Mousa Thesis

AMERICAN UNIVERSITY OF BEIRUT
DESIGN OF A MICROWAVE SENSOR FOR NON-INVASIVE

BLOOD GLUCOSE MONITORING
by
MOUSSA SEMAAN BTEICH
A thesis
submitted in partial fulfillment of the requirements
for the degree of Master of Engineering
to the Department of Electrical and Computer Engineering
of the Maroun Semaan Faculty of Engineering and Architecture
at the American University of Beirut
Beirut, Lebanon
April 2019
ACKNOWLEDGMENTS
I would like to genuinely acknowledge my thesis advisors Dr. Joseph Costantine

and Dr. Rouwaida Kanj for their support, advice, and encouragement to complete my
thesis. I recognize Dr. Joseph for helping me to improve my research skills during my
work on this thesis. I thank Dr. Rouwaida Kanj for her contribution to the data analytic
part.
I would like to thank Dr. Assad Eid, Dr. Ali Ramadan and Dr. Youssef Tawk for
their continuous input and for being on my thesis committee. I would also like to thank
Dr. Zaher Dawy for being on my thesis committee.
I would like to thank my colleagues for their help and support during my
research work.
Finally, I would like to thank my parents, brother, sister, and friends for
supporting and encouraging me during my studies.
v
AN ABSTRACT OF THE THESIS OF
Moussa Semaan Bteich for Master of Engineering

Major: Electrical and Computer Engineering
Title: Design of a Microwave Sensor for Non-Invasive Blood Glucose Monitoring
Diabetes is one of the most prevalent chronic diseases. The exponential rate of
increase in the number of diabetics urged researchers to search for new methods of
measuring blood glucose continuously and non-invasively. The ability of microwave
devices to extract the electrical parameters of material accurately and without direct
contact, makes them ideal for measuring glucose concentrations non-invasively.
Therefore, since the past decade, a lot of research work has focused on designing
microwave sensors that are capable of sensing the variation of glucose in blood.
Although some of the proposed devices have shown good sensitivity, however none of
them is accurate enough to replace the currently used glucometers.
This thesis addresses the design of non-invasive glucose sensors by relying on
microwave based components. Hence, the design of various types of radio frequency
(RF) circuits is presented to tackle this challenge. The behavior of the proposed RF
circuits as glucose sensing systems is tested using simulation in addition to in-vitro, ex-
vivo and in-vivo studies. A good correlation between the scattering parameters of
proposed sensors and the variations in glucose levels is attained. Several regression
models are also developed and applied on the collected data, where a selection of the
optimal model with the least prediction error is identified. Examined results using the
Clarke error grid demonstrate that 100% of the predicted glucose levels lie within the
clinically acceptable regions for the various proposed sensors.
vi
CONTENTS
ACKNOWLEDGEMENTS……………………………………………… v
ABSTRACT…………………………………………………………………… vi
LIST OF ILLUSTRATIONS……………………………………………. xii
LIST OF TABLES…………………………………………………………... xviii
Chapter
1. INTRODUCTION………………………………………………………. 1
1.1. Background …………………………………………………….…… 1

1.2. Project Motivation ………………………………………………….. 3
1.3. Thesis Structure ………………………………………………....….. 3
2. LITERATURE REVIEW……………………………………………... 5
2.1. Introduction …………………………………………………………. 5

2.2. Three Dimensional Measurements …………………………………. 5
2.2.1. Waveguides ……………………………………………… 5
2.2.2. Antennas …………………………………………………. 7
2.2.3. Resonators ……………………………………………….. 8
2.3. Planar Microwave Circuits …………………………………………. 9

2.3.1. Antennas ……….………………………………………… 9
2.3.2. Resonators ………………….……………………………. 11
2.3.3. Filters ………………….………………………………… 14
vii
3. THE CORRELATION BETWEEN MICROWAVE
AND BIOLOGICAL TISSUE …………………………………… 15
3.1. Introduction …………………………………………………….…… 15

3.2. Human Safety from Exposure to Electromagnetic Waves …………. 15
3.3. Dielectric Properties of Tissues……………………………………... 17
3.3.1. Polarization and Relaxation ……………………………... 17

3.3.2. Dispersion ……………………………………………….. 18
3.3.3. Dielectric Constants of Biological Tissues ……………… 19
3.3.3.1. Experimental Extraction of Dielectric
Constants………………………………………. 19
3.3.3.2. Mathematical Extraction of Dielectric
Constants……………………………………………… 21
3.3.3.3. Thickness of Biological Layers ………………. 23
3.4. Discussion ………………………………………………………….. 24
4. HIGH LOSS MATERIAL …………………………………………… 25
4.1. Introduction …………………..…………………………….……….. 25
4.2. Complex Permittivity………………………………………………... 25
4.3. Extraction of the Real Permittivity ε′r ……………………………… 26
4.4. Extraction of the Imaginary Permittivity ε′′r ………………………. 26
4.5. Effective Dielectric Permittivity ……………………………………. 27
4.6. Discussion ………………………………………………………….. 28
5. MATERIAL CHARACTERIZATION ……………………….. 29
5.1. Introduction …………………………………………………….…… 29

5.2. Microstrip Lines …………………………………………………….. 29
5.3. Characterization Methods…………………………………………… 31
5.3.1. Non-Resonant Methods ………………………………… 31
viii
5.3.1.1. Reflection Method……………………………. 31
5.3.1.2. Transmission/Reflection Method……………… 31
5.3.2. Resonant Methods………………………………………... 32
5.3.2.1. Resonant Method ……………………………... 32
5.3.2.2. The Resonant Perturbation Method …………... 33
5.4. Planar Resonators ………………………………………………….. 33
5.4.1. Straight Ribbon Resonator ………………………………. 33

5.4.2. T- Resonator ……………………………………………... 34
5.4.3. Ring Resonator …………………………………………... 35
5.5. Discussion …………………………………………………………. 37
6. DESIGN OF EM-BASED GLUCOSE SENSORS ………. 39
6.1. Introduction …………………………………………………….…… 39

6.2. Methodology ….…………………………………………………….. 39
6.3. Design Considerations ……….……………………………………... 41
6.4. Proposed Sensor #1………………………………………………….. 42
6.5. Proposed Sensor #2………………………………………………….. 46
6.5.1. Design Structure …………………………………………. 46

6.5.1.1. Top Layer……………………………………… 46
6.5.1.2. Bottom Layer………………………………….. 48
6.5.2. Design Features ………………………………………….. 50
6.5.2.1. Electric Field Distribution ……………………. 50
6.5.2.2. Size Reduction ……………………………...… 51
6.5.3. Fabrication and Measurements ………………………….. 51
6.5.4. Alternative Design ………………………………………. 53
6.6. Proposed Sensor #3 ………………………………………………... 56
6.6.1. Design Structure …………………………………………. 58

6.6.2. Sensitivity ………………………………………………... 63
6.6.3. Performance ……………………………………………… 64
ix
6.6.4. Fabrication and Measurements …………………………... 65
6.6.5. Tuning and Reconfiguration ……………………………... 67
6.6.5.1. Reconfigurable Microwave Circuits ………….. 68
6.6.5.2. Proposed Tunable Structure ………………….. 68
6.7. Discussion …………………………………………………………. 72
7. SIMULATION AND MEASUREMENT OF THE

PROPOSED EM-BASED GLUCOSE SENSORS……..…. 73
7.1. Introduction …………………………………………………….…… 73

7.2. Simulation ….……………………………………………………….. 73
7.3. Experimental Setup ……….………………………………………... 75
7.3.1. Serum Measurements ……………………………………... 78

7.3.2. Animal Tissues and Serum Measurements ……………….. 82
7.3.3. Clinical Measurements ……………………………………. 84
7.3.3.1. Study Design …………………………………… 84
7.3.3.2. Study Subject …………………………………... 84
7.3.3.3. Procedure ………………………………………. 84
7.3.3.4. Results ………………………………………….. 85
7.3.4. Measurements Accuracy ………………………………….. 88
7.3.4.1. Glucometer ……………………………………... 89
7.3.4.2. S-parameters …………………………………… 89
7.3.4.2.1. Vector Network Analyzer …………… 89
7.3.4.2.2. Calibration …………………………… 90
7.4. Discussion ……………..………………………………………….... 90
8. DEVELOPMENT OF A REGRESSION
MODEL…………………………………………………………………….. 91
8.1. Introduction.……………………………………………….………… 91
8.2. Regression Analysis ………………………………………………… 91
8.2.1. Preprocessing…………………………………………….. 93
8.2.2. Modeling Techniques…………………………………….. 94
x
8.2.3. Model and Features Selection ..………………………….. 95
8.2.4. K-fold Cross-validation ………………………………….. 96
8.3. Results ………………..……………………………………………... 97
8.3.1. Serum Measurements……………………………………. 97

8.3.2. Clinical Measurements.………………………………….. 99
8.4. Discussion …………………………………………………………. 105
9. CONCLUSION AND FUTURE WORK ………………… 106
REFERENCES …………………………………………………………. 107
xi
ILLUSTRATIONS
Figure Page
1. MMW measurement setup [10]………………………………………… 6
2. Proposed waveguides and measurement setup [12]…………………….. 6
3. Transmission variations with time [13]………………………………… 7
4. Tx and Rx antennas placed around pig’s ear [14]………………………. 8
5. Double ring resonators system [16]…………………………………….. 9
6. a) Antenna wrapped around subject's hand, and b) reference and

estimated glucose variations with time [20]……………………………. 10
7. Spiral resonator sensor [22]…………………………………………….. 11
8. Relation between permittivity and glucose concentrations in function of

frequency [23]…………………………………………………………... 12
9. Sensor structure embedded in the measurement setup [24]……………. 12
10. S21 magnitude and phase responses for different glucose

concentrations [24]……………………………………………………… 13
11. Prototype of the designed filter loaded by a human thumb [25]………... 14
12. S11 and S21 parameters variations with time [25]……………………... 14
13. Dispersion of biological tissues [28]……………………………………. 19
xii
14. Permittivity of muscle, fat and brain of rats in function of frequency
[29]……………………………………………………………………… 20
15. Measurement setup [33]………………………………………………… 22
16. Dielectric properties of blood for different concentrations of glucose

[33]……………………………………………………………………… 23
17. The geometry of microstrip transmission line………………………….. 30
18. Transmission line setup [39]……………………………………………. 31
19. Transmission/Reflection setup [39]…………………………………….. 32
20. Typical one-dimensional microstrip resonators: a) half-wave length line

resonator, b) hairpin resonator, and c) to f) other open loop resonators.
[40]……………………………………………………………………… 34
21. T-resonator [39]………………………………………………………… 35
22. Ring resonator [41]……………………………………………………... 36
23. Ring resonator measurement setup [39]………………………………... 37
24. Human model in ANSYS [44]………………………………………….. 40
25. Equivalent circuit model of SRRs [47]…………………………………. 43
26. Proposed split ring resonator design with enhanced coupling on

periphery. Image obtained from HFSS simulations…………………….. 43
27. Electric field distribution on the top layer of the proposed SRR-based
filter……………………………………………………………………... 44
28. Fabricated prototype of the SRR-based filter…………………………... 45
xiii
29. Simulated and measured response of the SRR-based filter…………….. 45
30. Proposed log-periodic broad-band reject filter…………………………. 47
31. Equivalent model of a defected ground structure………………………. 48
32. Electric field distribution of the traditional and modified OLR………... 50
33. Four turns uniform meander line……………………………………….. 51
34. Top and bottom layers of a fabricated prototype of the proposed log-
periodic BRF built on a 1.27 mm-thick Rogers 3006 substrate………… 52
35. Simulated and measured response of the proposed log-periodic BRF

built on a 1.27 mm-thick Rogers 3006 substrate……………………….. 52
36. Response of the log-periodic BRF using three flexible substrates……... 54
37. Top and bottom layers of a fabricated prototype of the proposed log-
periodic BRF built on a 0.25 mm-thick Rogers 3003 substrate………… 55
38. Simulated and measured response of the proposed log-periodic BRF

built on a 0.25 mm-thick Rogers 3003substrate………………………... 55
39. Top and bottom layers of the proposed dual-band reject filter………... 57
40. Response of the dual-band reject filter…………………………………. 57
41. Bottom layer of the proposed octa-band reject filter…………………… 59
42. Proposed sensing structure superposed with the topology of the lower
human arm……………………………………………………………… 59
43. Response of the proposed octa-band filter using a regular 50-ohms

transmission line………………………………………………………... 60
xiv
44. Tapered feeding line topology………………………………………….. 61
45. Proposed resonator-based feeding line topology……………………….. 62
46. Electric field distribution of the dual-band at f= 2.12 GHz and f= 2.34
GHz and octa-band filters at f=1.8 GHz and 2.2 GHz………………….. 63
47. Top and bottom layers of a fabricated prototype of the proposed octa-
band filter built on a 1.27 mm-thick Rogers 3006 substrate……………. 65
48. Top and bottom layers of a fabricated prototype of the proposed octa-
band filter built on a 0.25 mm-thick Rogers 3003 substrate……………. 66
49. Simulated and measured response of the proposed octa-band reject

filter built on a 1.27 mm-thick Rogers 3006 substrate………………….. 66
50. Simulated and measured response of the proposed octa-band reject

filter built on a 0.25 mm-thick Rogers 3003 substrate………………….. 67
51. Spice Model of the SMV 1705-079LF [56]…………………………….. 69
52. Topology of the proposed tunable filter………………………………… 70
53. Equivalent circuit of the proposed tunable filter………………………... 70
54. Fabricated Prototype of the proposed tunable filter…………………….. 71
55. Simulated and measured response of the proposed reconfigurable octa-

band reject filter built on a 1.27 mm-thick Rogers 3006 substrate. The
transmission coefficient is shown for a) V=0 Volts, b) V=5 Volts and c)
several voltage values…………………………………………………... 71
56. Multi-layered human tissues model…………………………………….. 74
57. S21 response of the proposed sensors for the single-layered (left) and
multi-layered models (right), a. narrowband sensor, b. broad-band
sensor, c. octa-band sensor……………………………………………… 76
xv
58. Magnitude and phase variation of the proposed sensors when loaded by
the multi-layered model, a) the narrowband sensor, b) the broad-band
sensor, and c) the octa-band sensor……………………………………... 77
59. Experimental setup for serum measurements…………………………... 79
60. Response of the proposed sensors for different glucose concentrations,

a) S21 magnitude of the broad-band sensor, b) S21 phase of the broad-
band sensor, c) S21 magnitude of the octa-band sensor, d) S21 phase of
the octa-band sensor…………………………………………………….. 80
61. Correlation between the response of the broad-band sensor and the
glucose concentrations at different frequencies………………………… 81
62. Correlation between the response of the octa-band sensor and the
glucose concentrations at different frequencies………………………… 82
63. Magnitude and phase variations of a) broad-band, and b) octa-band

sensors…………………………………………………………………... 83
64. Clinical trials setup……………………………………………………... 85
65. Correlation between the response of the broad-band sensor and the
blood glucose concentrations…………………………………………… 86
66. Correlation between the response of the octa-band sensor and the blood
glucose concentrations………………………………………………….. 87
67. Correlation between the response of a. broad-band sensor, and b. octa-

