3.1. An Empirical Approach
Different frequencies have been utilized for blood glucose level detection in the literature, ranging from radio frequencies to millimetre waves. Although some applications at terahertz and infrared range have also been investigated (e.g., [
44,
45]), the scope of this review is limited to RF/microwave and some millimetre-wave applications. Most of the work reported focusses on narrow-band applications, with a few reporting wide-band behaviour. Since microwave diagnostic and treatment applications are based on the dielectric property discrepancies between the anomalous and healthy tissues, the behaviour of dielectric properties that is dispersive with respect to frequency is one of the primary constraints for frequency of choice.
Another related factor in frequency selection is penetration depth: as conductivity increases with frequency for all tissues, the electromagnetic wave encounters greater loss at higher frequencies, which can be related directly to how much tissue the wave can pass through and still be detectable at the system’s minimum threshold for detection. (A related parameter is the skin depth; we do not discuss this here, but the penetration depth is always greater than the skin depth.) For most implementations of sensors for glucose monitoring, a reflection mode is used; a few use a transmission mode (e.g., for systems placed on the ear lobe). The penetration depth is essentially a reflection-mode parameter; transmission-mode systems can be expected to have a maximum allowed sample thickness approximately twice that of the penetration depth (because the reflected wave passes through the tissues twice). Penetration depth decreases with frequency for all tissues. At low frequencies (e.g., below 100 MHz), the penetration depths for skin, fat and muscle would be more than thicknesses typically encountered; above 10 GHz, however, very little penetration into muscle can be expected, while penetration depths for skin and fat are of less than or equal to typical thicknesses [
46]. We note that the limits on exposure, particularly the Specific Absorption Rate [SAR], place an upper bound on transmit power, which translates into a maximum penetration depth for a tissue of given loss.
As noted before, microwave frequencies have been employed for breast cancer imaging and treatment purposes due to the dielectric property discrepancy between the benign and malign tissues [
47]. In microwave imaging, the employed frequency range is between 3 GHz and 7 GHz. The resolution of the microwave images increases at higher frequencies; however, penetration depth and wavelength decrease. To illustrate the penetration depth and wavelength in tissue, we plotted the behaviour two values with respect to frequency in muscle tissue. The muscle tissue dielectric properties have been utilized before in the literature to represent the lossy medium of the human body [
40]. The relative permittivity and conductivity of muscle tissue is shown in
Figure 4a; the change of wavelength and penetration depth in muscle tissue medium with respect to frequency is shown in
Figure 4b. Similar to high water-content tissues, the relative permittivity of the muscle tissue decreases with the increase in frequency, while the conductivity of the muscle tissue increases. Both the dielectric property change and the increase in frequency affects the wavelength and penetration depth. From
Figure 4b, it can be seen that both the wavelength and the penetration depth is less than 5 mm. This demonstrates that, above 10 GHz, the body tissues will be even more lossy and the propagating wave will quickly attenuate.
One other criterion that can be considered while choosing the frequency of operation is looking into the utilization of the bands. For example, the US Federal Communications Commission’s MedRadio spectrum allocation covers the 413–419 MHz, 426–432 MHz, 438–444 MHz, 451–457 MHz and 2360–2400 MHz ranges. These bands are specifically useful for implants and body-worn devices for off-body, on-body and in-body communications, since the signal is still quite strong for these bands. Other possible bands include the license-exempt 2.4–2.5 GHz and 5.725–5.875 GHz ISM bands. At 2.45 GHz, the wavelength and penetration depth in muscle tissue is around 22 mm for both quantities. At 5.8 GHz, the wavelength and penetration depth in muscle tissue is around 7.6 mm and 7.4 mm, respectively.
Lastly, it can be concluded from
Section 2 that the glucose-dependent dielectric property change is very limited, especially in the microwave region, such that the glucose-dependent change does not display a significant variation between the frequencies. This indicates that the limited change in blood dielectric properties due to the glucose variations can only be measured by employing a highly sensitive technique. Broadband dielectric property measurement techniques, such as the open-ended coaxial probe, are known to suffer from accuracy and repeatability limitations, whereas narrow band measurement techniques are known to be more precise. Therefore, empirically we can conclude that a narrow-band technique will be more sensitive to glucose-dependent dielectric property changes. Ultimately, the sensitivity of a resonator also depends on the measurement technique and the performance of the employed technique at the operation frequency. The Q factor, which is indicative of the measurement sensitivity, of narrow-band resonators is expected to increase at higher frequencies (that is, higher order modes for a given resonator tend to have higher Q factors). Considering the constraints imposed by the lossy nature of the biological tissue media (higher loss and smaller penetration depth with increasing frequency) and the band availability, plus the fact that resonators tend to be some multiple of a half-wavelength in size (hence, physically larger at lower frequencies), a narrow-band resonator operating at a narrow band frequency between 4 GHz and 7 GHz can be a viable option. Of course, it is still possible to employ other frequencies, both higher and lower, as can been seen in the literature. We now review these choices, with a discussion following in
Section 3.3.
