . . . . . . . . . . . . . . . . . . . . . . . . . . 49

4 Mathematics of Glucose Control . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59

5 Closed-Loop Control Apparatus Example . . . . . . . . . . . . . . . . . . . . 109

A Mathematical Derivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 129

B Model Parameters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 133

Hyperglycaemia refers to an elevated glucose concentration in the circulating

1.2 Causes of Hyperglycaemia

1.2.1 Physiological Background

toward normal. Conversely, a decrease in blood glucose stimulates glucagon se-

1.2.2 Historical Background

1.2.3 Diabetic Patients

1.2.4 Surgical Patients

Surgical patients commonly enter a hypermetabolic stress state, induced by the

1.3 Importance of Tighter BG Level Control

BG level should be kept within the normal range, because:

long-term complications (in diabetic patients) and infectious complications (in

1.4 Achieving Tighter Control

Sensor Controller Pump

As the variable to be controlled is a patient’s BG level, a knowledge of the

2.2 BG Measurement – Invasive Techniques

2.3 BG Measurement – Minimally-Invasive Techniques

In “implantation”, miniaturised glucose sensors were inserted directly into the

Amperometric Glucose Sensor

Among the chemically-based methods of glucose sensing, the amperometric sen-

Table 2.1. Glucose-oxidase reaction

Glucose oxidase dehydrogenates glucose (C6 H12 O6 ), transforming it to glu-

Glucose + GO(F AD) −→ gluconolactone + GO(F ADH2 )

1. Gluconolactone reacts with water to form gluconic acid (C6 H12 O7 ),

Gluconolactone + H2 O −→ C6 H12 O7 ⇐⇒ C6 H11 O7− + H +

2. Glucose oxidase is then oxidised in the presence of oxygen, and hydrogen

GO(F ADH2 ) + O2 −→ GO(F AD) + H2 O2

By applying -600 mV to -700 mV (with respect to the working platinum

When amperometric sensors are used as implantable sensors, enzyme immo-

Implantable amperometric sensors have only recently become available as an

Fluorescence-based glucose sensor involves measuring a change in light emission

3. Ballerstadt et al [42] has proposed subcutaneous implantation of a self-

• Glucose oxidase catalytic reaction can be monitored by detecting oxygen con-

There are other ﬂuorescence and aﬃnity-binding methods in existence. Interested

Advantages of Implantable Sensors

Disadvantages of Implantable Sensors

2.3.2 Fluid Extraction

Some existing ﬂuid extraction techniques include:

Microdialysis is a method that allows blood or interstitial ﬂuid (ISF) to be

Advantages with Microdialysis

• Microdialysis methods prevent many of the other substances present in the

Disadvantages with Microdialysis

• Although microdialysis ﬁlters oﬀ much of the unwanted substances preventing

Reverse iontophoresis involves the application (or conduction) of a constant,

Advantages with Reverse Iontophoresis