band sensor between two different OGTTs for the same subject and for
the same frequency a) f=1.4 GHz and b) f=1.33 GHz………………….. 88
68. Estimated versus reference serum glucose levels using LW-PLS, GP

and LASSO, a. rigid broad-band, b. flexible broad-band, c. rigid octa-
band, and d. flexible octa-band…………………………………………. 98
69. BGL profiles of all the patients collected using the broad-band sensor
and estimated using GP. Each plot includes the invasively measured
BGL (solid lines), the estimated BGL using Gaussian Process (dashed
lines) and the prediction using the different experiments (dots)………... 101
70. BGL profiles of all the patients collected using the octa-band sensor
and estimated using GP. Each plot includes the invasively measured
BGL (solid lines), the estimated BGL using Gaussian Process (dashed
lines) and the prediction using the different experiments (dots)………... 102
xvi
71. Clarke error grid for the data collected using a) broad-band sensor, and
b) octa-band sensor……………………………………………………... 104
xvii
TABLES
Table Page
1. Standards values of specific absorption rate [26]………………………. 16
2. Permittivity of human biological tissues at specific frequencies [30]….. 20
3. Conductivity of human biological tissues at specific frequencies [30]… 21
4. Thickness of some biological tissues [34]……………………………… 24
5. FDA's accepatable accuracy for glucose measurements devices [45]….. 42
6. Dimensions of the proposed BPF………………………………………. 44
7. Dimensions of the designed tapered line……………………………….. 47
8. Dimensions of the proposed log-periodic BRF designed on a 1.27 mm-

thick Rogers 3006 substrate…………………………………………….. 49
9. Characteristics of several flexible substrates…………………………… 53
10. Dimensions of the proposed log-periodic BRF designed on a 0.25 mm-

thick Rogers 3003 substrate…………………………………………….. 56
11. Dimensions of the dual-band reject filter……………………………….. 58
12. Dimensions of the proposed tapered feeding line topology…………….. 61
13. Dimensions of the proposed resonator-based feeding line topology…… 62
xviii
14. Maximum attained electric field intensity for the octa-band filters at
different frequencies……………………………………………………. 64
15. Performance of different multi-band filters found in the literature…….. 65
16. Thickness of the layers used in the model……………………………… 78
17. Accuracy of the ACCU-CHEK Performa glucometer [57]…………….. 94
18. Mean percentage error for the glucose levels estimated using LW-PLS,
GP, and LASSO………………………………………………………… 103
xix
CHAPTER 1
INTRODUCTION
1.1. Background
During the past few decades, changes in lifestyle and nutrition made diabetes
one of the most prominent diseases among chronic conditions. By 2014, the number of
people diagnosed with diabetes reached 422 million worldwide. This number is
expected to increase dramatically in the upcoming years due to the exponential growth
of this disease [1]. In fact, diabetes is considered one of the most common metabolic
diseases. There are two main types of diabetes: type 1 and type 2, with type 1 being the
most serious condition. For patients with type 1 diabetes, the auto-immune system
destroys the cells that produce insulin. Consequently, the production of insulin
decreases or stops completely and hence the blood glucose level (BGL) increases. Only
10% of diabetics are diagnosed with type 1. Around 90% are diagnosed with type 2
diabetes, and they suffer from a slow production rate of insulin in the body, which in
turn decreases the effects of insulin resulting in chronic hyperglycemia [2].
Although the causes of these two categories are different, however their
consequences are dangerous and much alike. In fact, advanced stages of both types of
diabetes are associated with several complications such as kidney failure, blindness and
limb amputation. In addition, diabetes increases the risk of stroke and coronary heart
disease. In 2015, around 1.6 million deaths were caused directly by this condition. An
additional 3 million deaths were caused by the complications of the disease [1].
1
For treatment, accurate determination of BGL is a necessity. Currently, BGL is
measured using a glucometer. Although this device is highly accurate, however, it is
invasive. In fact, the patient has to prick a finger each time a glucose measurement is
required. This process is painful for diabetics especially with the need to measure
glucose levels several times during the day. In the long term, this method increases the
risk of infection and may damage the underlying tissue. Additionally, this measurement
technique does not provide continuous monitoring, which can result in missing serious
hypo/hyper-glycemic incidents that can happen between two measurements [3].
In the past decade, several new blood glucose measurement techniques have
been investigated, which fall within one of the two categories: minimally invasive and
non-invasive. In [4] and [5], a review of the most popular non-invasive techniques is
presented. This review includes infrared spectroscopy, excreted physiological fluid
analysis, electrodes, microcalorimetry and optical sensors. Additionally, minimally
invasive methods for BGL monitoring such as sonophoresis and iontophoresis aim to
extract the level of glucose from the skin [3]. All the proposed sensors have failed to
replace the current measurement method as they lacked accuracy. More recently,
alternative solutions are developed using electromagnetic measurement techniques, as a
mean to provide continuous and non-invasive monitoring of BGLs. These methods rely
on coupling EM waves on precise spots of the human body. The reflected waves are
collected back and monitored, as they include information concerning the electrical
properties of the underlying biological layers. Any variations in the collected waves
2
directly relates to parameters of blood. This information can be used to provide non-
invasive and continuous monitoring of BGLs [6].
The ability of EM-based devices to extract parameters accurately, without
perturbing the MUT, makes them ideal for measurements that require no direct contact
with the material. The robustness and accuracy of these devices have encouraged
researchers to investigate the possibility of designing microwave sensors for biomedical
applications. Consequently, multiple wearable devices are developed that are capable of
monitoring cardiac and respiratory activities [7], sensing bodies’ abnormalities and
disease prediction [8][9].
1.2. Project Motivation
In this research, we propose the design of three microwave sensors to detect
variations in BG. These sensors are designed and tested in several in vivo and in vitro
studies. A comparative study is also established to associate the performance of the
different devices. A regression model is finally built in the aim of predicting glucose
levels.
1.3. Thesis Structure
This thesis is divided into eight chapters. Chapter one includes an introduction.
Chapter two provides a literature review that presents the most recent research work on
non-invasive glucose monitoring. Chapter three discusses the electrical properties of
different human tissues and the corresponding safe absorption rates of electromagnetic
3
(EM) energy. Chapter four presents a theoretical background on the electrical properties
of lossy material. Chapter five discusses material characterization techniques that are
developed in literature. Chapter six presents the proposed sensors. Chapter seven
discusses all the conducted in vivo and in vitro studies. Chapter eight presents the
established regression model. Chapter nine concludes this work and presents future
research steps.
4
CHAPTER 2
LITERATURE REVIEW
2.1. Introduction
Throughout the last decade, research effort has focused on proposing microwave
devices that can be employed for the monitoring of blood glucose concentrations in a
continuous and non-invasive manner. These approaches include EM based waveguides,
antennas and resonators operating at different frequencies. In the following, a literature
review is provided. Topics associated to finding a relation between BGL and the
electrical parameters, in addition to investigating the sensitivity of several microwave
devices in sensing the BGL are covered in this section.
2.2. Three Dimensional Measurements
2.2.1. Waveguides
Using the setup in Fig. 1, a correlation between BGL and the electrical
properties of tissues is presented in [10] and [11]. This relation is generated based on
the fact that the relative permittivity of blood is directly related to the concentration of
glucose. In these experiments, electromagnetic waves operating at 30 GHz are incident
on body tissue. By monitoring the reflected waves from the tissue, it is shown that the
dielectric constant of blood is altered for various BGLs. The variation in the dielectric
constant are noted as a shift in frequency operation of the reflection coefficient.
5
Figure 1: MMW measurement setup [10].
Two rectangular waveguides operating between 50 – 75 GHz are designed and
tested in [12] using water-salt-glucose solutions as shown in Fig. 2. Glucose
concentrations as small as 0.025 (wt%) are detected with an accuracy of 0.22 dB per
wt%. The sensitivity of the system varies for different frequencies, and the best results
are reached at two different bands between 59-64 GHz and between 69-73 GHz.
Figure 2: Proposed waveguides and measurement setup [12].
Real-time direct correlative measurements of the blood glucose concentration is
performed using commercial blood test strips and millimeter-wave absorption for
several injections of glucose, insulin, and saline in a live anesthetized animal [13]. The
results of this work (Fig. 3) show an increase in the wave transmission after glucose
injection, and a decrease in transmission after insulin injection. The relatively slow time
6
for the observed changes (10-15 minutes for glucose and 20-40 minutes for insulin)
correlates well with the expected speed of glucose and insulin metabolism. The lack of
significant changes in MMW transmission upon injection of saline indicates the
selectivity of mm-wave absorption to the tissue concentration of glucose.
Figure 3: Transmission variations with time [13].
2.2.2. Antennas
A device is developed in which two antennas operating within the V-band are
placed around a pig’s ears as shown in Fig. 4 [14]. Transmission and reflection
parameters are measured using a Vector Network Analyzer (VNA). The frequency
range of interest is between 58.5 GHz and 61.5 GHz. It is reported that the best results
are obtained when the antennas are closest to the tissues and are placed in an area rich
with veins. Furthermore, results demonstrate that an increase in glucose concentration
produces an increase in transmission at specific frequencies.
Following the integration of temperature and motion sensors, the patch is tested
using a water-glucose phantoms. The proposed system is able to detect changes in
7
glucose as low as 24 mg/dl. Furthermore, the system is tested on 10 healthy male
subjects in [15]. These experiments verified its ability to successfully detect glucose
spikes.
Figure 4: Tx and Rx antennas placed around pig’s ear [14].
2.2.3. Resonators
The use of split ring resonators for glucose monitoring is examined in [16]. The
system consists of two rings operating at 1.4 GHz as shown in Fig. 5. One ring is used
for sensing and is placed at close proximity to the skin. The second one is placed far
away and is considered as a reference to regulate the temperature effect. Both in-vitro
and in-vivo measurements are carried out on one healthy patient and the results reached
are comparable to those attained by the commercial glucometer. Furthermore, the
proposed sensor is verified to be extremely selective, as it offers high sensitivity to
glucose variations, limited sensitivity to other sugars and no sensitivity to vitamins [17].
The device is then clinically tested on 24 volunteers. Promising results are reached as
210 of 214 data points lay in the clinically acceptable regions [18].
8
Figure 5: Double ring resonators system [16].
2.3. Planar Microwave Circuits
Planar circuits are also discussed in the literature. These structures offer a
considerable reduction in size compared to the previously discussed three dimensional
systems.
2.3.1. Antennas
One suggested system for planar circuits consists of implementing microstrip
antennas as in [19], [20]. Several types of planar antennas including spirals, serpentines
and dipoles are designed and tested in order to assess their ability to sense variations in
glucose levels. In one embodiment, the reflection coefficient of the proposed antennas is
monitored by placing the device near several glucose/water solutions of different
concentrations. This experiment proved the ability of the antennas to sense variations in
glucose concentration interpreted as shifts in the resonance frequency within the
reflection coefficient. It is also reported that the dipole outclassed the remaining types
of antennas. In another embodiment, the suggested device is mounted on the hand of a

9
diabetic patient and the volunteer’s glucose levels are monitored as shown in Fig. 6. A
good correlation is noted between the resonance frequency of the antenna and the
reference BGLs measured with a traditional glucometer. A sensitivity of 1 MHz per
14.62mg/dl is also reported. Another observation made relates to the high dependency
of the antenna’s response on the biological variables of each subject such as
metabolism, skin color, BMI and other elements.
Figure 6: a) Antenna wrapped around subject's hand, and b) reference and estimated
glucose variations with time [20].
The design of two ultra-wideband slot antennas of bandwidth equal to 12.5 GHz
is also considered in [21]. The proposed sensing structures are envisioned to be placed
around human’s earlobe to continuously monitor BGLs. Testing the suggested antennas
on glucose-water phantoms showed a linear correlation between the antenna parameters
and various glucose levels. Note that the range considered for glucose concentrations is
quite high (0-4000 mg/dl).
10
2.3.2. Resonators
A spiral sensor is presented in [22] to determine the BGL non-invasively (Fig.
7). In this research study, the relative permittivity of blood is investigated at different
frequencies between 0 and 2 GHz. One volunteer performed a soda test by fasting for
over eight hours before placing the device on his wrist. Then, the subject consumed a
sugared soft drink. Results show that variations in glucose levels caused shifts in the
resonance frequency of the sensor (Fig. 8). The fact that the shifts in frequency are not
linear over the whole range implies that some frequencies may be more sensitive to the
changes in BGL.
Another research is performed on the spiral sensor in [23]. In this experiment,
the concentration of sugar in water is varied by adding sugar into the solution. Results
of this work prove that the permittivity of the water decreases while adding sugar as
displayed in Fig. 8.
Figure 7: Spiral resonator sensor [22].
11
Figure 8: Relation between permittivity and glucose concentrations in function of
frequency [23].
A half-wavelength resonator operating at 2 GHz is designed in [24] to sense the
changes of glucose levels in water as shown in Fig. 9. By varying the concentrations of
glucose from 0 to 300 mg/dl, the magnitude and phase of the reflection and transmission
coefficients are recorded.
Figure 9: Sensor structure embedded in the measurement setup [24].
12
Fig. 10 presents the magnitude and phase of the S21 parameter for the various
concentrations of glucose. It is clear that the proposed structure is capable of detecting
changes in the concentrations of glucose with an accuracy in the range of 50 mg/dl.
Figure 10: S21 magnitude and phase responses for different glucose concentrations
[24].
13
2.3.3. Filters
A planar BPF is also proposed and is illustrated in Fig. 11 [25]. Parameters of
this structure such as resonant frequency and insertion loss are sensitive to the dielectric
constant of the superstrate. When contacted by a thumb, results show that the response
of the sensor changes.
Figure 11: Prototype of the designed filter loaded by a human thumb [25].
An experiment is conducted on the suggested sensor in which a volunteer
consumes sugar water with a high concentrations of glucose. Measurements show that
the permittivity of blood decreased continuously, which causes an increase in the
resonant frequency of the BPF (Fig. 12). After 600 seconds, a stable state is detected. It
is important to indicate that a linear operation is reported between 1.5 and 2 GHz.
Figure 12: S11 and S21 parameters variations with time [25].
14
CHAPTER 3
THE CORRELATION BETWEEN MICROWAVE AND

BIOLOGICAL TISSUES
3.1 Introduction
This chapter includes two main sections that illustrate the relation between
electromagnetic waves and human science. The first section addresses the safe levels of
human exposure to EM energy. The second section discusses the electrical properties of
the biological tissues, and the response of these tissues when excited with EM energy.
3.2. Human Safety from Exposure to Electromagnetic Waves
From a physiological point of view, concerns about the exposure of the human
body to electromagnetic waves have risen recently due to the abundance of such waves
within the ambient environment. In fact, nowadays, multiple EM sources such as cell
phones, routers and satellites are widely used. As a result, the human body is exposed
abundantly and more frequently to such radiation. In fact, several standards and
protocols are developed, by many researchers and health experts to limit as well as
control the exposure of the human body to EM waves.
The safe exposure standards are set by the Institute of Electrical and Electronics
Engineers (IEEE). According to the standard ‘C95.1-2005’, the specific absorption rate
is defined as “the time derivative of the incremental energy (dW) absorbed by an
incremental mass (dm) contained in a volume element (dV) of given density (ρ)” as
demonstrated in Eq. 1 [26].