3.2. Frequencies Employed in the Literature
Different frequencies have been employed in the literature to sense the glucose change. For example, in [
48], a monopole antenna operating between 1–6 GHz was designed. The antenna response between 1.5–3 GHz was observed to shift during simulations and while testing it with blood mimicking phantoms. However, it is known that the response of wideband and ultra-wideband antennas are less sensitive to the dielectric property changes in a medium. For instance, Vivaldi antennas are frequently employed for microwave medical imaging applications to both provide a wideband signal and prevent the antenna detuning due to close proximity to human body. Additionally broadband dielectric property measurement methods known to suffer from low measurement accuracy. Considering that the realistic dielectric property change with respect to change in glucose levels is very limited, there is a need to employ more sensitive techniques.
As mentioned above (
Section 2.3), an impedance spectroscopy approach was used in [
38]; here, we describe the second set of measurements using the bio-impedance parameter directly, measured using a system from Biopac. Initially, measurements were performed on agar phantoms with aqueous solutions using varying quantities of glucose: 0 mmol/L, 50 mmol/L (900 mg/dL), 100 mmol/L (1800 mg/dL) and 200 mmol/L (3600 mg/dL); again, it must be stated that these are not realistic values, so questions regarding the sensitivity are not addressed. The bio-impedance of glucose solutions with solutions having different glucose concentrations were calculated at 10 Hz and supported the expectation that the change in bio-impedance decreased with increasing glucose concentration. An additional set of measurements were performed on a non-diabetic test subject, in conjunction with direct measurements with a commercial portable blood glucose meter (ACCU-CHEK Performa). Measurements were taken over the course of a 135-minute period, at five-minute intervals for the bio-impedance and thirty-minute intervals for the blood meter, during a type of oral glucose tolerance test. After some signal processing and curve-fitting, it was observed that there was an inverse correlation between the measured bio-impedance and the measured glucose level [
38]. This work reported that the temperature, minimum and maximum blood volume and other components of blood (such as haemoglobin) might effect the bio-impedance calculations.
A spiral resonator operating between 628 and 677 MHz was introduced in [
30]. This resonator was not tested with realistic glucose values and the response was explored to retrieve the relative permittivity of the tissues. The sensitivity of the resonator can not be judged. It should be noted that the wavelength and penetration depth is quite large at these frequencies. Therefore, the response of the structures operating close to MedRadio bands can be affected by other factors, such as the size of the tissue loaded to the resonator. When the final application is considered, this could emerge as a problem when setting a calibration standard.
Another resonator was presented in [
26,
43], operating close to 2.45 GHz when loaded with four- and five-layer tissuemimicking materials (composed of dry skin, wet skin, fat, blood and muscle tissue). This resonator, which was not optimised for the wearable glucose monitor application, consisted of a microstrip patch resonator with two capacitively coupled feeding strips and had a Q-factor of around 4, making it relatively poor in terms of sensitivity. The penetration depth and wavelength at this frequency is still quite large (around 20 mm in muscle tissue); therefore, the calibration problem may still emerge at a smaller scale. One option to achieve matching for different loads is to utilize a impedance-matching circuit at these frequencies. However, this approach should be implemented so that glucose-dependent change in impedance will not be masked.
A serpentine-shaped capacitive structure operating between 3.0 GHz and 6.0 GHz was presented in [
33,
49]. The sensitivity of this structure was analysed for the best matching. Initially, the structure was designed to operate at 4.8 GHz; this frequency was chosen after analysing the penetration depth and reflections between different tissue boundaries. It is worth noting that most of the simulations were performed in commercial electromagnetic simulation programs; when the RF/microwave sensors are loaded with lossy materials (such as four-layered tissue-mimicking materials with frequency-dispersive dielectric properties), the simulations take longer to complete than simulations in air. Therefore, it is advised to run such simulations on workstations; even then, the cost of optimizing these sensors to operate at a certain frequency is high.