15
d dW
SAR = dt (ρdV) (W⁄Kg) (1)
Also, a relation between the absorption rate and the electric field at a specific
point is presented in Eq. 2, where σ and ρ are the conductivity (S/m) and mass density
(kg/m3 ) of the tissues and E represents the rms value of the electric field strength in the
tissues (V/m).
σ|E|2
SAR = (W⁄Kg) (2)
ρ
Accordingly, the safe levels of human exposure to RF fields in the spectrum
ranging between 3 kHz and 300 GHz are measured and tabulated in the standard [26].
Table 1 summarizes the SAR limits for several body parts, in both public and controlled
environments. Note that the SAR values in Table 1 are restricted for frequencies below
3 GHz. For higher frequencies, these values may differ and are beyond the scope of this
thesis work.
Table 1: Standards values of specific absorption rate [26].
Exposure Frequency Whole- Partial- Hands, Wrists,

Range Body Body Ankles and Feet
(W/Kg) (W/Kg) (W/Kg)
Controlled 100 KHz - 0.4 10 20
Environment 3 GHz
General Public 100 KHz - 0.08 2 4
3 GHz
16
3.3. Dielectric Properties of Tissues
Understanding the behavior of human tissue when exposed to EM energy is of
great importance for several areas of investigation such as electrical impedance
imaging, microwave hyperthermia and radiofrequency to name a few. Accordingly,
many researchers have focused on deriving a relation between the electrical properties
of the human tissues over a wide range of frequencies. In this process, it has been noted
that water is the major component of biological materials. Consequently, when
considering any part of the human body, water represents the main contributor to
permittivity. However, since biological materials are complex mixtures, therefore their
electrical response is not limited to only one component. In fact, each tissue has its own
contribution depending on the electrical properties and thickness of the layer.
In this section, the concepts of electrical polarization and dispersion for
biological tissues are presented, in addition to tables that summarize the dielectric
constants and thickness of several biological tissues.
3.3.1. Polarization and Relaxation
An induced electric field applied on biological tissues disturb the distribution
charges. This effect is known as electric polarization. The relation for the polarization
density in terms of the electric field and the dielectric constants of the material is given
by Eq. 3, where E is the induced electric field, P the dielectric polarization density, ε0
the permittivity of free space and χe the susceptibility of the tissues [27].
P = ε0 χe E (3)
17
When an electric field is applied on a structure, several non-idealities cause a
deferral between the polarization and the variations in the electric field. This delay is
recognized as a dielectric relaxation. To measure the relaxation time of a system, an
excitation should be applied on the structure. Then the relaxation time towards reaching
a new equilibrium is recorded [27].
3.3.2. Dispersion
Dielectric dispersion is the dependence of the permittivity of a structure on the
frequency of an applied electric field. Consequently, the dielectric constant value of a
given material is not fixed, but rather frequency dependent. For biological tissues,
dispersions are apparent. In [28] the electrical properties of human tissues are
categorized by three main dispersions: (1) the low frequency α-dispersion that is linked
with ionic diffusion processes, (2) the β-dispersion for radio frequencies, associated
with the polarization of cellular membranes, proteins and additional organic
macromolecules, and (3) the ϒ-dispersion for microwave frequencies, produced by the
polarization of water molecules. Additional dispersions may exist such as the δ-
dispersion which is a subset of β-dispersion (Fig. 13). Another observation from Fig. 13
concludes that biological tissues exhibit a relatively high dielectric constant at low
frequencies, and these values decrease at higher frequencies.
18
Figure 13: Dispersion of biological tissues [28].
3.3.3. Dielectric Constants of Biological Tissues
For glucose sensing, the objective is to detect the variations in blood glucose
while suppressing the effects of minerals in the blood as well as minimizing the
influence of the tissues surrounding the arteries and veins. Therefore, understanding the
characteristics and behavior of tissues such as skin, fat, blood, bones and muscle is of
great importance.
3.3.3.1. Experimental extraction of dielectric constants
Several researchers have investigated the extraction of the dielectric constants of
biological tissues in function of frequency. At first, experiments are conducted on
animals. In Fig. 14 for instance, the relative dielectric constants for rat muscle, rat brain
and canine fat with respect to frequency are shown [29]. For these tissues, the
19
permittivity of fat is the most consistent along the spectrum. However for the brain and
muscles, the values of permittivity dropped noticeably.
Figure 14: Permittivity of muscle, fat and brain of rats in function of frequency [29].
Later on, using advanced imaging techniques, researchers have been able to
characterize the dielectric constants of humans. Tables 2 and 3 summarize the values of
permittivity and conductivity for several biological tissues of the human body [30].
These values are collected from the most advanced researches in the field. From table 2,
it is again clear that permittivity drops at high frequencies.
Table 2: Permittivity of human biological tissues at specific frequencies [30].
Tissue f = 433 MHz f = 915 MHz f = 2.45 GHz
Skin 47 45 44
Fat 15 15 12
Blood 66 62 60
Muscle 57 55.4 49.6
Artery - - 43
20
Table 3: Conductivity (S/m) of human biological tissues at specific frequencies [30].
Tissue f = 433 MHz f = 915 MHz f = 2.45 GHz
Skin 0.84 0.97 -

Fat 0.26 0.35 0.82
Blood 1.27 1.41 2.04
Muscle 1.12 1.45 2.56
Artery - - 1.85
Several other researchers have worked on modeling the human tissues at pre-
defined locations in the body such as the neck, the ear, the leg and the arm. This
methodology is employed in order to improve the accuracy of the results [31], [32].
3.3.3.2. Mathematical extraction of dielectric constants
In addition to experiments, mathematical models are also developed to predict
the variation of the tissues’ permittivity as a function of frequency. The Cole-Cole
model is perhaps the most known model for dielectric relaxation. Using the Cole-Cole
as shown in Eq. 4, and with the appropriate choice of parameters for the tissue, the
dielectric behavior can be predicted over the desired frequency range.
Δεn σ
ε = ε'c (w)-jε''c (w) = ε∞ + ∑ 1+(jwτ (1-αn ) + jwεi (4)
n) 0
ω: Angular frequency.
ε′c (w): Frequency-dependent dielectric constant.
ε′′c (w): Frequency-dependent dielectric loss.
n: Order of model.
21
ε∞ : High frequency permittivity.
Δεn :Dispersion magnitude
τn : Relaxation time constant,
αn : Dispersion broadening parameter.
σi : Static ionic conductivity.
The values of ε∞ , Δεn , τn , αn , σi for the first four orders of the model are
summarized in [33]. In this reference, a study is conducted in order to build a relation
between the electrical properties and the BGLs. Measurements on blood samples are
performed using a dielectric probe kit and a vector network analyzer (Fig. 15). The
measurements are implemented at frequencies between 500 MHz and 20 GHz.
Figure 15: Measurement setup [33].
The glucose levels of the samples are varied between eight different glucose
concentrations, ranging from 0 mg/dl to 16,000 mg/dl. The resultant dielectric constant
and conductivity in function of frequency are presented in Fig. 16. Results show that the
real part of the permittivity in blood decreases for high concentrations of glucose. On
the other hand, the conductivity does not vary much between 0.5 GHz and 9 GHz.
However, for higher frequencies, small deviations are noticed.
22
The collected data is then fitted into the Cole–Cole model in order to build a
relation between the electrical properties of the blood and the glucose concentration.
The formulated model is able to successfully predict the dielectric properties in function
of frequency. Consequently, this work prove that the Cole-Cole model is a powerful
mathematical tool that can be used to extract the dielectric constants of the blood.
Figure 16: Dielectric properties of blood for different concentrations of glucose [33].
3.3.3.3. Thickness of biological layers
Concerning the thickness of the layers, it is hard to create a general relation that
includes all the population. In fact tissues such as fat and muscles are highly dependent
on age, gender and lifestyle. However, some approximations may be found in [34].
Also, a summary of the work developed in this process can be found in [35]. Table 4
presents the tissue layers’ thicknesses for some body parts. All the values are in
millimeters.
23
Table 4: Thickness of some biological tissues [34].
Tissue Limb Thorax Forehead Abdomen

Skin 2.5 2 2 2.5
Fat 8 - 2 10
Muscle 25 3 - 20
Bone - 10 7 -
3.4. Discussion
This chapter addressed the electrical properties of the biological tissues, and
the response of these tissues when excited with EM energy, evaluated using both
experimentations and mathematical equations such as the Cole-Cole model. These
procedures shows that the electrical properties of the biological tissues highly depend
on the frequency in addition to the physical and physiological conditions of the subject.
Accordingly, the design of any RF glucose sensor must account for these variations,
while respecting the standard safe absorption rates set by IEEE.
24
CHAPTER 4
HIGH LOSS MATERIAL
4.1. Introduction
Chapter 3 demonstrates that biological tissues have the characteristics of lossy
materials due to their relatively high permittivity. This chapter provides an overview on
high loss materials, which are defined by their complex permittivity. Furthermore, the
effect of having multi-layered high loss materials is considered.
4.2. Complex Permittivity
In a lossy media, the effective complex permittivity is given in Eq. 5 where ε'r is
the real permittivity that signifies the stored electric field energy, ε′′ r is the imaginary
permittivity that accounts for the losses in the medium and tanδ is the loss tangent of
the medium expressed in Eq. 6.
εeff = ε′r − jε′′ r (1 − jtanδ) (5)
ε′′ r
tanδ = (6)
ε′r
At resonance, the electric and magnetic field energy stored in any resonant
structure must be equal. When a material perturbs the stored energy, the field
distribution is perturbed and hence the resonance frequency shifts. This shift in the
resonance frequency is related to the properties of the sample based on Eq. 7.
Δfr ∫(ΔεE1 .E0 + ΔµH1.H0)dv

= (7)
fr ∫(ε0.|E0 |2 + µ0 .|H0 |2 )dv
25
fr and Δfr are the resonance frequency and the shift in the resonance frequency
respectively, Δε and Δµ represent the change in the permittivity and permeability.
E0 and H0 represent the field distributions in free space and E1 and H1 are the field
distributions with the perturbation of the MUT [36].
4.3. Extraction of the Real Permittivity 𝛆′𝐫
The resonant frequency is related to the structure’s properties using Eq. 8.

1
fr = (8)
2π√L(CSubtrate +CSUT )
Where L is the total inductance, CSubtrate is the substrate capacitance, and CSUT
is the capacitance of the sample under test. The only unknown in this equation is the
capacitance of the sample under test which is directly proportional to its real
permittivity ε′r according to Eq. 9. Consequently a relation between ε′r and fr can be
generated. The generated relation allows the characterization of the permittivity based
on the shifts in the resonant frequency [37].
ε′r d
C= (9)
A
4.4. Extraction of the Imaginary Permittivity 𝛆′′ 𝐫
The quality factor Q is related to the loss resistance using Eq. 10.
R
Q = 2πf (10)
rL
26
The variable in this case is the loss resistance of the sample under test R which
is directly proportional to its loss tangent tanδ according to Eq. 11. Consequently a
relation between ε′′r , fr and the S-parameters can be generated [37].
2πfr L
tanδ = (11)
R
4.5. Effective Dielectric Permittivity
For a stacked multi-layered structure (Fig. 16), the relative permittivity ‘εr ’ of
each layer is unique (εr = εMaterial ). However, the effective permittivity ‘εeff ’ of the
whole structure depends on the permittivity and thickness of each layer. Eq. 12 was
developed to compute the effective dielectric constant of multilayered structures with
thickness h ≅ λ⁄10 [38].
|d1 |+|d2 |+⋯+|dn |

εeff = d d d (12)
| 1|+| 2|+⋯+| n |
ε1 ε2 εn
With εn being the dielectric constant of the top layer. dn is calculated using Eq.
K(k )
13, k n is given in Eq. 14 and K′(kn ) in Eq. 15. hn refers to the thickness of the top layer
n
[38].
K(k ) K(k ) K(k )

dn = K′ (kn ) − K′ (kn−1 ) − K′ (k1 ) (13)
n n−1 1
1
kn = πw (14)
cosh( )
4(hn +hn−1+⋯+h1)
K(kn ) 1 1+√kn
= π ln (2 1−√kn) for 0.7 ≤ kn ≤ 1 (15)
K′ (kn )
27
4.6. Discussion
This chapter presented, from an EM perspective, the equations that define high
loss materials and their effect on any EM wave. Furthermore, it assessed the effect of
placing several high loss layers on waves’ propagation.
28
CHAPTER 5
MATERIAL CHARACTERIZATION
5.1. Introduction
Material characterization signifies measuring the structure and electrical
properties of a material under test (MUT). In Applied electromagnetics and RF systems,
two techniques are developed for material characterization purposes. The two
techniques are divided as: the resonant and non-resonant methods. These methods are
applied using waveguides or using planar circuit boards. However, since the objective is
to design a wearable device for BGL monitoring, the device must be as small as
possible. Consequently, it is essential to use printed circuit boards’ techniques in order
to decrease both the size and cost of the device [39].
The most known and used planar method relies on microstrip technology. This
chapter presents this approach and provides a review on the most common material
characterization methods using microstrips.
5.2. Microstrip Lines
Microstrip lines are planar transmission lines that are easy to manufacture and
integrate on low cost substrates. Fig. 17 shows the geometry of a microstrip line. The
conductor of width (W), is printed on the substrate of thickness (d) and dielectric
constant (ϵr) [36].
29
Figure 17: The geometry of microstrip transmission line.
The effective dielectric constant of a microstrip line depends on the conductor
width and the thickness of the substrate as in Eq. 16 [36].
εr+1 εr −1 1
εe = + × (16)
2 2 √1+12d⁄W
The characteristic impedance of a microstrip line depends on the substrate
thickness and dielectric constant and the conductor width. It is calculated using Eq. 17
[36].
60 8d W w
ln ( W + ) for d ≤ 1
√ εe 4d
Z0 = { 120π w (17)
W W for d ≥ 1
√εe [ d +1.393+0.667 ln( +1.444))]
d
The width to depth ratio is computed as shown in Eq. 18, and the parameters A
and B are provided in Eq. 19 and Eq. 20 respectively [36].
8eA w
for <2
W e2A −2 d
= {2 εr−1 0.61 w
(18)
d
[B − 1 − ln(2B − 1) + {ln(B − 1) + 0.39 − }] for >2
π 2εr εr d
Zo εr+1 ε −1 0.11
A = 60 √ + εr+1 (0.23 + ) (19)
2 r εr
377π
B = 2Zo (20)
√ε r
30
5.3. Characterization Methods
5.3.1. Non-Resonant Methods
Non-resonant characterization methods provide a general knowledge of EM
properties over a wide range of frequencies. For that, either the reflection or the
reflection/transmission parameters are used to extract either one or two EM properties
of the MUT [39].
5.3.1.1. Reflection method
This method uses a planar transmission line built on a substrate filled with the
MUT (Fig. 18). The EM properties of the sample are extracted from the S parameters.
The disadvantage of this method lies in the fact that it cannot be used for samples of
different thicknesses.
Figure 18: Transmission line setup [39].
5.3.1.2. Transmission/reflection method
This is another non resonant technique that is adequate for all the samples
regardless of the thickness. The determination of the complex permittivity and

31
permeability requires the measurement of the S11 and S12 parameters. During
measurement, the microstrip line is loaded by a film and its support (Fig. 19). The film
occupies a part of the cross section of the microstrip line. After determining the S11 and
S12 parameters, the transmission and reflection coefficients are calculated from Eq. 21
and Eq. 22. Consequently εr and µr can be extracted [39].
S11 = S22 = r1 exp(−2jγolo) (21), where r1 is the reflection coefficient.
S21 = S12 = t1 exp(−2jγolo) (22), where t1 is the transmission coefficient.
Figure 19: Transmission/Reflection setup [39].
5.3.2. Resonant Methods
Resonant methods provide accurate knowledge of the EM properties over one or
multiple discrete frequencies. Two techniques lie within the resonant methods, these
techniques are known as: the resonator and the resonant perturbation method.
5.3.2.1. Resonator method
The resonator method is used to extract the permittivity and the loss tangent
values based on the measurements of the resonant frequency and the quality factor. It
32
consists of placing the MUT between two conducting plates of a resonator. This
technique is highly accurate for only low loss dielectrics. For high dielectric constants’
materials, the resonant perturbation method is the most preferred.
5.3.2.2. The resonant perturbation method
This method consists of placing a dielectric MUT near a resonator, which causes
a shift in frequency and quality factor. Based on this shift the EM properties of the
material can be extracted. The sensitivity of this technique is highly dependent on the
type of resonator used. Several planar resonators are discussed in the following section
[39].
5.4. Planar Resonators
In microstrip technology, multiple resonators are used for material
characterization such as the straight ribbon, the T resonator and the ring resonator.
5.4.1. Straight Ribbon Resonator
The straight ribbon resonator (Fig. 20.a) is an open ended line of length l = n ×
λ
, n=1,2… The fields in a straight ribbon resonator extend beyond the ends of the line
2
causing radiation losses (fringing effect). This effect is modeled as a grounding
capacitance or consequently as a transmission line. As a result, the quality factor of such
33
type of resonators is relatively low. The ribbon resonator can be adjusted, as shown in
Fig. 20, to increase the quality factor [40].
Figure 20: Typical one-dimensional microstrip resonators: a) half-wave length line

resonator, b) hairpin resonator, and c) to f) other open loop resonators. [40].
5.4.2. T- Resonator
The T-resonator method is introduced as a technique that enhances the
transmission line characterization, reduces the radiation losses, and decreases the size of
the device. A T-resonator is a quarter wave long transmission line shown in Fig. 21.
The T pattern is an open-end transmission stub that resonates at odd integer multiples of
λ/4. By coupling the structure directly to the transmission line, and by having only one
open end, inaccuracies of the gaps are eliminated and the radiation and discontinuity
losses are reduced.
34
Figure 21: T-resonator [39].
Using this resonator, εreff is determined from fr using εreff =
nc
(4fr(l+lc)) with n = 1,3,5 … and hence the permittivity of the material is determined.
ε (εr −1)
The loss tangent is also determined from the quality factor using tanδ = Qdεreff(ε
r reff −1)
QQc
with Q d = Qc−Q, Q being the measured quality factor and Q c the calculated quality
factor.
5.4.3. Ring Resonator
The ring resonator, shown in Fig. 22, does not have open ends, which decreases
its radiation loss and enhances its quality factor even more. This fact made the ring
resonator one of the most accurate and sensitive planar resonators in material
characterization. The resonant condition for the first resonant mode is given by Eq. 23,
where r is the ring’s mean radius [39]. Therefore for a given radius of the ring, the
wavelength that produces the first resonant mode can be determined.
2πr = λg (23)
35
Figure 22: Ring resonator [41].
When loaded by a material with a specific dielectric constant, the resonant
frequency of the ring, where the peak occurs, decreases [41]. The larger the dielectric
values the higher is the shift. This behavior is due to the fact that the energy coupled
into the ring splits equally over the top and bottom sections of the ring. Therefore, a
standing wave will develop in a way such that when the ring is in resonance, the
maxima occur at the coupling gaps and the nulls are noticed at the top and bottom of the
ring [41].
To determine the properties of a sample, a single layer substrate can be used in
which the MUT is used as the substrate. Another possibility is multi-layer substrate in
which the ring acts as a measurement device and the MUT is placed as a cover on top of
the circuit. A PTFE block can be used to eliminate the air gap between the circuit and
the MUT. For the setup in Fig. 23, the permittivity ε1 and the thickness h of the
substrate are known. With no loading, ε2 = ε3 = 1. Therefore f0 can be measured and
εreff,0 is calculated from the substrate properties. With loading ε3 is known and ε2 must
be determined. To determine ε2 , εreff,1 is calculated using Eq. 24 and f1 is calculated
using Eq. 25 [39].
fo 2
εreff,1 = εreff,0 (f1 ) (24)
36
nc nc
πD = fo = f1 (25)
√εreff,0 √εreff,1
Figure 23: Ring resonator measurement setup [39].
Several resonator-based sensors are proposed in literature as highly sensitive
sensors for the aim of characterizing materials’ dielectric constant variations. The focus
has always been on designing narrow-band sensors with a high quality factor as in [42].
Such response increases the intensity of electric fields. The use of complementary split
ring resonators (CSRRs) with narrow responses has been examined in [42], and [43].
The proposed resonators are able to sense and predict the dielectric constant values of
low loss substrates with a percentage error that does not exceed 10% [42]. Also, by
increasing the number of resonators from two to three, the sensor is able to predict both
the dielectric constant and thickness of the substrates with lower error [43], compared to
[42].
5.5. Discussion
This chapter addressed several planar methods for materials characterization.
The resonant perturbation technique has the advantage of extracting the electrical
properties without perturbing the MUT. Its sensitivity however is highly dependent on
37
the type of resonator employed. Accordingly, several resonators are also presented in
this chapter, such as the open loop and ring resonators. Modified versions of these two
structures are considered in this thesis as discussed in the following chapter.
38
CHAPTER 6
DESIGN OF EM-BASED GLUCOSE SENSORS
6.1. Introduction
In this chapter, several two port RF filters that act as glucose sensors are
designed and tested. The performance of a narrow band SRR based band-pass filter is
initially evaluated. Also two novel broad-band and tunable octa-band reject filters are
proposed. The narrowband BPF operates at 2.4GHz. The broad-band reject filter covers
the whole frequency range between 1.25 GHz and 2.25 GHz. Finally, the tunable octa-
band reject filter covers multiple bands ranging between 1.5 GHz and 2.4 GHz.
The designs characteristics of these filters are discussed in this chapter. The
proposed structures are fabricated and tested using different substrates. A comparison
between the simulated and the measured results is also presented, and a good agreement
is noticed.
6.2. Methodology
The proposed sensors are initially designed, simulated and tested in free space
scenarios. To examine the performance of the filters, the scattering parameters, S11,
S21, and S22, are selected. S11 and S22 represent the reflection losses at ports 1 and 2,
and S21 refers to the insertion loss from port 1 to port 2. From an RF point of view, for
the design of a band stop filter, in the operating band, it is desirable to have the
39
magnitude of |S21| value lower than − 10 dB and a |S11| value near 0 dB. For the case
of band pass filters, these requirements are the opposite.
After reaching the required response, the filters’ behavior near human tissues
was examined. This allows to estimate and compare the performance of these sensors as
glucose measurement devices. For this purpose, a model of the human body is
considered within the simulator as shown in Fig. 24 [44]. To reduce the simulation time,
the human model is dissected into smaller and simpler layers as explained in the
following chapter.
Figure 24: Human model in ANSYS [44].
40
6.3. Design Considerations
Many challenges need to be addressed, while developing any non-invasive and
continuous glucose measurements technique. First, high sensitivity and selectivity to
glucose variations is needed; it is essential that the proposed system is capable of
sensing and detecting BGLs in spite of possible variations in other biological
constituents that may induce undesired effects on the measurements. The sensitivity of
the sensors is linked to both distribution and magnitude of the induced electric field
across the resonating structure. In fact, better sensitivity is achieved by inducing
strengthened fields across the largest possible area [42]. The main target of this research
is to maximize the sensitivity of the proposed sensors by increasing the distribution and
intensity of the induced electric fields.
The proposed device must also meet the standard accuracy as required by the
international organization for standardization (ISO). The most recent version,
ISO:15197:2013, specified that for BGLs lower than 100 mg/dL, an accuracy of ± 15
mg/dl should be reached and for BGLs of 100 mg/dL or more, an accuracy of ± 15% is
acceptable as shown in Table 5 [45].
Additionally, the suggested device should be compact, light and wearable.
Furthermore, several environmental factors must also be considered, including ambient
and body temperature, humidity as well as body movements.
41
Table 5: FDA's accepatable accuracy for glucose measurements devices [45].
Glucose Concentrations Criteria

Requirements ≥ 100 mg/dL 95% within ± 15%
for blood
glucose
monitoring
< 100 mg/dL 95% within ± 15 mg/dl
systems for
self-testing
Consensus error Entire Range 99% in Zones A and B

grid analysis
6.4. Proposed Sensor #1
Initially, a narrow band pass filter is designed and tested. This sensor is an SRR-
based band pass filter that consists of two gap-coupled split ring resonators. SRRs are
left handed metamaterials that exhibit negative values of magnetic permeability and
permittivity near the resonant frequency. These resonators are electrically small
structures, which means that the ring’s perimeter is less than λ/2 at resonance. For these
small quasi-static resonators, resonances are initiated due to a combination of
inductances and capacitances [46]. The equivalent circuit of the structure is illustrated in
Fig 25. In addition, the resonant frequency is calculated using Eq. 26 [47]. For the
proposed design, Ls = 20.6 nH, Cs = 0.21 pF and f = 2.4 GHz.

1
f = 2π×√Ls×Cs (26)
42
Figure 25: Equivalent circuit model of SRRs [47].
In order to maximize coupling, the gap between the adjacent structures must be
minimized. However, due to fabrication constraints, we are restricted to a minimum
value of 0.18 mm. To further increase coupling from the transmission line to the largest
resonator, an enhanced coupling periphery is considered as shown in Fig. 26. This
structure improves coupling by increasing the coverage area between the two entities
[41]. The dimensions of proposed design using a 1.6 mm-thick Rogers 5880 substrate
are summarized in Table 6. The distribution of the induced fields over the top layer is
illustrated in Fig 27. The maximum achieved electric field intensity is 4.85 × 104 𝑉/𝑚.
ro
ri
W0
W
Figure 26: Proposed split ring resonator design with enhanced coupling on periphery.
Image obtained from HFSS simulations.
43
Table 6: Dimensions of the proposed BPF.
Parameter Dimensions (mm)
𝐿 16
𝑊 16
𝑟𝑜 5.2
𝑟𝑖 4.6
𝑔 0.2
𝑠 2
𝑊𝑜 5
Figure 27: Electric field distribution on the top layer of the proposed SRR-based filter.
A prototype of the SRR-based BPF is realized for verification of the simulated
results. The prototype is fabricated through chemical etching technique on a Roger
RT/Duroid5880 substrate as shown in Fig. 28. Fig. 29 demonstrates a good agreement

44
between the simulated and measured results of the transmission and reflection
coefficients.
Figure 28: Fabricated prototype of the SRR-based filter.
Figure 29: Simulated and measured response of the SRR-based filter.
45
The second proposed sensor is a novel broad-band reject filter design that
employs log-periodic distributed complementary open loop resonators (OLRs). The
proposed filter is designed to be implemented as a sensitive, non-destructive and
compact sensor for BG monitoring over a broad-band frequency range. It also enables
estimating the dielectric constant using multiple features, which leads to a low
prediction error.
Log periodic structures are widely used in order to increase the bandwidth of a
microwave structure. In the literature, one design that is discussed in [48] resorts to
three complementary circular rings that are etched at the top layer in order to produce a
broad rejection band. The filter in the corresponding study exhibits a large scaling factor
of 0.98, which limits its bandwidth (fractional bandwidth ~ 20%).
6.5.1. Design Structure
The proposed design is a double-sided microstrip structure that operates as a
broad-band reject filter. The top and bottom layers of the design are shown in Fig. 30.
6.5.1.1. Top layer
The top layer consists of an exponentially tapered transmission line that
couples the magnetic flux density to the underneath resonators. The feed line is
optimized based on the tapering techniques discussed by the author in [36] to better
46
enhance the broad-band operation of the filter based on Eq. 27 and Eq. 28.
Z(z) = Zoea×z (27)
1 Z
a = L × ln(Z l ) (28)
o
Zl and Zo are the impedances to be matched and 𝐿 is the length of the line. By
setting Zl = 100 ohms, Zo = 50 ohms and L = 3cm, the impedance and width of the
line at a specific position can be calculated. The values of Z and W for some positions
along the line are shown in Table 7. The width is computed based on Eq. 18.
L
Wo
W
Wi
W1
W2 W3 W4
S1
S2 S3
L1
L2
L3
L4
g
Figure 30: Proposed log-periodic broad-band reject filter.
Table 7: Dimensions of the designed tapered line.
L (cm) Z (ohms) W (mm)

0 50 1.88
1 63 1.208
2 79.37 0.709
3 100 0.367
47
6.5.1.2. Bottom layer
The bottom layer of the filter is a defected ground plane (DGS) that includes
four complementary OLRs. A DGS has a defect integrated in the ground plane which
alters the uniformity of the plane. This defect manifested as a slot disturbs the shielding
current distribution, which increases the inductance and capacitance of the line. The
circuit area of DGS is relatively small compared to other structures. DGS provides
sharp selectivity at cutoff frequencies with excellent rejection in the stop band and
minimum ripples in the pass band. The stop band response can be further enhanced by
increasing the number of cells (slots). In this design the number of cells is equal to four.
DGS is modeled as RLC equivalent components in series with the transmission line to
which it is coupled as demonstrated in Fig. 31. The input and output impedances are
those of the line section, and the values of the RLC model are determined by the
dimensions of the introduced slots as well as their positions relative to the transmission
line. The LC components determine the resonant frequency of the structure.
Figure 31: Equivalent model of a defected ground structure.
48
In the proposed design, the dimensions and spacing of the OLRs follow a log-
periodic distribution as given in Eq. 29, where τ is a scaling factor that affects the
desired impedance bandwidth B for the four required OLRs in the proposed design [49].
Moreover, the electrical length of the largest OLR is taken to be one-half the
wavelength of the lowest desired frequency of operation as shown in Eq. 30. The
dimensions of the suggested filter configuration, for τ=0.88 , and using a 1.27 mm-thick
Rogers 3006 substrate are presented in Table 8.
Wn+1 Ln+1 Sn+1 1