Two patch antennas operating at 2.45 GHz and 5.8 GHz were designed and tested with dextrose solutions in an attempt compare the performance of the antennas at these two frequencies [
50]. The antennas are mounted at the bottom of two 3D printed cups. The cups were filled with dextrose solutions having concentrations ranging from 0 mg/dL to 5000 mg/dL. From 0 mg/dL to 1000 mg/dL, the amount of dextrose was increased by 200 mg/dL. When the behaviour of the patch antenna was analysed while changing the dextrose amount, it was observed the operation frequency did not change. However, the matching of the antennas changed with the increase in dextrose levels. Although the change was not linear, it was observed that the matching of the antenna operating at higher frequency was more sensitive to the dextrose change. The response of both antennas are given in
Table 4. Since the antenna types were identical, the glucose-dependent change in antenna matching is attributed to the frequency of operation.
A resonator operating at 1.4 GHz was proposed in [
51]. The sensor, which had a Q-factor of about 800 in air [
51], was proposed to be placed on the abdomen region of the body, where its Q-factor was reduced to about 80 [
51]. The sensor was tested with humans and the response compared with the commercial glucose sensors. Also, an in vitro interference test technique was proposed to test the sensor performance with glucose and other materials. The resonator response shifted 600 kHz when the glucose levels were increased from 0 mg/dL to 600 mg/dL. The in vitro performance of this resonator does not only depend on the frequency; the structure itself also has an important role. Therefore, in the abdomen region the tissue should not be considered homogeneous in the 1.4 GHz frequency range. Changes in other parameters are likely to affect the resonator response. In [
51], the effect of other parameters was mitigated with a reference structure that was separated spatially from the tissue and main resonator, but otherwise identical to it. Hence, the change in resonant response (frequency and bandwidth) for the reference resonator can be used to track temperature via a calibration curve, thus allowing detection of permittivity changes with the main resonator caused by other factors. A clinical trial of this sensor involving 24 human subjects (eight non-diabetics, four Type-1 diabetics and 12 Type-2 diabetics) undergoing an oral tolerance test was reported [
52] and assessed using the Clarke Error Grid and the mean absolute relative difference (MARD) parameter. Two versions of the sensor were used (12 subjects per sensor); some differences between sensors were possibly evident, based on MARD values of 11% and 14% for the respective test groups, although no detail is provided on the test subject groups to allow identification of other possible causes. An overall MARD value of 12.5% was calculated. Although the majority of test samples were in regions A and B of the Clarke Error Grid, there were some values in the upper C region (attributed to unexpected movement by the test subjects in [
52]), demonstrating further work is necessary to enhance robustness. The comparison in time between the sensor and the reference glucose readings (taken using a YSI 2300 Glucose and Lactate Analyzer) was visually close in both papers [
51,
52]; curve-fitting models were developed in [
51] that have presumably been used in [
52] to produce estimated blood glucose levels (in mg/dL) directly.
Another microwave resonator operating at 6.53 GHz was proposed in [
53]. When a container of de-ionized water–glucose solution, with a concentration of 0.75 mg/mL, was placed on the resonator, the resonance frequency shifted to 3.43 GHz. The glucose concentration was then increased to 1 mg/mL, then in 1 mg/mL increments to 5 mg/mL. The resonance frequency shifted to 3.53, 3.93, 4.23 and 5.03 GHz for glucose concentrations of 1 mg/mL, 2 mg/mL, 3 mg/mL, 4 mg/mL and 5 mg/mL, respectively. Although a very good resonance shift is observed, it should be noted that the sensitivity can not be merely attributed to frequency of operation. Both the resonator structure and the frequency of operation, thus, the Q factor, are all parameters that needs to be considered in this work. The readers should note the units used, which have a factor of 100 difference to those used in this work, such that the normal range of glucose concentration is stated as from 0.75 mg/mL to 2.16 mg/mL by the authors of [
53]. The concentrations used are therefore similar to the values used in this review for quantitative comparisons.
A microstrip-line-based multi-band resonator, operating between 100 MHz and 500 MHz and 1.4 GHz to 1.8 GHz, was proposed in [
54]. The resonator was tested using glucose solutions, with concentrations from 0 mg/dL to 5000 mg/dL. The resonance shift, as well as the matching of the resonator, was observed for both frequencies. It was concluded that the resonator displayed a better sensitivity to the glucose change at higher frequencies.