= = =τ (29)
Wn Ln Sn
λmin vp
Lmax = = (30)
2 2×f
Table 8. Dimensions of the proposed log-periodic BRF designed on a 1.27 mm-thick

Rogers 3006 substrate.
Parameter Dimensions (mm) Parameter Dimensions (mm)
L 60 W1 20
W 18 W2 17.6
WO 1.9 W3 15.5
WI 0.35 W4 13.6
L1 14.7 S1 13.2
L2 13 S2 11.6
L3 11.4 S3 10.2
L4 10 𝑔 2
49
6.5.2. Design Features
6.5.2.1. Electric field distribution
To upsurge the distribution of fields, the configuration of the embedded resonators is
modified as shown in Fig. 32. This helps spread the induced fields across the ground
plane, and hence causes a higher interaction with the loading MUT. Furthermore, by
perturbing the resonators, the magnitude of the induced fields tends to increase thereby
leading to enhanced sensitivity levels. The advantage of the modified OLR in terms of
sensitivity is illustrated in Fig. 32, where the maximum attained value of electric fields
increased from 8 × 103 𝑉/𝑚 to 5 × 105 𝑉/𝑚.
Perturbation
Figure 32: Electric field distribution of the traditional and modified OLR.
50
6.5.2.2. Size reduction
Reducing the overall size of the filter requires the implementation of
miniaturization techniques such as line meandering. This concept is based on folding a
conductor back and forth to have a miniaturized structure. By executing this approach,
the wave is not able to cross the specified distance in a straightforward fashion. Instead
it must traverse the straight-line several times. This increases the curvature of the lines
resulting in an increase in the fringing of fields, which makes the microstrip line appear
electrically longer. Therefore, a smaller physical length is required for the same
resonant frequency, leading to a smaller, more compact structure. A Meander line
includes multiple turns comprising vertical and horizontal sections (Fig. 33). In this
design, the proposed resonators consist of eight turns uniform meander lines and the
dimensions of the turns are optimized by simulation. The size of the modified OLR is
30 % less than that of the conventional structure at 1.43 GHz. In addition, the relatively
high dielectric constant of the substrate reduces further the size of the filter.
Figure 33: Four turns uniform meander line.
6.5.3. Fabrication and Measurements
To validate the performance of the proposed filter in carrying out glucose
sensing processes, a prototype is fabricated using the Computer Numerical Control

51
milling machine on a 1.27 mm-thick Rogers 3006 substrate as shown in Fig. 34. The
complete size of the design is 1.8 cm × 6 cm. A good agreement between the simulated
and measured S-parameters of the fabricated filter is attained as demonstrated in Fig.
35.
Figure 34: Top and bottom layers of a fabricated prototype of the proposed log-periodic
BRF built on a 1.27 mm-thick Rogers 3006 substrate.
Figure 35: Simulated and measured response of the proposed log-periodic BRF built on
a 1.27 mm-thick Rogers 3006 substrate.
52
6.5.4. Alternative Design
The proposed filter is envisioned as a future wearable glucose sensor to be
placed on the human body. Although the size of the device is quite compact, however it
is interesting to build it on a thinner and more flexible substrate. For this purpose, a
modified version of the design was simulated on three different substrates: Rogers 3003,
Polyethylene and Polyimide. The properties of these substrates are presented in Table 9.
The dielectric constants of the considered substrates are quite similar. However, the
properties of these substrates highly differ in terms of thickness and loss tangent.
Accordingly, the dimensions of the feeding line and resonators are adjusted and
optimized using HFSS in order to achieve a similar response to the one realized with
Rogers 3006. It is worthy to mention that for PET and polyimide, silver is used instead
of copper to model the conductive traces. This is due to the fact that traditional
manufacturing methods cannot be used to etch designs on these substrates. Instead,
inkjet printing could be considered as a fabrication method and this process uses silver
nanoparticle as a conductive ink. This ink has a lower conductivity (5 × 106 𝑆/𝑚)
compared to copper (5.8 × 107 𝑆/𝑚).
Table 9: Characteristics of several flexible substrates.
Substrate Thickness ( ) Permittivity Tan ( )

Rogers 3003 250 3 1e-3
PET 136 2.99 5.7e-3
Polyimide 25 3.5 8e-3
53
The return loss and insertion loss responses of the filter designed using the three
substrates are shown in Fig. 36. The filter designed using the Rogers 3003 substrate
provides the best response in terms of bandwidth and return loss levels. This is mainly
attributed to the relatively lower loss tangent of the substrate and the high conductivity
of copper compared to silver ink.
Figure 36: Response of the log-periodic BRF using three flexible substrates.
54
A prototype of the log-periodic filter is therefore realized on a 0.25 mm-thick
Rogers 3003 substrate as shown in Fig. 37. The dimensions of the design are presented
in Table 10. A good agreement between the simulated and measured S-parameters of
the fabricated filter is attained as illustrated in Fig 38.
Figure 37: Top and bottom layers of a fabricated prototype of the proposed log-periodic
BRF built on a 0.25 mm-thick Rogers 3003 substrate.
Figure 38: Simulated and measured response of the proposed log-periodic BRF built on
a 0.25 mm-thick Rogers 3003substrate.
55
Table 10: Dimensions of the proposed log-periodic BRF designed on a 0.25 mm-thick
Rogers 3003 substrate.
Parameter Dimensions (mm) Parameter Dimensions (mm)
L 65 W1 16.15
W 30 W2 14.2
WO 0.63 W3 12.5
WI 0.35 W4 11
L1 22.8 S1 15.7
L2 20 S2 13.8
L3 17.6 S3 12.15
L4 15.5 𝑔 1.3
The third sensor proposed in this thesis is a biologically inspired tunable octa-
band reject filter which consists of a feed line on the top layer and eight slots embedded
in the defect ground plane. These complementary resonators are oriented in such a way
to produce multiple narrow bands instead of just one wide band. The eight stop bands
are distributed between 1.5 GHz and 2.4 GHz and are separated by seven pass bands.
These multiple bands are used to sense glucose variations.
As a first design iteration, only two slots are employed as illustrated in Fig. 39.
These slots follow the distribution of the arms’ ulnar arteries. Accordingly, the width of
each resonating structure must relate to the diameter of the arteries. In [50], a research
study conducted on 251 adult patients showed that the average diameter of the ulnar
artery is 2.4 ± 0.4 mm for the right arm and 2.3 ± 0.3 mm for the left one. Furthermore,
56
the minimum diameters encountered for this artery are 1.3 mm and 1.5 mm for the right
and left arms respectively. Based on that, the dimensions of the slots are optimized and
are presented in Table 11. This structure exhibits a dual-band reject response as
illustrated in Fig 40.
L L
Wsb
W0
W
W
Wss
Figure 39: Top and bottom layers of the proposed dual-band reject filter.
Figure 40: Response of the dual-band reject filter.
57
Table 11: Dimensions of the dual-band reject filter.
L 30
W 25
Wo 1.9
Wsb 1.8
Wss 1.6
6.6.1. Design Structure
To increase the number of bands, each slot is subdivided into four slots of equal
width but different lengths. This configuration increases the number of resonances for
the same physical size. Further increase in the number of slots within the same space
(diameter of veins) would require decreasing the width of the slots below 0.18 mm,
which makes the fabrication process quite complex. The length of each resonator is
optimized to achieve the target resonant frequency. The modified ground plane structure
is presented in Fig. 41. The same structure is superposed with the distribution of arteries
of a human arm as shown in Fig. 42.
58
L
Wsb
W
Wss
Figure 41: Bottom layer of the proposed octa-band reject filter.
Figure 42: Proposed sensing structure superposed with the topology of the lower human
arm.
The simulated response of the proposed structure using the previously used
transmission line is shown in Fig. 43. From this figure, it is clear that the S11 level
drifted apart from the desired 0 dB level, especially for higher frequencies. This could
only mean that the regular 50 ohms transmission line is not suitable to feed the eight
59
slots. Accordingly, the challenge in this design is to provide a simple structure capable
of efficiently feeding the slots in order to enhance the return loss levels. This is
addressed by relying on two approaches.
Figure 43: Response of the proposed octa-band filter using a regular 50-ohms
transmission line.
The first approach is based on increasing the width of the transmission line from
1.9 mm to 3.4 mm as illustrated in Fig. 44. This is equivalent to decreasing the
impedance of the line from 50 ohms to 35 ohm. The dimensions of the proposed
structure are presented in Table 12. Using this topology the feedline is able to cover all
the slots. This is essential to enhance the levels of the reflection coefficients.
60
Wo
Wi
Figure 44: Tapered feeding line topology.
Table 12: Dimensions of the proposed tapered feeding line topology.
Wo 1.9
Wi 3.4
Another topology used to enhance the response of the octa-band filter consists of
implementing a rectangular resonator at the top layer near the transmission line as
illustrated in Fig. 45. This resonator acts as a relay that receives the electric field from
the 50 ohms line in order to feed the eight slots in the ground plane. The dimensions of
this feeding network are presented in Table 13.
61
L
Wo
W
Figure 45: Proposed resonator-based feeding line topology.
Table 13: Dimensions of the proposed resonator-based feeding line topology.
𝐿 35
𝑊 20
𝐿𝑓 20
𝐿𝑐 7.5
𝑊𝑜 0.63
𝑊𝑐 4.66
𝑔 0.2
62
6.6.2. Sensitivity
The electric fields distribution of the two-slot structure and the proposed design
are illustrated in Fig 46 at several resonant frequencies. This distribution proves that the
proposed eight-slot configuration increases the electric field intensity of the structure.
This is mainly attributed to the high density of the eight concentrated nested slots. In
fact, each slot, on its own, contributes to a high field intensity at a specific frequency
band. Some values of the electric field intensity for the octa-band structure are
summarized in Table 14.
Figure 46: Electric field distribution of the dual-band at f= 2.12 GHz and f= 2.34 GHz
and octa-band filters at f=1.8 GHz and 2.2 GHz.
63
Table 14: Maximum attained electric field intensity for the octa-band filters at different
frequencies.
Frequency (GHz) E-Field (V/m)

1.6 2.79 × 105
1.85 3.76 × 105
2 5.6 × 105
2.1 6.97 × 105
2.25 1 × 105
2.5 1.35 × 105
6.6.3. Performance
From an RF point of view, the performance of the octa-band filter can only be
compared with some quint-band and sext-band filters implemented in wireless
communication systems. The proposed methods suffer in simultaneously satisfying all
the required design conditions. This is mainly attributed to the limited degrees of
freedom in the design parameters. In some publications, researchers have worked on
enhancing the system performance, but at the expense of large circuit sizes [51], or
complex structures [52], [53] and [54] or both [55] as they usually rely on implementing
a pair of resonators to generate each band. The characteristics and performance of
several multiband filters and our proposed design are summarized in Table 15.
Compared to the literature, the design developed in this work provides higher number of
bands with comparable levels of S-parameters, using a simple and more compact
configuration.
64
Table 15: Performance of different multi-band filters found in the literature.
Bands (GHz) IL (dB) Bands / Size (cm)

Resonator
[51] 0.6/0.9/1.2/1.5/1.8 2.8/2.9/2.9/2.6/2.3 1 17.1x1.49
[52] 1.5/2.5/3.5/4.5/5.8 1.5/1.8/0.9/1.2/2.5 0.5 3.55x2.52
[53] 0.9/1.2/1.4/1.7/2/2.4 2.3/2/2.3/2.7/2.2/2 0.5 5.1x3.1
[54] 2.1/3/4/4.7/7.2 0.98/1.78/1.22/1.77/2.39 1 2.28x1.07
[55] 0.8/1.2/1.4/1.8/2.2/2.5 2.9/2.34/2.59/2.24/2.67/2.64 0.5 17.1x1.43
Proposed 1.55/1.66/1.75/1.84/ <2 dB 1 2x1.5
Design 2/2.15/2.3/2.4
6.6.4. Fabrication and Measurements
The first feeding topology is used to realize a prototype of the proposed octa-
band filter on a 1.27 mm-thick Rogers 3006 substrate as shown in Fig. 47. The second
feeding topology is used to realize a prototype of the proposed filter on a 0.25 mm-thick
Rogers 3003 substrate as shown in Fig. 48. A good agreement between the simulated
and measured scattering parameters is reached as shown in Fig 49 and Fig. 50.
Figure 47: Top and bottom layers of a fabricated prototype of the proposed octa-band
filter built on a 1.27 mm-thick Rogers 3006 substrate.
65
Figure 48: Top and bottom layers of a fabricated prototype of the proposed octa-band
filter built on a 0.25 mm-thick Rogers 3003 substrate.
Figure 49: Simulated and measured response of the proposed octa-band reject filter built
on a 1.27 mm-thick Rogers 3006 substrate.
66
Figure 50: Simulated and measured response of the proposed octa-band reject filter built
on a 0.25 mm-thick Rogers 3003 substrate.
6.6.5. Tuning and Reconfiguration
The eight bands of the filter are reconfigured by implementing a varactor diode
as discussed in this section.
67
6.6.5.1. Reconfigurable microwave circuits
Reconfigurable microwave circuits are multifunctional devices that have the
ability to change their characteristics such as the frequency of operation, and are used in
many applications that require agility and dynamic response. Reconfiguration is
achieved using electronic components such as PIN diodes, RF- (Micro Electro
Mechanical Switches) MEMS and varactors that are connected to the circuitry in order
to change the electromagnetic behavior of the device. These electronic components
change the electrical length of the RF structure by either redistributing the currents (PIN
diode, RF-MEMS) or loading the structure by a variable capacitance (varactor). The
change in the electrical length of the RF circuit causes a shift in the frequency of
operation. The switch position and the biasing network are critical in order to achieve
the best possible tuning.
For the proposed filter design, the equivalent circuit of the bottom layer is a
parallel RLC resonator. This suggests that by adding a varactor diode between the
internal and external metallic regions of the slots, the equivalent capacitance of the
structure can be tuned. In practice this can be better achieved by placing the varactor
diode on the top substrate side and connection to the bottom side through metallic vias.
6.6.5.2. Proposed tunable structure circuits
In the proposed design, the SMV 1705-079LF is used to reconfigure the
operating frequencies [56]. Fig. 51 shows the spice model of the varactor. The values of
the components are Ls = 0.8 nH, R s = 0.32 ohms and Cp = 0.5 pF. The capacitance
68
value of the varactor diode Cj can be tuned from 31.5 pF to 5.2 pF, by varying the
reverse voltage from 0 to 5 Volts. Biasing the varactor diode requires the use of
an 470 nH inductor and a 10 pF capacitor. The basic function of the inductor is to
prevent the RF signal from passing to the power supply, and the capacitor is used
prevent shorting DC current.
Figure 51: Spice Model of the SMV 1705-079LF [56].
The topology of the proposed tunable filter is represented in Fig. 52. It
comprises the previously discussed eight slots etched in the ground plane beneath the
feeding line, a varactor diode with a variable capacitance, an RF choke and a lumped
capacitance for biasing. All the electrical components are soldered in the upper substrate
side to prevent any interference with the sensing area. The varactor diode and the
capacitance are connected in parallel and this combination is placed between the
internal and external metallic regions of one set of slots using two vias. The Cathode of
the varactor is connected to the power supply through the RF choke. The equivalent
69
circuit of this structure is demonstrated in Fig. 53. The fabricated prototype is presented
in Fig. 54, and the simulated and measured results are presented in Fig. 55.
Figure 52: Topology of the proposed tunable filter.
Figure 53: Equivalent circuit of the proposed tunable filter.
70
Figure 54: Fabricated Prototype of the proposed tunable filter.
(a)
(b)
Figure 55: Simulated and measured response of the proposed reconfigurable octa-band
reject filter built on a 1.27 mm-thick Rogers 3006 substrate. The transmission
coefficient is shown for a) V=0 Volts, b) V=5 Volts and c) several voltage values.
71
6.7. Discussion
This chapter presents the design of three different RF filters. These circuits are
initially designed using HFSS and then built on a variety of flexible and rigid substrates.
The performance of these filters as glucose sensors is assessed by monitoring their E-field
distribution over the sensing area. It was noted that the octa-band filter achieved the
highest E-field intensity, followed by the broad-band filter. The reflection and
transmission coefficients of the different filters were also validated by measurements.
72
CHAPTER 7
SIMULATION AND MEASUREMENT OF THE PROPOSED

EM-BASED GLUCOSE SENSORS
7.1. Introduction
This chapter presents the simulation results of the proposed RF sensors along
with the measurement results for in-vitro and ex-vivo and in-vivo studies. Simulation
results include the integration of both a single and multi-layered tissues model. For in-
vitro measurements, serum is used to mimic the blood. For ex-vivo studies, rat tissues
are placed as a separation between serum and the sensor. Finally, for in-vivo studies,
OGTT is performed twice for six different patients.
7.2. Simulation
To investigate the ability of the filters to perform as glucose meters, several
human tissue models were considered for analysis. The common act in these
simulations is to vary the relative permittivity of the blood layer, so that it corresponds
to the change of the BGL. For the simplest single-layered model, blood is modeled as a
rectangular box of height h=4mm placed at a distance of 2 mm from the sensing area of
the filter. Rectangular shapes are utilized for purposes of reduced simulation time. A
more complex model is the one shown in Fig 56. It includes the main biological layers
encountered in a human arm: skin, fat, blood, and bones. The thickness of each of these
four layers is presented in Table 16. This model was placed 4.4 mm away from the
73
sensing area of the filter, meaning that the blood layer is distanced by 7.4 mm away
from the sensor.
Figure 56: Multi-layered human tissues model.
Table 16: Thickness of the layers used in the model.
Layer Height (mm)
Skin 1.5
Fat 1.5
Blood 4
Bones 10
Placing these layers as superstrates near the sensors causes a shift in the
operating frequency of the device under test. When analyzing the device response, we
sweep the dielectric constant of the blood layer from 60 to 75 to reflect some variations
in the BGLs. The S21 responses of the sensors for the different values of the dielectric
constant are presented in Fig. 57. These include the results for both cases of the device
loaded by the single-layered and multi-layered models. The S21 phase and magnitude
variations of the proposed sensors, when loaded by the multi-layered model, are
illustrated in Fig 58 at one of the operating frequencies. It is worth noting that similar
trends were obtained at other operating frequencies (not illustrated here).

74
Figure 58 demonstrates a linear and monotonic behavior for the proposed
sensors in response to the variations in relative permittivity; the sensors’ responses
exhibit a clear correlation with the material’s dielectric constant. For the SRR-based
filter, this behavior is restricted to the narrow operating band. For the broad-band filter,
this linear trend is observed over a wide range of frequencies especially between 1.75
GHz - 2.75 GHz. In the case of the multi-band filters, the linear behavior is observed
around all the resonant frequencies. Based on the observed trends, we find that the
performances of the broad-band and octa-band sensors to be very favorable for purposes
of glucose sensing applications. Furthermore, their sensitivity can be sampled across
several frequencies in comparison to narrowband filters. Accordingly these two sensors
are considered for further analysis as discussed in the following sections.
7.3. Experimental Setup
The proposed broad-band and octa-band sensors are tested in in-vitro, ex-vivo
and in-vivo scenarios. The basic experimental setup consists of three elements, a
portable vector network (VNA, N9914A, Keysight Technologies), RF cables and a
sensor. The VNA applies to the sensor an RF signal whose frequency is swept over a
predefined frequency range with an output power of about -15 dBm. For each
frequency in the specified range, the VNA measures the reflected signal at both ports of
the sensor along with the transmitted signal between the two ports. The changes in the
microwave parameters such as resonance frequency, reflection coefficient and insertion
75
loss are then tracked using an algorithm. The S-parameters data is represented using
smith chart formats, and is collected in Cartesian forms.
(a)
(b)
(c)
Figure 57: S21 response of the proposed sensors for the single-layered (left) and multi-
layered models (right), a. narrowband sensor, b. broad-band sensor, c. octa-band sensor.
76
(a)
(b)
(c)
Figure 58: Magnitude and phase variation of the proposed sensors when loaded by the
multi-layered model, a) the narrowband sensor, b) the broad-band sensor, and c) the
octa-band sensor.
77
At each frequency, ten different measurements and hence microwave data points
are obtained from the VNA and their average is reported in order to minimize noise
effects and reduce the randomness in the measurements. The same process was repeated
for three different experiments.
 For the first experiment, the sensor is calibrated to monitor variations in
glucose inside a foam container filled with 7 mL of serum whose glucose
density is being varied gradually.
 In the second experiment we introduced rat tissue as a separation layer
between the serum and the sensor.
 Finally, in the third experiment we tested the ability of the sensors to detect
variations in human BGLs. This experiment was conducted for six different
volunteers.
7.3.1. Serum Measurements
We propose for our setup placing a fixed, serum-filled foam container on top of
the sensor as shown in Fig. 59. Serum is a liquid that is similar in composition to the
blood plasma; however, it excludes the clotting factors of blood. Dextrose powder is
then added to the solution to alter the glucose concentrations. The considered
experimental procedure consists of extracting part of the liquid, adding dextrose,
applying vortex mixing to accelerate the dissolving process, adding the mixture to the
container, manually mixing the whole solution, and finally wait ten minutes to ensure a
homogeneous entity before reading the S-parameters. This process ensures that the
78
setup and volume of the serum remains almost fixed during the whole experiment,
while the glucose level is incremented gradually.
VNA
Serum
Foam
Container 7 mL of Serum
2 mm-thick Foam
Proposed (Filter) Sensor
Figure 59: Experimental setup for serum measurements.
As a first test, baseline measurements were performed. This was conducted by
measuring the sensors’ S-parameters in the absence of the foam container, then by
placing an empty container on the sensor, and finally by filling the container with
serum. From these measurements, it was verified that the sensors under test maintain
their free space responses, with and without the empty foam container. The shift in
frequency occurs only when serum is added.
The response of the proposed sensors for different glucose concentrations is
shown in Fig. 60. Their response displays a clear correlation with glucose levels at
different frequencies as illustrated in Fig. 61 for the broad-band and in Fig. 62 for the
octa-band sensors.
79
(a) (b)
(c) (d)
Figure 60: Response of the proposed sensors for different glucose concentrations, a)
S21 magnitude of the broad-band sensor, b) S21 phase of the broad-band sensor, c) S21
magnitude of the octa-band sensor, d) S21 phase of the octa-band sensor.
80
Figure 61: Correlation between the response of the broad-band sensor and the glucose
concentrations at different frequencies.
81
Figure 62: Correlation between the response of the octa-band sensor and the glucose
concentrations at different frequencies.
7.3.2. Animal Tissues and Serum Measurements
The setup of this experiment is almost identical to the one presented in the
previous section. The same foam container was considered, and glucose levels were
varied in the exact fashion. The only difference is that now the serum-glucose solution
is being tested in a multi-layer setup, where the animal tissues separate the sensor from
82
the liquid. The magnitude and phase variations are shown in Fig. 63, and it’s clear that
the linear correlation is still maintained.
(a)
(b)
Figure 63: Magnitude and phase variations of a) broad-band, and b) octa-band sensors.
83
7.3.3. Clinical Measurements
7.3.3.1. Study design
Subjects were recruited to participate in the clinical trial after signing a consent
form, previously approved by the Institutional Review Board.
7.3.3.2. Study subject
Subjects were considered eligible for the study if they were between 18 and 70
years of age, and able to provide informed consent. There were no restrictions on either
race, sex or ethnicity. Substance abuse, lactation, pregnancy, and being part of an
interventional trial were the exclusives criteria. In phase one of the study, only healthy
subjects with HbA1c levels less than 6%, normal blood pressure and no sign of
dyslipidemia were included.
7.3.3.3. Procedure
The patients arrived to the clinical study unit in the morning after fasting for at
least 8 hours. Measurement of the blood glucose levels was initially performed using
the standard techniques that assess glycemia, using a glucometer of ACCU-Check.
Afterwards, the sensor under test was attached to the lower arm of the volunteer as
illustrated in Fig. 64. Subsequently, glucose was orally ingested as a concentrated
glucose drink that contains 75 g glucose dissolved in 200 mL of water. This induces a
hyperglycemic excursion to a target BGL of 170 -220 mg/dL. These levels are expected
to fall back within 2 hours. Readings from the sensor were collected each 5 min, and
84
reference BGLs were measured at intervals of 15 minutes using the glucometer. During
the process, patients were asked to stay tranquil and with no physical movements. The
room temperature was 23 ± 0.5 ◦C.
Figure 64: Clinical trials setup.
7.3.3.4. Results
The response of the proposed sensors showed a clear correlation with the BGLs
as demonstrated in Fig. 65 and Fig. 66 for all the patients. The solid red line represents
the reference glucose levels, and the dashed blue line represents the normalized
response of the sensor at one frequency. Moreover, we note that for a given patient there
was a clear correlation between the response at some specific frequencies and the
glucose levels not only for the first OGTT but also for the second test as demonstrated
in Fig 67.
Furthermore, a statistical prediction model was generated on the basis of the
sensor signals collected during the two visits of the volunteers and is discussed in the
next chapter.
85
Figure 65: Correlation between the response of the broad-band sensor and the blood
glucose concentrations.
86
Figure 66: Correlation between the response of the octa-band sensor and the blood
glucose concentrations
87
(a)
(b)
Figure 67: Correlation between the response of a. broad-band sensor, and b. octa-band
sensor between two different OGTTs for the same subject and for the same frequency a)
f=1.4 GHz and b) f=1.33 GHz.
7.3.4. Measurements Accuracy
In the presented experimentation, there exist several sources of error. For
instance the position of the sensor might vary between the measurements. Furthermore,
the data collected is always bounded to the accuracy of the underlying measurement
tools both in terms of the utilized glucometer and the lab vector network analyzer.
88
7.3.4.1. Glucometer
The ACCU-CHEK Performa glucometer is a calibrated system used to measure
blood glucose concentrations within the 10-600 mg/dL range [57]. However, the meter's
accuracy depends upon a lot of physical and pharmacological factors in addition to the
condition of the patient and the strips used. For the considered meter, the accuracy of
the system is shown in Table 17.
Table 17: Accuracy of the ACCU-CHEK Performa glucometer [57].
Glucose Concentrations Glucose Concentrations

< 𝟏𝟎𝟎 𝒎𝒈/𝒅 ≥ 𝟏𝟎𝟎 𝒎𝒈/𝒅
Within Within Within Within Within Within

±5 mg/dL ±10 mg/dL ±15 mg/dL ±𝟓 𝒎𝒈/ ±10 mg/dL ±𝟏𝟓 𝒎𝒈/
𝒅 𝒅
81.5% 97.0% 99.4% 59.3% 91.0% 99.1%
7.3.4.2. S-parameters
The accuracy of S-parameters measurements depends on the precision of the
VNA, the calibration, in addition to the quality of the cables.
7.3.4.2.1. Vector Network Analyzer
Keysight FieldFox RF analyzer N9914A was used for the measurements. The
S-parameter value for a given frequency may fluctuate around the correct one. We
treated this as a random source of error, and to eliminate this randomness, ten readings
of the signal were collected subsequently at a specific frequency. The averaged value
was then considered for further analysis.
89
7.3.4.2.2. Calibration
A typical calibration will move the measurement reference planes to the very
ends of the test cables to account for the phase difference and losses in the cables.
However, when dealing with lengthy measurements as the ones presented in this work,
the quality of calibration deteriorates with time courtesy of large cables’ movement and
bending. The collected data is therefore smoothed by applying a moving average filter.
7.4. Discussion
In this chapter, the previously designed RF sensors are tested using simulation
models and measurements. The responses of the proposed sensors showed good
correlation with the variations in the dielectric constant and glucose levels. The
collected data from the in-vitro and in-vivo studies will be used to develop and test
several regression models in chapter eight.
90
CHAPTER 8
DEVELOPMENT OF A REGRESSION MODEL
8.1. Introduction
This chapter addresses the development of a mathematical model for predicting
glucose levels from the data recorded by the proposed sensors. Particularly, several
linear and nonlinear methods are applied to the collected data to build the model. The
aim here is to assess the performance of these regression methods, and to find the best
prediction model.
8.2. Regression Analysis
Regression is the estimation of an output variable (blood glucose level in our
case) from a set of measured input variables (measured sensor device physical
parameters such as the S-parameters phase and magnitude at different frequencies). This
is performed by establishing a mathematical relation between the input and output
variables. In case of linear regression, the mathematical relation is a simple linear
equation that uses either one (univariate) or several (multivariate) explanatory variables
x to describe the behavior of the dependent variable y. For the case of k explanatory
variables, the linear function is denoted as in Eq. 31.
𝑦 = 𝛽1 𝑥1 + 𝛽2 𝑥2 + ⋯ + 𝛽𝑘 𝑥𝑘 + 𝑒(𝛽1 , … , 𝛽𝑘 ) (31)
Alternatively, assuming a sample of 𝑇 observations, Eq. 31 can be expressed in
its general form as in Eq. 32. Each column of 𝑿 contains 𝑇 observations of an
91
explanatory variable xk . 𝒚 is the vector of responses, 𝑒(𝛽) is the vector of error terms,
and 𝛽 is the vector of unknown parameters to be estimated. The objective of regression
is to find the most suitable values for 𝛽 that minimize the resultant prediction error.
𝒚 = 𝑿𝛽 + 𝑒(𝛽) (32)
𝑦1
𝑥11 ⋯ 𝑥1𝑘
𝑦2
𝒚 = [ ⋮ ], 𝑿=[ ⋮ ⋱ ⋮ ]
𝑥 𝑇1 ⋯ 𝑥 𝑇𝑘
𝑦𝑇
For the case of ordinary least squares (OLS), the parameters are estimated by
minimizing the residual sum of squares ||𝒚 − 𝑿𝜷||𝟐𝟐. When the system is
overdetermined and 𝑿𝑻 𝑿 is nonsingular, and hence can be inverted, the unknown
parameters vector β can then be obtained according to Eq. 33.
̂ = (𝑿𝑻 𝑿 )−𝟏 𝑿𝑻 𝒚 (33)