In [
55], three versions of a resonator, operating at 1.92 GHz, 5.16 GHz and 7.16 GHz, was designed for measurement of glucose concentrations in microlitre volume solutions. A dielectric sensing cup with a microlitre volume was designed and integrated to these structures to hold the liquid. The Q factor of all three structures was investigated with solutions having different glucose concentrations. The Q factor changed by up to five units for glucose concentration ranging from 0% to 10%.
A spiral resonator was proposed in [
56], operating at 7.65 GHz when placed in aqueous glucose solution and operating at 7.77 GHz when placed in a sample of pig blood. During the tests, the sample under test was put into a Petri dish with a diameter of 8 mm, which was placed in turn on the resonator (first, the aqueous solutions were tested, followed by the blood samples). Glucose concentrations for the aqueous solutions were from 0 mg/dL to 600 mg/dL; for the pig blood samples, concentrations from 100 mg/dL to 600 mg/dL were tested. The observed shift in operating frequency was negligible; thus, the authors reported the change in matching of the resonator. For the aqueous samples, with concentrations ranging from 0 mg/dL to 600 mg/dL; the S
response decreased from −40 dB to −55 dB; for pig blood samples (concentrations ranging from 100 mg/dL to 600 mg/dL), the S
response decreased from −18 dB to −25 dB. The change in the S
response with respect to volume was also reported in this work. To the best of the authors’ knowledge, the sample volume should be chosen in such way that the electromagnetic energy completely attenuates within the material under test (MUT) at the operating frequency, in order to explore the true performance of the structure during such experiments. The effective permittivity of the medium measured by the resonator will then only depend on the permittivity of the substrate and the permittivity of the MUT. Since the glucose-dependent dielectric property change is very limited, this is a paramount parameter to explore the full potential of a microwave sensor for blood glucose monitoring.
A metamaterial-based resonator operating at 2 GHz was proposed in [
57]. In this work, the change in both amplitude and phase of S
was tracked. The resonator was simulated with digital phantoms, with the relative permittivity of the digital phantoms changed based on the glucose-dependent dielectric property change equations described in [
24]. To simulate the change in blood glucose levels between 0 mg/dL and 250 mg/dL, the relative permittivity of the digital blood-mimicking phantom was changed from 69.4 units to 47.5 units, respectively. When compared to the conclusions drawn in [
26], where the change in relative permittivity was expected to be 1 unit for glucose levels from 0 mg/dL to 216 mg/dL, these changes in permittivity values are very large. Since no experimental validation was performed, the performance of the proposed structure can not be fully judged.
The millimetre-wave part of the spectrum, specifically around 60 GHz, is the operating band selected by a company called MediWise for their GlucoWise system [
58,
59]. This was chosen “
…as the wavelength is short enough for a relatively compact antenna sensor and the penetration depth is large enough for interrogation of thin human tissue regions with sufficient blood concentration” (Saha et al., 2017 [
59]). The developed sensors are intended to work either on the ear lobe or the fleshy part of the hand between thumb and first finger and are based on a pair of patch antennas (resonators) acting in transmission mode. Standard in vitro measurements were conducted using aqueous solutions of glucose and the authors stated the system “
can detect as low as 0.025 wt% of glucose in water” [
59]. Ten non-diabetic male subjects underwent an intravenous glucose tolerance test (IVGTT) while wearing the system. Results for two subjects showed reasonable correlation; poor results for the other test subjects were attributed to hand motion and “
…gradual sliding of the holder during the session, possibly due to fatigue or stress” [
59], emphasising the challenges introduced by external factors. The lag between direct blood measurements and indirect tissue measurements was also evident [
59]. A more recent study involving an anaesthetised pig was reported [
58], again using an IVGTT to introduce glucose ‘spikes’ to the blood stream. The sensor was compared with a “
spectrophotometric clinical blood chemistry analyzer (iLab 650 by Diamond Diagnostics) and a commercially available glucometer (Contour Next EZ by Bayer)” [
58]. The antennas were located at different positions on the ear of the pig and with varying separations, to investigate variability and the detuning effect. The spikes in glucose level were evident in the sensor response, with a lag of about thirteen minutes. This lag was attributed, in part, to the distance from the injection site and to the known lag between venous blood samples and interstitial fluid. It was also stated that “
although the area is convenient for the sensor placement, it is not particularly rich in blood and contains significant amounts of interstitial fluid” [
58].