𝜷
For the case of glucose sensing, it is quite difficult to collect a large number of
reference glucose points. Accordingly, while the number of covariates is high, the
number observation points is quite low. This means that we are dealing with an
undetermined system where K>>T. Consequently, OLS can’t be used for prediction. To
deal with this problem, sparse regression and/or feature selection methodologies are
employed. We particularly focus on linear regression techniques such as partial least
squares (PLS) [58], and the least absolute shrinkage and selection operator (LASSO)
[59] method as well as non-linear regression methodologies such as the Gaussian
processes (GP) [60]. Furthermore, dimensionality reduction methods such as principal
component analysis can be employed [61].
92
The procedure used to develop the mathematical regression model is presented in
the following. The main steps include preprocessing, dimensionality reduction
processes, model generation, and model performance evaluation.
8.2.1. Preprocessing
Real measurements data may be inconsistent, incomplete, and containing
measurement errors. The data preprocessing step assembles raw data and transforms it
into a clear format for additional processing. Preprocessing includes multiple steps such
as data collection, transformation and reduction. For purposes of measurement error, we
rely on averaging and smoothing techniques to eliminate random sources of error.
In our case, measurements are performed over 201 frequencies taken over the
operating regions of the sensors. These frequencies range between 0.5 GHz and 3.5
GHz for the broad-band filter, and between 1 GHz and 3 GHz for the octa-band filter.
However, we noticed that neighboring frequencies have similar trends, and to eliminate
redundancy, we sampled uniformly 20 frequencies over the operating ranges. This
resulted in a total of 120 features corresponding to the magnitude and phase of the
sensor S11, S21, S22 parameters at the different frequencies over the operating range of
the device.
Sampled data is then normalized for consistency between the different feature
vectors. The filter’s measured data are scaled and shifted to [0 1]. This is performed for
each measured physical parameter at a given frequency, by subtracting for each
observation point the value of the physical parameter at that frequency from the
93
reference (fasting) glucose level of a specific OGTT at the same frequency. The value is
then divided by the absolute maximum of the difference for all the observation points at
that frequency.
8.2.2. Modeling Techniques
Several regression techniques like Partial Least Squares (PLS), Least Absolute
Shrinkage and Selection Operator (LASSO) and Gaussian Processes (GP), were
considered for prediction. Cross-validation methods were then employed to find the best
model for purposes of prediction.
PLS aims to predict a set of dependent variables (target), from a set of
independent variables (predictors). This is realized by extracting, from the predictors, a
reduced set of orthogonal features while maximizing correlation with the dependent
variable Y. These factors are linearly combined from the original variables, and are used
for estimation [58]. A more advanced version of PLS is locally weighted PLS [62].
LASSO is a linear estimation method that uses the 𝐿1 penalty as a regularization
technique. This penalized regression process shrink the regression coefficients toward
zero, introducing some bias to reduce variability. This is performed by adding a penalty
term to the residual sum of squares as in Eq. 34. The tuning parameter λ controls the
amount of shrinkage. Choosing the right λ is essential to minimize the error in
prediction [59].
̂ = 𝒂𝒓𝒈𝒎𝒊𝒏{ ||𝒚 − 𝑿𝜷||𝟐𝟐 + ∑𝑷𝒋=𝟏 𝑱𝝀 |𝜷𝒋 |}, 𝜷 ∈ ℝ𝑷 (34)

𝜷
94
GP produces flexible, nonlinear, and nonparametric Bayesian models. Prediction
using this method is probabilistic and is related to uncertainties. In contrast to traditional
regression methods, GP does not regulate a unique function on the dataset, but rather
produces a probabilistic distribution over a space of functions with respect to the
dataset. To achieve this objective, the covariance function parameters are adjusted to
maximize the likelihood of the observation points. Consider a GP function h(x), the goal
is to predict its value for a random input vector x. A Gaussian process, like a Gaussian
distribution, is completely specified by a mean and a covariance (kernel) function.
Consequently, the predictive distribution is also Gaussian distributed with mean and
variance [60].
8.2.3. Model and Features Selection
From section 8.2.1., it is obvious that we have a large number of features (120
features compared to 35-46 observations) and we have an undetermined system.
Accordingly, to build a good model, we need to identify the critical features to build our
model.
LASSO inherently identifies the critically features iteratively. For PLS and GP
we rely respectively on feature extraction and selection methods such as the wrapper or
the filter method (based on correlation significance for example. For all these methods,
K-fold cross validation can be employed to identify the model, i.e., set of features (and
kernel functions in case of GP) that result in the lowest error.
95
Feature extraction translates the features from the high-dimensional space into a
lower dimensional space by a set of linear or nonlinear transformations. The most
common linear feature extraction method is principal component analysis (PCA) [61].
Feature selection concentrates on selecting a small subset of features based on
some predefined criteria. For this purpose, we rely on a wrapper method that identifies
the next best feature for a given kernel function in case of GP. We then determine,
based on cross-validation error, the minimum number of features and best kernel
function that results in lowest cross-validation error and hence that determine the best
model. Alternate feature selection methods include sorting the features based on their
maximum correlation or maximum relevance and then performing cross-validation to
find the best set of features. This is referred to as the filter method (not shown here.
8.2.4. K-fold Cross-validation
K-fold cross-validation divides the dataset D into K partitions (folds) of nearly
equal size. For K iterations, K-1 folds are used to train the model, and the remaining
fold is used for testing. During each iteration, the error of the testing fold is calculated.
By averaging these values, the average error is used to help select the best model with
minimum cross-validation error.
96
8.3. Results
8.3.1. Serum Measurements
The data collected from serum measurements using the rigid and flexible broad-
band and octa-band filters were used for testing the different modeling techniques.
Preliminary results showed that GP offers the best performance among the considered
approaches. This is visualized in Fig. 68 where the estimated glucose levels are plotted
versus the reference glucose levels using LW-PLS, GP and LASSO are presented.
Moreover, the mean percentage error for the different modeling methods is shown in
Table 18.
Table 18: Mean percentage error for the glucose levels estimated using LW-PLS, GP,
and LASSO.
Sensor Type LW-PLS GP LASSO
Broad-band Rigid 25.12 16.95 26.07

Broad-band 22.88 11.37 16.18
Flexible
Octa-band Rigid 44.13 16.37 112.53
Octa-band Flexible 19.85 12.82 15.07
97
(a)
(b)
(c)
(d)
Figure 68: Estimated versus reference serum glucose levels using LW-PLS, GP and
LASSO, a. rigid broad-band, b. flexible broad-band, c. rigid octa-band, and d. flexible
octa-band.
98
8.3.2. Clinical Measurements
We developed a statistical prediction model using the sensor signals collected
from in-vivo measurements during the two visits of the volunteers. The model considers
the BGLs as the dependent variable. Without loss of generality in this initial
implementation we do not consider the physical and physiological status of the patients.
The number of observations per patient was quite small in comparison to the number of
features (S-parameter magnitude and phase obtained at the different frequencies). We
had a total of 23 observations per test. For the corresponding glucose levels obtained
from the ACCU-Check glucometer, we have measured only 9 value points recorded
corresponding to 9 out of the 23 observations. The remaining 14 reference glucose
points are interpolated using cubic spline interpolation. The measured sensor data
includes some fluctuations or ripples, courtesy of bending the RF cables for a long
duration. The S-parameters are therefore smoothed by applying a moving average filter
with a span equal to 5%.
We rely on the wrapper method and 10-fold cross-validation to determine the
best features and best suitable kernel functions for a given patient in order to come up
with the best model (lowest error). Five different kernel functions are considered in the
simulations exponential, squared exponential, matern32, matern52, and rational
quadratic. Around 8 to 15 critical features were identified for a given patient to build the
model. To test the performance of the model, we randomly divide the observation points
into 2/3 for training and 1/3 for testing. We report the test error for ten random data sets
using the predetermined features and kernel functions. For this step, the data is initially
99
stratified into three homogeneous groups. This step is essential to make sure that each
partition includes low, medium and high values of reference glucose levels. Following
that, the data is randomly divided into two thirds (31 points) for training and one third
(15) for testing. The 46 points correspond to two OGTT experiments. This step is
repeated ten times to make sure that the majority of the data is used for both training
and testing.
The results of the model for individual BGL profiles are shown in Fig. 69 and
Fig. 70, plotted as function of time and compared with the reference blood glucose
concentration. These profiles show two successive peaks corresponding to the two
OGTT experiments. For each OGTT, we note a rapid increase in glucose values from
the fasting level to a maximum value, and then a decrease. The solid lines are the
invasively measured BGL, and the dashed lines are the estimated values using Gaussian
Process; the dots correspond to the prediction using the different experiments. It is
obvious that the estimated glucose concentration by the proposed sensor matches well
the rate of increase and decrease of glucose concentration. It is also noted that there is
no time delay between the two measurement systems. This proves that the proposed
sensor directly interacts with blood. Furthermore, the mean average percentage error in
prediction is noted to be 3.19% for the broad-band sensor and 1.83% for the case of the
octa-band sensor.
100
Figure 69: BGL profiles of all the patients collected using the broad-band sensor and
estimated using GP. Each plot includes the invasively measured BGL (solid lines), the
estimated BGL using Gaussian Process (dashed lines) and the prediction using the
different experiments (dots).
101
Figure 70: BGL profiles of all the patients collected using the octa-band sensor and
estimated using GP. Each plot includes the invasively measured BGL (solid lines), the
estimated BGL using Gaussian Process (dashed lines) and the prediction using the
different experiments (dots).
102
To better assess the results, the Clarke Error Grid is considered [63]. This
approach is used to evaluate the clinical significance of differences between a given
glucose measurement technique under test (RF sensor in this case) and the intravenous
blood glucose reference measurements. In this work, the reference measurements are
considered to be those detected using the commercial glucometer. The Clarke Grid is a
two-dimensional Cartesian illustration, where the reference values are displayed on the
x-axis and the predicted values are presented on the y-axis. The points laying on the
diagonal signify a perfect agreement between the reference and predicted values. The
points below and above this line designate, respectively, overestimation and
underestimation of the actual values. The grid is also divided into five zones. Points
might lay within region A when the deviancy in the predicted values from the
references doesn’t exceed 20%, or when both the predicted and reference values are in
the hypoglycemic range (<70 mg/dl). The values of this zone are labeled clinically
exact, and are thus described by correct clinical treatment. Region B is linked to benign
errors and is located around zone A. This area includes points that deviates by more
than 20% from the reference values, but don’t lead to inappropriate treatment. Points in
region C lead to unnecessary treatment. Points in region D indicate failure in detecting
hypoglycemia and or hyperglycemia (>180 mg/dl). Finally, points in region E are those
that confuse the treatment of hypoglycemia for hyperglycemia or the other way around.
Clinically, all the values within areas A and B are considered acceptable, whereas the
values in zones C, D and E are potentially dangerous [63].
103
The generated Clarke grids using the data collected from the two sensors are
shown in Fig. 71. For the broad-band log-periodic filter 89.5% of the data are lying in
zone A and only 1.5% in zone B. For the octa-band filter, all the data are in zone A.
These results show exceptional accuracy of the proposed measurement and prediction
methods.
(a)
(b)
Figure 71: Clarke error grid for the data collected using a) broad-band sensor, and b)
octa-band sensor.
104
8.4. Discussion
Several regression models were used in this chapter for estimation purposes.
Gaussian process showed better abilities to predict the glucose levels, in comparison to
LW-PLS and LASSO, based on in-vitro measurements. Furthermore, GP was applied
on the data collected from the clinical trials. The Clarke error grid was used to assess
the results. All the predicted points are clinically acceptable, proving the high sensitivity
of the proposed sensors and the accuracy of GP as a modeling method.
105
CHAPTER 9
CONCLUSION AND FUTURE WORK
There is a need for a method of measuring blood glucose continuously and non-
invasively. The ability of microwave devices to extract the electrical parameters of
material accurately and without direct contact, makes them ideal for this application. In
this thesis several microwave sensors are designed and tested: an SRR-based narrow
band filter, a log-periodic based broad-band reject filter and a biologically inspired
tunable octa-band reject filter. The behavior of the proposed RF circuits as glucose
sensing systems is tested using simulation in addition to in-vitro, ex-vivo and in-vivo
studies. A good correlation between the scattering parameters of proposed sensors and
the variations in glucose levels is attained. Several regression models are also developed
and applied on the collected data. In this context, Gaussian Processes helped achieved
the lowest error in prediction. Examined results using the Clarke error grid demonstrate
that for the broad-band and octa-band sensors 100% of the predicted glucose levels lay
in the clinically acceptable regions.
For future work, the broad-band and reconfigurable octa-band sensors will be
clinically tested on a larger number of volunteers. The regression model can also be
developed further based on the newly collected data. Furthermore, a miniaturized mean
to analyze the data and visualize glucose levels will be considered.
106
REFERENCES
[1] "WHO," 30 october 2018. [Online]. Available:

http://www.who.int/mediacentre/factsheets/fs312/en/. [Accessed 29 March 2019].
[2] "Diabetes," [Online]. Available: https://www.diabetes.org.uk/Diabetes-the-basics.

[Accessed 2019 March 2019].
[3] R. A. Gabbay, "New Developments in Home Glucose Monitoring: Minimizing the

Pain," Canadian Journal of Diabetes, vol. 27, no. 3, pp. 271-276, 2003.
[4] U. Fischer, "Fundamentals of Glucose Sensors," Diabetic Medicine, vol. 8, no. 4,

pp. 309-321, 1991.
[5] E. Wilkins and P. Atanasov, "Wilkins E, Atanasov P. Glucose monitoring: State of

the art and future possibilities," Medical Engineering & Physics, vol. 18, no. 1, pp.
273-288, 1996.
[6] H. Choi, "Recent developments in minimally and truly non-invasive blood glucose
monitoring techniques," in 2017 IEEE SENSORS, Glasgow, 2017.
[7] F. Barbon G. Vinci S. Lindner R. Weigel A. Koelpin "A six-port interferometer

based micrometer-accuracy displacement and vibration measurement radar", IEEE
MTT-S Int. Microwave Symp., 2012.
[8] S. Salman "Pulmonary edema monitoring sensor with integrated body-area network
for remote medical sensing" IEEE Trans. Antennas Propag. vol. 62 no. 5 pp. 2787-
2794 May 2014.
[9] H. Acikgoz and I. Turer, “A novel microwave coaxial slot antenna for liver tumor
ablation”, Advanced Electromagnetics, Vol. 3, No. 1, April 2014.
[10] Nikawa, Y., Michiyama, T., ‘Non-invasive measurement of blood-sugar level by

reflection of millimeter-waves’. Proc. Asia-Pacific Microwave Conf., 2006.
[11] Nikawa, Y., Someya, D., ‘Application of millimeter waves to measure blood sugar
level’. Microwave Conf., 2001. APMC 2001. Asia-Pacific, 2001, vol. 3, pp. 1303–1306.
[12] H. Cano-Garcia, P. Kosmas, I. Sotiriou, K. P. Ioannis, C. Parini, I. Gouzouasis, G.

Palikara and E. Kallos, "Detection of glucose variability in saline solutions from
transmission and reflection measurements using V-band waveguides," Measurement
Science and Technology, vol. 26, no. 12, 2015.
107
[13] Siegel, P. H., L. Yoon, and V. Pikov, "Millimeter-wave non-invasive monitoring of
glucose in anesthetized rats, "in Infrared, Millimeter, and Terahertz waves
(IRMMWTHz), 2014 39th International Conference on, 2014, pp.1-2.
[14] H. Cano-Garcia, I. Gouzouasis, I. Sotiriou, S. Saha, G. Palikaras, P. Kosmas and E.

Kallos, "Reflection and transmission measurements using 60 GHz patch antennas in the
presence of animal tissue for noninvasive glucose sensing," in 2016 10th European
Conference on Antennas and Propagation (EuCAP), 2016.
[15] S. Saha, H. Cano-Garcia, I. Sotiriou, O. Lipscombe, I. Gouzouasis, M.

Koutsoupidou, G. Palikaras, R. Mackenzie, T. Reeve, P. Kosmas and E. Kallos, "A
Glucose Sensing System Based on Transmission Measurements at Millimetre Waves
using Micro strip Patch Antennas," Scientific Reports, vol. 7, no. 1, 2017.
[16] H. Choi, J. Nylon, S. Luzio, J. Beutler and A. Porch, "Design of continuous non-
invasive blood glucose monitoring sensor based on a microwave split ring resonator,"
2014 IEEE MTT-S International Microwave Workshop Series on RF and Wireless
Technologies for Biomedical and Healthcare Applications (IMWS-Bio2014), London,
2014, pp. 1-3.
[17] H. Choi, J. Naylon, S. Luzio, J. Beutler, J. Birchall, C. Martin and A. Porch,

"Design and In Vitro Interference Test of Microwave Noninvasive Blood Glucose
Monitoring Sensor," in IEEE Transactions on Microwave Theory and Techniques, vol.
63, no. 10, pp. 3016-3025, Oct. 2015.
[18] H. Choi, S. Luzio, J. Beutler and A. Porch, "Microwave Noninvasive Blood

Glucose Monitoring Sensor: Human Clinical Trial Results," 2017 IEEE MTT-S
International Microwave Symposium (IMS), 2017.
[19] M. Sidley and J. Venkataraman, “Calibration for Real-time Noninvasive Blood

Glucose Monitoring.” Masters’ Thesis, Department of Electrical and Microelectronic
Engineering, Rochester Institute of Technology, 2013.
[20] M. Sidley and J. Venkataraman, "Feasibility of blood glucose estimation from real
time monitoring," 2013 IEEE Antennas and Propagation Society International
Symposium (APSURSI), Orlando, FL, 2013, pp.2055-2056.
[21] X. Xiao and Q. Li, "A Noninvasive Measurement of Blood Glucose Concentration
by UWB Microwave Spectrum," in IEEE Antennas and Wireless Propagation Letters,
vol. 16, pp. 1040-1043, 2017.
[22] Jean, B.R., Green, E.C., McClung, M.J.: ‘A microwave frequency sensor for non-
invasive blood-glucose measurement’. IEEE Sensors Appl. Symp., Atlanta, GA, 2008.
108
[23] Yilmaz, T., Hao, Y.: ‘Electrical property characterization of blood glucose for on-
body sensors’. Proc. 5th European Conf. on Antennas and Propagation (EUCAP),
Bolivar, 2011, pp. 3659–3662.
[24] Schwerthoeffer, U., Weigel, R., and Kissinger D., “Highly Sensitive Microwave
Resonant Near-Field Sensor for Precise Aqueous Glucose Detection in Microfluidic
Medical Applications”, I2MTC Proceedings, 2014.
[25] R. Baghbani, M. A. Rad and A. Pourziad, "Microwave sensor for noninvasive

glucose measurements design and implementation of a novel linear," IET Wireless
Sensor Systems, vol. 5, no. 2, pp. 51-57, 2015.
[26] "IEEE standard for safety levels with respect to human exposure to radio frequency
electromagnetic fields, 3 kHz to 300 GHz," IEEE Std C95.1, 2005 Edition, 2005.
[27] O.G. Martinsen, S. Grimnes and H.P. Schwan, "Interface Phenomena and
Dielectric Properties of Biological Tissue," in Encyclopedia of Surface and Colloid
Science, Marcel Dekker, Inc., 2002, pp. 2643-2643-2652.
[28] Gabriel, C., Gabriel, S., and Corhout E. “The dielectric properties of biological
tissues: I. Literature survey” Phys. Med. Biol. 41 2231–49, 1996.
[29] R. Pethig, "Dielectric properties of body tissues," Clinical Physics and

Physiological Measurement, vol. 8, pp. 5-12, 1987.
[30] H.P. Schwan, "Interaction of Microwave and Radio Frequency Radiation with
Biological Systems," Microwave Theory and Techniques, vol. 19, pp. 146-152, 1968.
[31] L.A. Geddes, and L.E. Baker. Principles of Applied Biomedical Instrumentation.
3rded. New York: Wiley, 1968.
[32] C. Furse, A. Douglas and C. H. Durney, Basic Introduction to Bioelectromagnetics.

CRC Press, 2009.
[33] T. Karacolak, E. C. Moreland, and E. Topsakal, “Cole–Cole model for glucose-

dependent dielectric properties of blood plasma for continuous glucose monitoring,”
Microwave Optical Technol. Lett., vol. 55, no. 5, pp. 1160–1164, 2013.
[34] S. Amendola, S. Milici, and G. Marrocco, “Performance of epidermal RFID dual-

loop tag and on-skin retuning,” IEEE Trans. Antennas Propag., vol. 63, no. 8, pp. 3672–
3680, Aug. 2015.
109
[35] Mobashsher, A.T.; Abbosh, A.M., "Artificial Human Phantoms: Human Proxy in
Testing Microwave Apparatuses That Have Electromagnetic Interaction with the
Human Body," Microwave Magazine, IEEE, vol.16, no.6, pp.42, 62, July 2015.
[36] D. Pozar, Microwave Engineering, 4th ed. Hoboken, New Jersey: John Wiley & Sons,
Inc., 2012.
[37] C.-S. Lee and C.-L. Yang, “Complementary split-ring resonators for measuring
dielectric constants and loss tangents,” IEEE Microwave Wireless Components. Letter,
vol. 24, no. 8, pp. 563–565, Aug. 2014.
[38] Collin, R.E.: Foundation for Microwave Engineering, 2nd ed. McGraw-Hill, New
York, 1992.
[39] Chen LF, Ong CK, Neo CP, Varadan VV, Varadan VK. Microwave electronics;
measurement and materials characterization. West Sussex, England: John Wiley, 2004.
[40] He, Yusheng & Li, Chunguang. (2010). Superconducting Receiver Front-End and
Its Application In Meteorological Radar. 10.5772/7189.
[41] K. Chang, Microwave Ring Circuits and Antennas, New York: John Wiley and
Sons, 1996.
[42] M. S. Boybay and O. M. Ramahi, “Non-destructive thickness Measurement using

quasi-static resonators,” IEEE Microw. Wireless Compon. Lett., vol. 23, no. 4, pp. 217–
219, Apr. 2013.
[43] C. Hsu, K. Chen, C. Lee and C. Yang, "Improved approach using multiple planar
complementary split-ring resonators for accurate measurement of permittivity," 2016
IEEE MTT-S International Wireless Symposium (IWS), Shanghai, 2016, pp. 1-4.
[44] ANSYS, Inc. ANSYSS HFSS Features [Online].

Available: http://www.ansoft.com/products/hf/hfss/new.cfm
[45] "ISO 15197:2013 In vitro diagnostic test systems -- Requirements for blood-
glucose monitoring systems for self-testing in managing diabetes mellitus,"
International Organization for Standardization, 2013. [Online]. Available:
https://www.iso.org/standard/54976.html.
[46] M. S. Boybay and O. M. Ramahi, "Material Characterization Using

Complementary Split-Ring Resonators," in IEEE Transactions on Instrumentation and
Measurement, vol. 61, no. 11, pp. 3039-3046, Nov. 2012.
110
[47] Houssam Ayad. Controle des performances des antennes par les metamateriaux.
Universite de Grenoble, 2012. English.
[48] S. R. Choudhury, S. K. Parui and S. Das, "Design of a novel bandstop filter using
log periodic based circular split ring slots," 2012 Students Conference on Engineering
and Systems, Allahabad, Uttar Pradesh, 2012, pp. 1-4.
[49] C. Balanis, Antenna Theory and Design, 3rd ed. Hoboken, New Jersey: John Wiley &
Sons, Inc., 2005, pp. 811-815.
[50] Ashraf T, Panhwar Z, Habib S, et al. Size of radial and ulnar artery in local
population. J Pak Med Assoc 2010;60:817–819.
[51] C. Chen, “Design of a compact microstrip quint-band filter based on the tri-mode
stub-loaded stepped-impedance resonators,” IEEE Microw. Wireless Compon. Lett.,
vol. 22, no. 7, pp. 357–359, Jul. 2012.
[52] K. Hsu, W. Hung, and W. Tu, “Compact quint-band microstrip bandpass filter
using double-layered substrate,” in Proc. IEEE MTT-S, Int. Dig., Seattle, WA, USA,
2013, pp. 1–4.
[53] K. W. Hsu, J.-H. Lin, and W. H. Tu, “Compact sext-band bandpass filter with
sharp rejection response,” IEEE Microw. Wireless Compon. Lett.,
vol. 24, no. 9, pp. 593–595, Sep. 2014.
[54] J. Ai, Y. Zhang, K. D. Xu, D. Li and Y. Fan, "Miniaturized Quint-Band Bandpass

Filter Based on Multi-Mode Resonator and$\lambda/4$Resonators With Mixed Electric
and Magnetic Coupling," in IEEE Microwave and Wireless Components Letters, vol.
26, no. 5, pp. 343-345, May 2016.
[55] W. H. Tu and K. W. Hsu, “Design of sext-band bandpass filter and sextaplexer

using semi-lumped resonators for system in a package,” IEEE Trans. Compon. Pack.
Manuf. Techn., vol. 5, no. 2, pp. 265–273, Feb. 2015.
[56] Skyworks [online]. Available:

http://www.skyworksinc.com/uploads/documents/SMV1705_Series_200076O.pdf
[57] Accu-Check PERFORMA [online]. Available: https://www.accu-chek.com
[58] Robins JA. Partial Least Squares Revisited. Long Range Planning. 2014
Jun;47(3):131.
111
[59] S. Lee, H. Lee, P. Abbeel, and A. Ng, “Efficient L1-regularized logistic
regression,” in proceedings of the national conference on artificial intelligence, vol. 21,
no. 1, 2006, p. 401.
[60] Gaussian processes for machine learning. Cambridge, Mass: MIT Press; 2006.
[61] Abdi, Hervé and Lynne J. Williams. “Principal Component

Analysis.” Encyclopedia of Biometrics (2009).
[62] Alin A, Agostinelli C. Robust iteratively reweighted SIMPLS. Journal of

Chemometrics. 2017 Mar;31(3):e2881.
[63] Clarke WL, Cox D, Gonder-Frederick LA, Carter W, Pohl SL. Evaluating clinical
accuracy of systems for self-monitoring of blood glucose. Diabetes Care. 1987;
10(5):622-8.
112

Mousa Thesis

Uploaded by

Copyright:

Available Formats

Mousa Thesis

Uploaded by

Document Information

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

Mousa Thesis

Uploaded by

Copyright:

Available Formats

AMERICAN UNIVERSITY OF BEIRUT

DESIGN OF A MICROWAVE SENSOR FOR NON-INVASIVE

I would like to genuinely acknowledge my thesis advisors Dr. Joseph Costantine

Moussa Semaan Bteich for Master of Engineering

Title: Design of a Microwave Sensor for Non-Invasive Blood Glucose Monitoring

LIST OF ILLUSTRATIONS……………………………………………. xii

LIST OF TABLES…………………………………………………………... xviii

1.1. Background …………………………………………………….…… 1

2.1. Introduction …………………………………………………………. 5

2.3. Planar Microwave Circuits …………………………………………. 9

3.1. Introduction …………………………………………………….…… 15

3.3.1. Polarization and Relaxation ……………………………... 17

3.4. Discussion ………………………………………………………….. 24

4. HIGH LOSS MATERIAL …………………………………………… 25

4.1. Introduction …………………..…………………………….……….. 25

4.2. Complex Permittivity………………………………………………... 25

4.3. Extraction of the Real Permittivity ε′r ……………………………… 26

4.4. Extraction of the Imaginary Permittivity ε′′r ………………………. 26

4.5. Effective Dielectric Permittivity ……………………………………. 27

4.6. Discussion ………………………………………………………….. 28

5. MATERIAL CHARACTERIZATION ……………………….. 29

5.1. Introduction …………………………………………………….…… 29

5.3.1. Non-Resonant Methods ………………………………… 31

5.4. Planar Resonators ………………………………………………….. 33

5.4.1. Straight Ribbon Resonator ………………………………. 33

5.5. Discussion …………………………………………………………. 37

6. DESIGN OF EM-BASED GLUCOSE SENSORS ………. 39

6.1. Introduction …………………………………………………….…… 39

6.4. Proposed Sensor #1………………………………………………….. 42

6.5. Proposed Sensor #2………………………………………………….. 46

6.5.1. Design Structure …………………………………………. 46

6.6. Proposed Sensor #3 ………………………………………………... 56

6.6.1. Design Structure …………………………………………. 58

6.7. Discussion …………………………………………………………. 72

7. SIMULATION AND MEASUREMENT OF THE

7.1. Introduction …………………………………………………….…… 73

7.3.1. Serum Measurements ……………………………………... 78

8.2. Regression Analysis ………………………………………………… 91

8.3. Results ………………..……………………………………………... 97

8.3.1. Serum Measurements……………………………………. 97

8.4. Discussion …………………………………………………………. 105

9. CONCLUSION AND FUTURE WORK ………………… 106

REFERENCES …………………………………………………………. 107

1. MMW measurement setup [10]………………………………………… 6

2. Proposed waveguides and measurement setup [12]…………………….. 6

3. Transmission variations with time [13]………………………………… 7

4. Tx and Rx antennas placed around pig’s ear [14]………………………. 8

5. Double ring resonators system [16]…………………………………….. 9

6. a) Antenna wrapped around subject's hand, and b) reference and

7. Spiral resonator sensor [22]…………………………………………….. 11

8. Relation between permittivity and glucose concentrations in function of

9. Sensor structure embedded in the measurement setup [24]……………. 12

10. S21 magnitude and phase responses for different glucose

11. Prototype of the designed filter loaded by a human thumb [25]………... 14

12. S11 and S21 parameters variations with time [25]……………………... 14

13. Dispersion of biological tissues [28]……………………………………. 19

15. Measurement setup [33]………………………………………………… 22

16. Dielectric properties of blood for different concentrations of glucose

17. The geometry of microstrip transmission line………………………….. 30

18. Transmission line setup [39]……………………………………………. 31